Integrating mathematical modelling and artificial intelligence to combat emerging viral syndemics: A systematic review

Abstract

Background

Global health security is increasingly threatened by emerging viral syndemics, where infectious viruses interact synergistically with socio-economic-structural vulnerabilities. Traditional reactive surveillance is increasingly insufficient to address the complexity of these interconnected crises. This systematic review aims to synthesize evidence for the “Silicon Shield,” a proactive computational framework integrating mathematical modelling and artificial intelligence (AI) to bridge a unified computational framework, and critical literature gaps for pandemic preparedness and response.

Methods

Following updated PRISMA 2020 guidelines, we searched PubMed, Scopus, and Web of Science (January 1990-July 2026) using Medical Subject Headings (MeSH). We evaluated the convergence of mechanistic frameworks (SIR, SEIR, ABM) with deep learning architectures (CNNs, Transformers, pLMs). Analysis prioritized Uncertainty Quantification (UQ) via Bayesian updating and causal inference through Directed Acyclic Graphs (DAGs) to mitigate ecological biases when modelling biosocial determinants.

Results

Synthesizing 364 sources, we formalize the “Silicon Shield” three-layer architecture: Data Pipes, Model Pipelines, and Decision Interfaces. Integrating real-time genomic and mobility data enables anomaly detection via ViraMiner and TCINet. EVEscape forecasts immune escape, while Social Vulnerability Indices (SVI) optimize resource allocation. The framework yields predictive performance, improved AUROC and probabilistic metrics including Weighted Interval Score (WIS) and Continuous Ranked Probability Score (CRPS). Explainable AI (xHAIM) translates outputs into clinical decision support.

Conclusions

The convergence of AI and mathematical modelling facilitates a fundamental transformation into a proactive global health defence system. Implementing the “Silicon Shield” requires interdisciplinary collaboration, standardized clinical validation, and a focus on global health equity to effectively mitigate future emerging virus syndemic threats.

Highlights

• •AI and math models form a kind of “Silicon Shield” against emerging viral syndemics.
• •AI models embed social determinants of health to map syndemic risks of emerging viruses.
• •Deep learning advances agnostic emerging virus discovery from mNGS data.
• •The EVEscape model forecasts viral immune escape before emergence.
• •Frameworks like CFIR and RE-AIM guide sustainable public health NPIs against emerging viruses.

1. Introduction

1.1. The fragility of global health in the 21st century: the shift to viral syndemics

The twenty-first century has witnessed a remarkable acceleration in the emergence of viral diseases, fundamentally transforming the landscape of global health security. Emerging infectious diseases (EIDs) have risen substantially, with approximately 60% originating as zoonoses and 72% of these derived from wildlife reservoirs [1]. Over the past two decades, multiple emerging viral outbreaks, including Severe Acute Respiratory Syndrome Coronavirus (SARS-CoV, 2002), H1N1 influenza (2009), Middle East Respiratory Syndrome Coronavirus (MERS-CoV, 2012), Ebola virus (EBOV) disease (2013-2016), Zika virus (ZIKV) disease (2015), and Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2, 2019-present), have collectively resulted in substantial global morbidity, mortality, and economic disruption [[2], [3], [4]]. More recently, rapid transmission of Bundibugyo virus (BV) within the Ituri Province of the Democratic Republic of the Congo (DRC) and its subsequent international spread to Kampala, Uganda [5,6], prompted the World Health Organization (WHO) to declare a Public Health Emergency of International Concern (PHEIC) on May 17, 2026 [7]. This epidemic of BV disease, a severe and frequently fatal form of EBOV disease for which no strain-specific vaccines or therapeutics are currently approved, is exacerbated by high population mobility, dense urban transmission hubs, and ongoing humanitarian crises in the region, thereby heightening the risk of further regional dissemination and underscoring the urgent need for coordinated international response and enhanced cross-border surveillance [[7], [8], [9]]. Recently, a highly pathogenic hantavirus outbreak was detected on the cruise vessel MV Hondius [10,11]. Genomic sequencing of the cases by the European Centre for Disease Prevention and Control (ECDC) confirmed that the etiological agent was Andes virus (ANDV) [12,13], a New World, negative-sense RNA virus transmitted mainly by the long-tailed pygmy rice rat (Oligoryzomys longicaudatus) [12]. ANDV is notable for its capacity for person-to-person transmission and for causing severe hantavirus cardiopulmonary syndrome (HCPS) [12]. During the event, the Oceanwide Expeditions vessel operated under technical guidance and public health supervision from Spanish Health Authorities, in coordination with the World Health Organization (WHO), while in Tenerife, Canary Islands [10,14]. The MV Hondius crisis exemplifies how structural environments, such as the confinement inherent to maritime travel and the complexities of international health coordination, can function as potent catalysts for viral dissemination [15,16].

These outbreaks [17,18] underscore the persistent challenge that emerging viruses pose in the 21st century and demonstrate that contemporary viral threats diverge from historical patterns through their fundamentally syndemic character, in which pathogens interact synergistically with pre-existing social and health vulnerabilities.

Singer et al. provided the foundational definition: a syndemic requires (1) two or more diseases clustering within a specific population; (2) contextual and social factors creating conditions for this clustering; and (3) adverse disease interaction that increases the health burden of affected populations [19]. Thus, the scientific community has paid increasing attention to the named “theory of syndemics” since the onset of the COVID-19 pandemic, due to the synergistic interactions of the disease with pre-existing political, structural, social, and health conditions [[20], [21], [22]].

Unlike traditional concepts of comorbidity, syndemics involve deleterious interactions between pathogens and social determinants of health, altering disease progression, treatment efficacy, and population-level outcomes [19]. The Coronavirus disease of 2019 (COVID-19) pandemic exemplifies this paradigm. The virus amplified existing fault lines of structural inequality, poverty, racism, and chronic disease burden [20,23,24]. Social determinants of health, including race, ethnicity, socioeconomic deprivation, educational attainment, occupation, and housing status, emerged as powerful predictors of COVID-19 incidence, hospitalisation, and mortality [23,[25], [26], [27], [28]]. Underlying this vulnerability, the drivers of viral emergence are multifaceted (Fig. 1A). Climate change has emerged as a critical amplifier of infectious disease risk, expanding both the geographic range and transmission seasons of vector-borne pathogens [[29], [30], [31]]. For instance, climatic suitability for Dengue virus (DENV) vectors has increased significantly, leading to over 5 million cases in 2023 [31]. Urbanisation, in low- and middle-income countries, has created ideal conditions for arboviral emergence. Globalisation and international travel have facilitated the rapid dissemination of pathogens across continents, often within hours [3]. Land-use changes, including deforestation, agricultural intensification, and extractive industries, bring wildlife populations into unprecedented proximity with human populations, facilitating zoonotic spillover events [29,32,33].

Fig. 1.

The paradigm shifts from reactive to proactive global health security in an era of viral syndemics. (A)Drivers of Viral Emergence: This panel represents the drivers that could favour an emerging virus outbreak (novel zoonotic spillover). It first outlines the environmental outcomes of climate change and land-use changes. These environmental factors include increasing desertification, rising temperatures, elevated atmospheric CO₂, more frequent and prolonged droughts, erratic rainfall patterns, progressive salinisation, intensified flooding, stronger storms, and rising sea levels. In addition, accelerating urbanisation and globalisation may further facilitate the emergence of this virus and catalyse local outbreaks or larger epidemics, with the potential to escalate into a pandemic. (B) Syndemic Vulnerability (The Crisis): This section illustrates how an emerging virus from a novel zoonotic spillover acts in synergy with existing vulnerabilities (i.e., social determinants and countermeasure gaps), which in turn raises the threat and favours the viral outbreak's impact on health and economy. This crisis is exacerbated by social determinants, including poverty, structural inequality, chronic disease burden, and critical countermeasure gaps, specifically when there are no established therapeutics, no reliable vaccines, and no rapid diagnostics against the viral emergence. (C) The Proactive Paradigm (The Solution): This panel represents a systems-based approach to global health consisting of a One Health approach, high-throughput mNGS surveillance, platform vaccines development, and the search for new and specific broad-spectrum therapeutics, where AI and mathematical modelling are integrated to help accelerate genomic interpretation, predict outbreak trajectories, and support agile, data-driven rapid response systems. Designs and templates were created with BioRender, and licenses were obtained for publication.

In response to these escalating threats, this convergence of ecological disruption, demographic shifts, and structural inequalities has required a fundamental paradigm shift in global health, from reactive clinical responses to proactive, systems-based approaches [34,35] (Fig. 1B and C). The failures of the COVID-19 response, most notably the profound inequities in access to diagnostics, therapeutics, and vaccines between high-income and low- and middle-income countries [35,36], have catalysed urgent calls to transform the global health architecture [35]. This transformation requires moving beyond traditional vertical disease control programmes toward integrated approaches that build pandemic preparedness capacity [34,36]. The establishment of high-level multisectoral task forces, investment in big data, and proposals for pandemic treaties and financing mechanisms represent initial steps toward this proactive paradigm [35,37].

However, sustaining this shift demands recognition that viral syndemics are fundamentally biosocial phenomena shaped by political, economic, and structural determinants [25,38]. Effective intervention requires systems-level approaches that address not only pathogen control but also the social, economic, and environmental conditions that enable syndemic clustering [24,38] (Fig. 1). Integrating advanced mathematical modelling and AI into this proactive framework enables forecasting outbreak trajectories (Fig. 1C), identifying geographic hotspots of syndemic vulnerability, optimizing resource allocation, and evaluating intervention strategies [24,25,38,39]. This establishes a form of “Silicon Shield”: a computational, AI-driven modelling defence against emerging viral syndemics.

Current literature lacks integrative frameworks that merge computational epidemiology and artificial intelligence to address the biosocial complexities of viral syndemics. This review bridges this critical gap by synthesizing predictive computational modelling, genomic surveillance, and advanced AI architectures into a cohesive “Silicon Shield” paradigm. This unified framework explicitly connects molecular tracking, population transmission dynamics, and socioeconomic vulnerabilities to establish a proactive, equity-informed global health defence strategy against syndemic outbreaks of emerging viruses.

1.2. The challenge of emerging pathogens

The emergence of novel viral pathogens presents a profound vulnerability for global populations particularly during the critical early stages of an outbreak, when effective medical countermeasures, rapid diagnostics, therapeutics, and vaccines are not available [40,41] (Fig. 1C). The World Health Organization incorporated “Disease X″ into its Research and Development Blueprint priority pathogen list, formally acknowledging the threat posed by a previously unknown infectious agent with epidemic or pandemic potential for which medical countermeasures are insufficient or non-existent [[40], [41], [42], [43], [44]]. This designation represents the awareness that a novel pathogen not currently known to cause human disease could precipitate a major public health emergency [40,43]. The subsequent emergence of SARS-CoV-2 exemplified this prescient concern, demonstrating the catastrophic impact such pathogens can exert on global health security [41,44].

Traditional vaccine development timelines span 10-15 years, rendering them essentially useless for outbreak response [45,46]. Even with accelerated development pathways demonstrated during COVID-19, vaccine development required approximately 11 months [47,48]. Therapeutic development faces similar constraints, with median clinical development times of 7.3 years for infectious disease drugs [48].Even in accelerated outbreak settings, populations remain vulnerable for months or years before interventions are ready [48,49].

Diagnostic challenges during the early phases of viral outbreaks further amplify this vulnerability by delaying pathogen identification, hindering case detection, and preventing the implementation of targeted control measures [[50], [51], [52]] (Fig. 1C). When a novel pathogen emerges, molecular or serological diagnostic assays are initially non-existent [53]. The early COVID-19 response exemplified these diagnostic limitations: delays in the availability of PCR reagents and narrow case definitions led to unrecognised human-to-human transmission and systematic under-ascertainment of cases [36,54,55]. Emerging viruses often present with non-specific febrile illness or symptoms overlapping with endemic diseases, creating diagnostic confusion that can persist for weeks or months [55,56]. DENV and Chikungunya virus (CHKV), for example, share mosquito vectors, endemic areas, and nearly identical initial clinical presentations, yet require vastly different management strategies [56,57], illustrating how diagnostic uncertainty can compromise patient care even for known pathogens. Consequently, the application of bioinformatic tools to metagenomic next-generation sequencing (mNGS) holds the potential to refine the clinical diagnosis of emerging, re-emerging, and new viruses [56].

From a technological standpoint, current diagnostic technologies, while increasingly sophisticated, face inherent limitations when confronting unknown pathogens. Targeted diagnostic approaches, including multiplex PCR and conventional serological assays, rely fundamentally on a priori knowledge of a pathogen's genome or antigenic structure, meaning an untargeted or unknown pathogen will not be detected [50,51]. mNGS offers an agnostic approach that can non-specifically sequence all detectable nucleic acids in a sample without prior knowledge, theoretically enabling detection of novel viruses [50,53,58]. However, mNGS faces substantial barriers to widespread clinical deployment, including lower sensitivity compared to targeted PCR (particularly at low viral loads), requirements for large amounts of genetic material, complex bioinformatic analyses, high costs, and limited availability in resource-constrained settings where many zoonotic spillovers originate [50,53,58,59] (Fig. 1C).

These infrastructure gaps for pandemic preparedness are particularly acute in low- and middle-income countries (LMICs), where most emerging infectious diseases originate [59,60]. Surveillance systems in these settings remain under-resourced, fragmented, and reactive, with limited laboratory capacity, inadequately trained personnel, insufficient funding, and weak data integration systems [59,61]. The COVID-19 pandemic similarly exposed these critical weaknesses, with early testing delays in multiple countries hindering containment efforts [62]. Furthermore, beyond surveillance, the absence of established treatments for emerging pathogens creates profound clinical and ethical challenges. When patients present with severe illness caused by a novel pathogen, clinicians have no evidence-based therapeutic options and must rely primarily on supportive care [63,64]. The 2014-2016 EBOV outbreak starkly illustrated this therapeutic vacuum: despite over 28,000 cases, no licensed treatments existed, and experimental therapeutics could only be evaluated through clinical trials conducted under extraordinarily challenging outbreak conditions [49]. Since pivotal clinical trials can only be conducted during active outbreaks when cases are available, this demands substantial pre-outbreak preparation [45,49,63].

To proactively address these risks, the development of broad-spectrum antiviral agents and platform vaccine technologies constitutes a key approach to closing this critical gap. Vaccine platforms, including mRNA, viral vector, and protein subunit technologies, enable rapid adaptation to new pathogens by substituting antigenic sequences while maintaining proven platform backbones [47,[64], [65], [66]] (Fig. 1C). The COVID-19 pandemic validated this approach: mRNA vaccines progressed from viral genome sequencing to emergency authorisation in under one year, a timeline that would have been unattainable with traditional vaccine platforms [47,66].

However, technological innovation alone cannot overcome the structural barriers that perpetuate vulnerability. Regulatory frameworks, although increasingly flexible, still require substantial time for safety and efficacy evaluation [45,46,66,67]. Perhaps most critically, the economic incentives for developing medical countermeasures against emerging pathogens with uncertain market potential remain fundamentally misaligned with public health needs, resulting in chronic underinvestment in preparedness research and development [46,63,68]. Hence, the COVID-19 pandemic demonstrated both the transformative potential of accelerated medical countermeasure development and the persistent inequities that undermine global health security. While high-income countries achieved vaccine coverage exceeding 70% within 18 months of vaccine authorisation, many LMICs struggled to secure even basic supplies, with some countries reporting less than 10% coverage two years into the pandemic [35,47]. This vaccine inequity reflected deeper structural problems, including intellectual property barriers, concentrated manufacturing capacity, inadequate financing mechanisms for procurement in LMICs, and fragmented global governance structures [35,66].

Therefore, addressing the vulnerability created by emerging pathogens requires a fundamental transformation in how the global community approaches pandemic preparedness. This includes sustained investment in pathogen research [53,60], alongside strengthening laboratory capacity and diagnostic infrastructure in regions where emergence is most likely. Emphasis must be placed on agnostic surveillance approaches, including metagenomic sequencing and wastewater monitoring, that can detect novel pathogens without prior knowledge [50,[59], [60], [61]]. Furthermore, it necessitates the development and stockpiling of broad-spectrum therapeutics and platform-based vaccine candidates [40,[63], [64], [65]]. To support these scientific efforts, reforming regulatory pathways to enable faster evaluation while maintaining safety standards is imperative. This must be paired with the creation of sustainable financing mechanisms that de-risk investment in medical countermeasures targeting pathogens, with uncertain commercial markets [45,46,66]. As highlighted previously, these investments must integrate a One Health approach and address structural drivers of vulnerability [3,29,[32], [33], [34],38,60].

To bridge these persistent gaps in preparedness, the integration of advanced mathematical modelling and artificial intelligence offers powerful tools for early detection, forecasting, and intervention optimisation, which is supported by the synergy between artificial intelligence, nanotechnology, and biosensing for the design of pandemic prediction and preparedness even for future Disease X [[69], [70], [71]], as explored in subsequent sections of this review. However, the success of these computational approaches depends on pandemic responses that prioritise social equity and address the root causes of vulnerability within healthcare systems [72].

2. Methodology and literature search strategy

To ensure a transparent, rigorous, and reproducible synthesis of this multidisciplinary literature, this systematic review followed updated PRISMA 2020 reporting standards [73] (Fig. 2 briefly outlines the overall methodological workflow).

Fig. 2.

PRISMA 2020 Flow Diagram for the Present Systematic Review. Schematic of the identification, screening, eligibility, and inclusion phases following updated PRISMA 2020 guidelines [73]. The search covers PubMed, Scopus, and Web of Science (1990-2026), prioritizing 2020-2026 literature on viral outbreak preparedness, surveillance, diagnostics, mathematical models, and AI. It incorporates 2025-2026 emerging outbreak data from Organizations and news sources. Quantitative placeholders ([n]) indicate record counts at each stage.

2.1. Information sources and search strategy

A systematic search across three primary electronic databases, PubMed, Scopus, and Web of Science, was executed for literature published. Thus, in accordance with updated PRISMA 2020 reporting guidelines, the literature search covered publications spanning from 1990 to 2026, with primary emphasis on literature published between 2020 and 2026 to capture current advancements in viral outbreak preparedness, disease surveillance, diagnostic innovations, and mathematical models’ interventions and artificial intelligence applications. The search strategy combined Medical Subject Headings (MeSH) and free-text terms tailored to three core conceptual domains:

• -Pathogen & Epidemic Context: “viral syndemic” OR “emerging infectious diseases” OR “pandemic preparedness” OR “biosocial threats";
• -Computational & Mathematical Modelling: “mathematical model” OR “SEIR” OR “agent-based modelling” OR “stochastic simulation");
• -Artificial Intelligence & Data Integration: “artificial intelligence” OR “machine learning” OR “deep learning” OR “natural language processing” OR “protein language models";
• -All the keywords used for this Systematic Review: “Emerging viruses; Epidemics/pandemics; Syndemics; Predictive mathematical models; Artificial intelligence; Preparedness”;
• -We have combined all the above presented MeSH with the indicated keywords;
• -Furthermore, we have incorporated data on emerging viral outbreaks from 2025 through 2026 up to the date of submission: The PRISMA 2020 flow diagram (Fig. 2) indicates that these references have been obtained from different sources, such as Organizations, viral outbreak news, etc.

2.2. Eligibility criteria

Studies were selected based on pre-defined inclusion and exclusion criteria:

Inclusion Criteria: (i) Peer-reviewed primary research articles, systematic reviews, or validated technical reports; (ii) Studies integrating mathematical, computational, or AI frameworks with epidemiological or clinical data; (iii) Literature evaluating viral pathogens with epidemic or pandemic potential; and (iv) Articles published in English.

Exclusion Criteria: (i) Publications with a non-viral focus (e.g., strictly bacterial or fungal infections without syndemic relevance); (ii) Articles lacking epidemiological or clinical integration (e.g., purely theoretical mathematical formulations without biological context); (iii) Studies lacking computational or methodological validation; and (iv) non-peer-reviewed commentaries, editorials, or conference abstracts without detailed methods.

2.3. Selection process and PRISMA flow framework

The identification, screening, eligibility, and inclusion processes are briefly illustrated in the PRISMA 2020 flow diagram (Fig. 2).

In our review of the literature, the primary indicators compared vary by computational domain: predictive sensitivity/specificity for AI-driven variant escape and host-pathogen models (particularly within literature associated with emerging viruses and viral outbreaks); mathematical, machine learning, and deep learning architectures for reproduction parameters (R₀, R_t) and epidemiology studies; and dedicated metrics for diagnostic, environmental, and social conditions. While probabilistic forecasting scores, such as Weighted Interval Score (WIS), Continuous Ranked Probability Score (CRPS), empirical calibration, and decision utility, are relevant for ensemble models, the extreme methodological and data heterogeneity across the included literature precluded a quantitative meta-analysis using a single metric. Instead, results were combined using a narrative synthesis framework organized into the main functional layers of our proposed “Silicon Shield” architecture: (1) molecular evolution and immune escape against virus outbreaks, (2) population-level transmission dynamics, and (3) AI-assisted surveillance and clinical diagnostics for virus outbreaks. More specifically, for AI-driven molecular evolution and immune escape models (layer 1), the primary evaluation metrics reported in the included literature were predictive sensitivity, specificity, and area under the receiver operating characteristic curve (AUC/AUROC), which are the standard measures for assessing classification and diagnostic performance in machine learning applied to infectious diseases [74,75]. For population-level transmission dynamics (layer 2), the dominant metrics were epidemiological parameter estimates, particularly the basic and effective reproduction numbers (R₀, R_t), alongside point forecast accuracy measures such as mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²) [76,77]. Where probabilistic epidemic forecasts were available, proper scoring rules were employed: the Weighted Interval Score (WIS), a discrete approximation of the Continuous Ranked Probability Score (CRPS), served as the primary metric for evaluating the calibration and sharpness of quantile-based forecasts, as established by the US and European COVID-19 Forecast Hubs [[78], [79], [80]]. The WIS can be decomposed into components measuring sharpness (concentration of prediction intervals) and penalties for over- and underprediction, and can be applied on transformed (e.g., logarithmic) scales to yield epidemiologically interpretable relative errors that reflect accuracy in predicting time-varying epidemic growth rates [81]. For AI-assisted diagnostics (layer 3), evaluation centred on standard classification metrics, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and AUROC, as reported in domain-specific systematic reviews and meta-analyses [74,82].

Despite the relevance of these individual metrics, the extreme methodological heterogeneity across computational domains, encompassing fundamentally different model classes (mechanistic ODEs, agent-based simulations, convolutional neural networks, gradient boosting, large language models), distinct data modalities (genomic sequences, medical images, time-series surveillance data, environmental variables), and incompatible outcome definitions, precluded a pooled quantitative meta-analysis using any single metric [83,84]. This challenge is well recognised in systematic reviews of AI applications in healthcare, where substantial heterogeneity (I²>98%) frequently persists even within relatively homogeneous subgroups [75,85]. Accordingly, results were synthesised using a structured narrative approach following the Synthesis Without Meta-analysis (SWiM) reporting guideline, which provides a transparent framework for organising, presenting, and interpreting findings when statistical pooling is inappropriate due to diversity in study characteristics [86,87]. Within this framework, studies were grouped thematically according to the three functional layers of the proposed ‘Silicon Shield’ architecture, and findings were summarized by direction and magnitude of effect within each layer, with explicit reporting of the metrics used by each study to enable cross-study comparison while preserving methodological integrity.

Records retrieved from initial database queries were deduplicated and uploaded for screening. Titles and abstracts were evaluated independently to exclude non-relevant literature. Full-text articles were subsequently assessed against the eligibility criteria, with reasons for exclusion recorded at each stage (e.g., lack of biosocial integration, purely theoretical focus, or non-viral scope). All records were imported into EndNote 20 and duplicates were removed. The selection process was summarized using a PRISMA 2020 flow diagram (Fig. 2).

3. Results

3.1. Socio-economic and governance impacts: navigating vulnerabilities

3.1.1. Multisectoral consequences

Viral syndemics generate cascading socioeconomic consequences that propagate through interconnected economic systems, affecting households, small businesses, and large corporate infrastructure [[88], [89], [90], [91], [92]] (i.e., this concept is represented in Fig. 1B). The COVID-19 pandemic provided an unprecedented natural experiment demonstrating these multisectoral impacts, with global consumption losses amounting to $3.8 trillion, triggering 147 million full-time equivalent job losses within the first months [89,90]. These primary effects were compounded by global supply chain disruptions, which propagated economic shocks across interconnected markets independently of local disease burden [89,90,93].

At the microeconomic level, viral syndemics disproportionately affect economically vulnerable populations, exacerbating pre-existing inequalities and threatening long-term social mobility [94,95]. These impacts were profoundly regressive, disproportionately affecting women, youth, and workers with lower educational attainment. These cohorts experienced higher rates of job displacement and income loss, with self-employed and informal labour, the most precarious segments in developing economies, facing the greatest vulnerability, who ultimately bore the brunt of pandemic-induced income losses [94].

Pandemic-induced transitions into unemployment have exacerbated food insecurity, precipitating a deleterious cycle wherein economic deprivation heightens susceptibility to infection, subsequently compounding systemic poverty and financial volatility [94,95].

Parallel to these household effects, small businesses experienced particularly severe impacts, with the number of active business owners in the United States plummeting by 3.3 million (22%) between February and April 2020, the largest drop on record [96]. Historically underrepresented groups experienced disproportionate losses: African-American businesses experienced a 41% drop in activity, Latinx businesses 32%, Asian businesses 26%, and immigrant-owned businesses 36% [96]. The pandemic's economic shocks were amplified by the underlying financial fragility of small enterprises. Empirical evidence from a survey of over 5800 firms indicates that those with monthly expenditures exceeding $10,000 typically maintain only a two-week liquidity buffer, underscoring a critical lack of operational resilience [97].

At the macroeconomic level, viral syndemics disrupt both supply and demand, generating cascading effects throughout interconnected global production networks [88,[91], [92], [93]]. Direct socioeconomic impacts of airborne and droplet-borne infectious disease outbreaks include increased morbidity, mortality, and healthcare expenditures. These, in turn, drive indirect effects such as employment crises, reductions in workforce size, demand shocks, supply chain disruptions, rising production costs, service and business interruptions, and GDP losses [88]. Together, these factors constitute what we term “syndemic vulnerability”: the compounding fragility arising from the convergence of biological, economic, and structural deficits during emerging virus outbreaks, as illustrated in Fig. 1B.

In this context, the literature consistently demonstrates that prompt outbreak containment and the strategic allocation of resources are fundamental to reconciling pandemic mitigation with macroeconomic stability, thereby preventing supply chain volatility from further compromising public health outcomes [93,98].

Correspondingly, the tourism, transportation, and service industries sustained disproportionate losses, characterised by an unprecedented contraction in the travel and hospitality sectors [91,92,99]. Trade shocks extended beyond immediate pandemic periods, with evidence that epidemic diseases can cause trade diversion and permanent restructuring of trade relationships, particularly for countries with prolonged outbreaks [99]. This suggests that many nations are confronting more than transient trade disruptions; rather, they face enduring structural challenges stemming from the fundamental reorganisation of global value chains [99].

3.1.2. Community resilience and policy

To mitigate these profound multisectoral impacts, predictive tools must incorporate socio-economic factors (Fig. 1B). Thus, the integration of social vulnerability indices into epidemiological modelling and public health planning represents a critical advance toward equity-centred pandemic response, enabling the identification of communities at heightened risk and guiding targeted resource allocation [[100], [101], [102], [103], [104]]. The CDC (Centers for Disease Control and Prevention) Social Vulnerability Index (SVI) has emerged as a primary tool for assessing community-level vulnerability to adverse health outcomes during public health emergencies [101,103].

The predictive value of social vulnerability indices extends beyond hotspot identification to forecasting disease severity and mortality outcomes [100,102,105]. Comparative analysis reveals that various geographic social disadvantage indices consistently demonstrate significant positive associations with COVID-19 incidence and mortality [100]. Despite divergent methodologies regarding component measures and weighting schemas, these tools provide a robust consensus on the link between socioeconomic vulnerability and health outcomes [100]. This predictive reliability was observed across all four major metrics evaluated: the SVI (Social Vulnerability Index), ADI (Area Deprivation Index), CCVI (COVID-19 Community Vulnerability Index), and MH-SVI (Minority Health-Social Vulnerability Index). This convergent validity suggests that multiple indices can assist public health leaders in targeting vaccines, therapeutics, and other resources to regions most vulnerable to negative outcomes [106,107], with selection among indices potentially guided by local policy frameworks and operational feasibility [100].

However, demographic risk is not static. Spatial and temporal analyses reveal that relationships between social vulnerability and disease outcomes are dynamic, shifting as epidemics evolve [101,102,104]. While the initial phase of the COVID-19 pandemic was concentrated within highly connected urban centers with lower baseline vulnerability, the subsequent geographical expansion into socioeconomically disadvantaged counties resulted in disproportionately elevated incidence and mortality rates [102,107]. Geographically weighted regression analyses demonstrate that relationships between social vulnerability and disease outcomes vary substantially across regions, necessitating locally tailored rather than uniform national approaches [101]. Furthermore, vulnerable communities often faced a double burden of increased risk and lower vaccination rates, highlighting the need to analyse individual SVI themes to design highly targeted interventions [108].

Despite their clear utility, integrating social vulnerability indices into predictive models for pandemic preparedness necessitates addressing several methodological and practical challenges [104]. First, indices developed for general disaster preparedness may not capture vulnerability dimensions specific to infectious disease outbreaks, such as occupational exposure risk, household transmission potential, or access to testing and treatment [100]. Second, county-level indices may mask substantial within-county heterogeneity, with vulnerable populations concentrated in specific neighbourhoods or communities [101]. Third, the dynamic nature of vulnerability, as epidemics evolve and interventions are deployed, requires models capable of updating in real-time [102]. Fourth, translating vulnerability assessments into actionable resource-allocation decisions requires clear prioritisation frameworks that balance equity considerations with epidemiological efficiency [100].

3.1.3. Implementation science frameworks

Evaluating the application of the RE-AIM and CFIR frameworks to support the sustainability and cultural adaptability of community-based interventions in highly vulnerable social contexts.

Bridging the gap between identifying vulnerable populations and delivering effective interventions necessitates robust methodological approaches. The successful translation of evidence-based interventions into sustainable community practice requires systematic attention to implementation processes, contextual factors, and adaptation needs, addressed by established implementation science frameworks including RE-AIM and the Consolidated Framework for Implementation Research (CFIR) [[109], [110], [111], [112]]. These frameworks are particularly relevant for community-based health interventions targeting vulnerable populations during pandemics [109,[113], [114], [115], [116], [117]].

The RE-AIM framework, encompassing Reach, Effectiveness, Adoption, Implementation, and Maintenance, provides a comprehensive evaluation model for assessing the public health impact of interventions [110,117]. RE-AIM has been applied extensively in public health and health behavior change research, with over 2800 citations since its introduction two decades ago [117]. The framework has demonstrated utility across clinical, community, and corporate settings, with increasing application in diverse domains beyond its original focus on health promotion [114,117]. A systematic review of RE-AIM applications found that Reach was the most frequently reported dimension (92.9%), followed by Implementation (90.3%), Adoption (89.7%), Effectiveness (84.5%), and Maintenance (77.4%) [113]. While users find the framework practical and well-established, obtaining data for all its dimensions remains challenging, underscoring the need for technical capacity building [118].

In contrast to RE-AIM's evaluative focus, the CFIR provides a complementary determinant framework for identifying barriers and facilitators to implementation success [111,112]. CFIR consolidates constructs from multiple implementation theories, models, and frameworks into a comprehensive taxonomy organized across five domains: Innovation (characteristics of the intervention being implemented); External context (including policies, incentives, and external pressures); Organizational context (including culture, climate, and readiness); Individuals (characteristics of those involved in implementation), and Strategies and activities for implementation [111,112].

The application of CFIR in community health interventions has demonstrated utility for identifying context-specific barriers and facilitators that influence implementation success [115,116]. CFIR evaluations in low- and middle-income countries highlight that addressing barriers (such as cultural incompatibility or mistrust) and leveraging facilitators (like community engagement and health system strengthening) are prerequisites for successful implementation [116].

Integrating CFIR with implementation strategy taxonomies, particularly the ERIC (Expert Recommendations for Implementing Change) compilation, enables systematic matching of identified barriers with evidence-based strategies, significantly improving the efficiency of implementation efforts compared to non-systematic approaches [111,119]. The CFIR-ERIC matching tool provides a structured approach for selecting implementation strategies based on specific barriers [119].

To maximize their utility during global health emergencies, applying these frameworks to pandemic preparedness requires critical adaptations: timelines must be compressed using rapid assessments, strategies must be highly adaptive to evolving epidemiology, equity must be centred to avoid exacerbating disparities, and sustainability planning must begin immediately [109,111,112,116,118].

By synthesizing implementation science with advanced mathematical modelling and AI, the proposed ‘Silicon Shield’ constitutes a formidable, multifaceted framework for global pandemic preparedness. Predictive models incorporating SVIs identify vulnerable communities, while CFIR and RE-AIM guide the cultural adaptation and delivery of interventions. Supported by AI-powered real-time monitoring, this integration, central to the “Silicon Shield”, helps build resilient health systems that could protect the most vulnerable populations during future viral syndemics [120,121]. All these data are summarized in Table 1, which covers the application of AI to socio-economic and Governance impacts of syndemic-based models.

Table 1.

Socio-Economic and Governance Impacts of syndemic-based models.

Key Theme (Subsection)	Main Concepts and Models	Advantages	Potential Evolution	Cited References
Multisectoral Consequences (3.1.1).	Syndemic vulnerability; supply chain disruption modelling; macroeconomic shock analysis; microeconomic fragility; regressive impact assessment.	Explicit quantification of inequity amplification across sectors; identification of “bidirectional” poverty-infection cycles; mapping of global trade restructuring.	Real-time tracking of informal labour market shifts; automated impact assessments on global production networks; long-term socio-economic recovery forecasting.	[[88], [89], [90], [91], [92], [93], [94], [95],99]
Community Resilience and Policy (3.1.2).	Social Vulnerability Indices (SVI, ADI, CCVI, MH-SVI); Geographically Weighted Regression (GWR); Dynamic vulnerability mapping.	Evidence-based targeted resource allocation; predictive reliability for incidence and mortality; enables locally tailored policy rather than uniform national approaches.	Integration of dynamic, real-time spatial data to update risk profiles; transition from county-level to hyperlocal (neighbourhood) analysis; automated equity-efficiency balancing.	[[100], [101], [102], [103], [104], [105], [106], [107]]
Implementation Science Frameworks (3.1.3).	RE-AIM (Reach, Effectiveness, Adoption, Implementation, Maintenance); CFIR Taxonomy; ERIC matching tool; Cultural adaptation frameworks.	Systematic identification of implementation barriers; structured methodology to bridge the “know-do” gap; ensures cultural suitability and intervention sustainability.	Development of rapid implementation assessment tools for crises; automated matching of barriers to evidence-based strategies; community-led feedback loops for AI model refinement.	[[109], [110], [111], [112], [113],[116], [117], [118], [119], [120], [121]]

3.2. Epidemiological perspectives: the evolution of mathematical modelling

3.2.1. Foundations of transmission dynamics

How deterministic and stochastic mathematical models remain fundamental in quantifying transmission mechanisms (e.g., R₀) and estimating vital epidemiological parameters.

Mathematical modelling of infectious diseases evolved significantly in the early 1900s. Sir William Heaton Hamer (1862-1936) was a British physician and a pivotal figure in the transition of epidemiology from a descriptive field to a mathematical science [122,123]. Ross's Nobel Prize-winning research proved the transmission link between mosquitoes and humans [124]. Ross also demonstrated, through mathematical modelling, that reducing mosquito density could lead to malaria elimination, although these conclusions were initially met with scepticism within the scientific community. Over the subsequent century, mathematical models have provided an explicit framework for understanding malaria transmission dynamics in human populations [125]. One of the principal conclusions of the Ross-Macdonald model is that malaria endemicity is highly sensitive to changes in adult mosquito survival, implying that malaria can be controlled more efficiently with adulticides than with larvicides [126].

Mathematical modelling provides critical quantitative insights into infectious disease dynamics that transcend observational epidemiology [[127], [128], [129]]. At the foundation of this approach are compartmental models [130], which partition populations into discrete states and describe their dynamics using differential equations [[131], [132], [133]]. The classic Susceptible-Infectious-Recovered (SIR) model, first formulated by Kermack and McKendrick in 1927, remains the conceptual cornerstone of transmission modelling, dividing populations into susceptible individuals who can acquire infection, infectious individuals who could transmit the pathogen, and recovered individuals who have gained immunity [131,134]. Its extension, the Susceptible-Exposed-Infectious-Recovered (SEIR) model (Fig. 3), provides more realistic representations of diseases with incubation periods, including influenza, COVID-19, and most viral pathogens [131,132,135]. Almost every contagious disease generates its own SIR-type model [130].

Fig. 3.

Mathematical Modelling and Simulating Intervention Scenarios. (A) Foundations of Transmission Dynamics: This panel details the theoretical framework of disease spread. (a.1) Compartmental Models (SIR/SEIR): Illustrates the progression of individuals through Susceptible [S], Exposed (Latent) [E], Infectious [I], and Recovered/Immune [R] states. The transitions between these states are governed by specific rates (-βSI, σE, γI). (a.2) Deterministic vs. Stochastic Modelling: Compares deterministic models, which are used for large populations to calculate average paths using differential equations (dI/dt), against stochastic models, which account for randomness and probabilistic outcomes critical during early outbreaks or rare events. (a.3) Fundamental Epidemiological Parameters: Highlights key metrics bridging theory and applied simulation, including the Basic Reproduction Number (R₀), the Effective Reproduction Number (R_t), and the Herd Immunity Threshold (1-1/R₀). R₀ is defined as an emergent property of the pathogen, host, and environment. (B) Simulating Intervention Scenarios (‘What-If’ Analyses): This panel demonstrates how models are applied for systems-level optimisation and resource allocation. (b.1) Agent-Based Modelling (ABM): Highlights models like Covasim that simulate individual interactions, activity patterns, and contact networks. (b.2) Intervention and Decision-Making Scenarios: A decision tree outlines various interventions, including Non-Pharmaceutical Interventions (NPIs) like school closures, mask-wearing, and physical distancing; Targeted Vaccination Strategies utilising mRNA platforms; and Quarantine & Testing protocols using PCR diagnostics, rapid tests, and isolation. Designs and templates were created with BioRender, and licenses were obtained for publication.

Compartmental epidemiological models are typically characterised by systems of non-linear autonomous ordinary differential equations (ODEs) [130]. The non-linearity inherent in these systems arises from the interaction between classes, where the transmission dynamics are primarily governed by the contact rate between susceptible (S) and infected (I) populations, mathematically represented by the product “β(SI)”, where “β” denotes the effective infection rate. As autonomous systems, the governing parameters remain time-invariant, ensuring that the population dynamics are determined solely by the current state variables. Because these non-linear systems generally lack explicit closed-form solutions, their behaviour is analysed through numerical methods and qualitative dynamical systems theory. Closed-form analytical solutions to these systems are generally not obtainable, necessitating numerical integration to characterise the temporal evolution of each compartment, particularly the Exposed (E) and Infected (I) classes that govern epidemic dynamics [130,131]. The qualitative behaviour of these solutions is therefore analysed through equilibrium and stability theory.

Central to this analysis is the identification of equilibria, states where all derivatives are set to zero, specifically the Disease-Free Equilibrium (DFE), where I = 0, and the Endemic Equilibrium, where infection persists at a stationary level. These states are characterised by fundamental parameters, including contact, mortality, natality, and immunity rates. Thus, the long-term trajectory of an outbreak is determined by the stability of the resulting equilibria. Given the system's non-linearity, stability is therefore assessed via linearisation around these equilibrium points. By evaluating the Jacobian matrix at each equilibrium, the local dynamics can be approximated by a linear system whose eigenvalues dictate the behaviour of the model. Then, an equilibrium is locally asymptotically stable if all eigenvalues have negative real parts, indicating the outbreak will dissipate; conversely, if any eigenvalue is positive, the equilibrium is unstable, suggesting that the introduction of a single infected individual may precipitate a widespread epidemic.

These models generate epidemic curves that are determined by fundamental parameters, most notably the basic reproduction number (R₀) [130,131,[136], [137], [138]]. R₀ represents the average number of secondary infections generated by a single infectious individual [139,140]. This metric calculates the herd immunity threshold as “1 - 1/R₀”, providing quantitative targets for vaccination programmes [139] (Fig. 3).

The R₀ number is intended to be an indicator of the contagiousness or transmissibility of infectious and parasitic agents (measles, polio, influenza, Ebola virus disease, HIV disease, a diversity of vector-borne infectious diseases, and many other communicable diseases) [137]. Notably, in the case of snail-borne trematode infection as the fascioliasis disease with an environmental reservoir, R₀ represents a multi-host/environmental threshold rather than a direct person-to-person transmission metric [141,142]. Then, optimal study on fractional fascioliasis disease models could be performed by generalized Fibonacci polynomials and/or by Katugampola Fractional Derivative [143], which builds a nine-compartment human-animal-snail-environment fractional system with the Caputo-Katugampola derivative solved by a Chebyshev spectral method, and explicitly derives the basic reproduction number from disease-free and endemic equilibrium stability analysis, with sensitivity analysis of transmission parameters [143,144].

R₀ is commonly defined as the expected number of secondary infections arising from a single individual during their entire infectious period, in a population of susceptible individuals [145,146]. From this definition, it follows that when R₀ < 1, each infected individual generates, on average, fewer than one new infection. Consequently, the infection is expected to be eliminated from the host population, or the microorganism cleared from the individual. If R₀ > 1, the pathogen could invade the susceptible population. This threshold behaviour is the most important and useful aspect of the R₀ concept [138,145]. The potential size of an outbreak or epidemic is often based on the magnitude of the R₀ value for that event [140,147,148], and R₀ can be used to estimate the proportion of the population that must be vaccinated to eliminate an infection from that population [139,145,149].

However, R₀ is an emergent property of the pathogen-host-environment system, varying with biological characteristics, social patterns, and environmental conditions [137]. R₀ could be easily misrepresented, misinterpreted, and misapplied. R₀ is not a biological constant for a pathogen, a rate over time, or a measure of disease severity. R₀ is a population-level theoretical construct, not a biological constant, and should not be conflated with real-time transmissibility estimates such as R_t, the effective reproduction number. Thus, R₀ is rarely measured directly, and modelled R₀ values are dependent on model structures and assumptions. Some R₀ values reported in the scientific literature are likely obsolete. This parameter must be estimated, reported, and applied with great caution as it is a fundamental metric that is far from simple [137]. Moreover, as an outbreak progresses, the effective reproduction number, R_t, typically declines from its initial value R₀. In epidemiology, R_t denotes the average number of secondary cases generated by a single infected individual at a specific time point, whereas R₀ represents the initial reproduction number in a completely susceptible, unmitigated population [150,151]. This shift highlights the importance of time-varying estimates, which are essential for accurately tracking and understanding epidemic trajectories as interventions and behavioural changes alter transmission dynamics over time [140,148]. The calculation of R₀ from empirical data presents substantial methodological challenges [148,152,153]. Complexities such as clustered social structures, age-stratified mixing, and spatial heterogeneity render a single population-level R₀ problematic and necessitate cautious interpretation [137,148] (Fig. 3).

In this context, modellers must choose between deterministic and stochastic approaches [134,[154], [155], [156]]. Deterministic models, formulated as systems of ordinary differential equations, are computationally efficient and mathematically tractable, facilitating analytical characterisation and estimation of average epidemic trajectories [131,133,134,154,155]. In contrast, stochastic models explicitly incorporate random variation in transmission events, accounting for chance effects that deterministic models average out [[154], [155], [156], [157]]. This stochasticity becomes critically important for small populations, rare events, or the early stages of outbreaks [134,154]. Critically, both approaches can produce qualitatively different predictions regarding the probability of large outbreaks and the initial growth rate of epidemics [154,158] (Fig. 3).

Parallel to the evolution of mechanistic, differential-equation modelling, computational epidemiology underwent a major transformation with the birth of internet-era digital syndromic surveillance [159]. Early pioneering initiatives demonstrated that non-traditional, real-time digital data streams could capture outbreak signals far faster than conventional clinical reporting [160,161]. In 1994, the Program for Monitoring Emerging Diseases (ProMED-mail) established a foundational model for internet-based global threat detection by aggregating and human-curating unstructured, narrative epidemiological reports [[162], [163], [164]]. By the late 2000s, this approach evolved toward automated big-data paradigms, epitomised by Google Flu Trends (GFT) in 2008, which attempted to estimate regional influenza-like illness (ILI) activity in real time using aggregated search query logs [165,166]. Although GFT ultimately demonstrated key vulnerabilities, such as sensitivity to algorithmic drift, news-driven search panics, and overestimation of outbreak magnitude, it served as a watershed moment for the field [[167], [168], [169], [170]]. These early digital surveillance efforts demonstrated both the immense potential and the methodological pitfalls of uncalibrated digital proxies, creating the direct impetus for today's advanced natural language processing (NLP), multimodal data fusion, and transformer architectures that extract fine-grained, robust syndemic intelligence from global real-world data [39,171,172].

Recent advances have extended classical compartmental frameworks to incorporate greater biological realism [132,135,173,174]. Techniques such as delay-differential equations, time-varying parameters to capture temporal heterogeneity, and non-first-order transmission kinetics enable a more accurate representation of complex infection dynamics, including multiple epidemic peaks [132,135,174]. Recent modelling of lung cancer tumour-immune interactions using generalized Laguerre polynomials (fractional-order differential equations) illustrates the utility of fractional calculus for capturing complex immune system dynamics [175]. Thus, the COVID-19 pandemic catalysed rapid development of compartmental models incorporating asymptomatic transmission, pre-symptomatic infectiousness, and quarantine [135,173]. Models explicitly representing pre-symptomatic (P) and asymptomatic (A) compartments alongside quarantined (Q) individuals revealed that these often-overlooked populations drive substantial transmission, fundamentally altering optimal control strategies [173]. Critically, identifying and quarantining these individuals through widespread testing proved far more effective than symptom-based surveillance alone [173].

3.2.2. Simulating intervention scenarios

The capacity to simulate complex scenarios, to explore “what if” questions about interventions that have not yet been implemented, and to compare alternative strategies under identical conditions (Fig. 3), constitutes a central contribution of modelling to public health decision-making [127,128,176]. Unlike clinical trials or observational studies that can only evaluate interventions actually deployed, computational models enable systematic exploration of the entire intervention landscape, quantifying trade-offs and identifying optimal strategies before committing resources [128,176,177]. This capability proved indispensable during the COVID-19 pandemic, when governments faced urgent decisions about implementing non-pharmaceutical interventions (NPIs) in the absence of vaccines or effective therapeutics [[177], [178], [179], [180]] (Fig. 3).

Computational models for evaluating NPIs span a spectrum of complexity and realism, from parsimonious compartmental models with a few parameters to highly detailed agent-based simulations [178,[181], [182], [183]]. Age-structured compartmental models that stratify populations by age groups and incorporate age-specific contact matrices enable evaluation of age-targeted interventions, including school closures and prioritized vaccination [178,180]. These models revealed critical insights during COVID-19, demonstrating that the interaction between vaccination rollout speed, vaccine targeting strategies, and the timing of NPI relaxation would determine whether healthcare systems would be overwhelmed by surges in hospitalizations and deaths [180]. Modelling results provided quantitative bounds on expected deaths and hospitalisations under multiple scenarios, enabling policymakers to balance public health protection against economic and social disruption [180].

To operationalise these theoretical foundations for public health decision-making, researchers increasingly turn to more granular simulation techniques. Agent-based models (ABMs) represent individual people as autonomous computational agents with specific demographic characteristics, daily activity patterns, and contact networks, simulating disease transmission through explicit interactions [181,182,184,185] (Fig. 3).

Covasim, an open-source agent-based simulator developed specifically for COVID-19, incorporates country-specific demographic data, realistic transmission networks, age-specific disease outcomes, and viral-load-based transmissibility [181] (Fig. 3). Critically, Covasim supports extensive intervention modelling including physical distancing, mask wearing, testing strategies (symptomatic and asymptomatic), isolation, contact tracing, quarantine, and vaccination, with each intervention incorporating realistic implementation details including delays, loss-to-follow-up, and imperfect compliance [181]. Despite this complexity, the model can simulate realistic scenarios on standard laptops in under 1 min, enabling rapid iteration and extensive sensitivity analyses [181]. Covasim has been applied, in collaboration with health agencies across more than a dozen countries, to examine epidemic dynamics and inform policy decisions in real time [[186], [187], [188], [189]].

In summary, mathematical modelling provides critical quantitative insights into infectious disease dynamics (Fig. 3). The foundation of this approach relies on compartmental models, such as SIR and SEIR, which partition populations into discrete states to evaluate transmission mechanisms. These frameworks generate epidemic curves governed by key parameters, most notably the basic and effective reproduction numbers, R₀ and R_t (Fig. 3). Model selection is context-dependent: deterministic models are typically used to characterise average trajectories in large populations, whereas stochastic models are preferred to capture inherent randomness during early outbreak phases. Advanced computational frameworks, including agent-based models (ABMs) such as Covasim, incorporate realistic contact networks and demographic data, thereby strengthening the “Silicon Shield.” These advanced models empower policymakers to simulate counterfactual “what-if” scenarios, enabling the systematic evaluation and optimisation of NPIs before real-world implementation (Fig. 3).

Altogether, this section synthesises the core themes about the foundational and applied aspects of mathematical modelling in infectious disease epidemiology, as summarized in Table 2.

Table 2.

Epidemiological perspectives. The evolution of mathematical modelling.

Key Theme (Subsection)	Main Concepts and Models	Advantages	Potential Evolution	Cited References
Foundations of Transmission Dynamics (3.2.1).	Deterministic compartmental models (SIR/SEIR); stochastic extensions; basic reproduction number (R0 = βcD); effective reproduction number (Rt); herd-immunity threshold (1 – 1/R0); delay-differential and non-first-order variants.	Quantify transmission mechanisms and intervention efficacy with analytical tractability; produce universal epidemic-curve shapes; distinguish average trajectories from probabilistic outcomes; computationally efficient for large populations.	Greater incorporation of real-world heterogeneity (age-stratified mixing, clustered contacts, spatial structure); improved handling of stochastic extinction risk in early/outbreak phases; systematic integration of social-behavioural parameters.	[129,131,132,[135], [136], [137],173,174]
Simulating Intervention Scenarios (3.2.2).	Agent-based models (e.g., Covasim); age-structured compartmental models; counterfactual “what-if” scenario testing for NPIs (quarantine, school closure, vaccination prioritisation).	Enable rapid, ethical evaluation of untested policies; quantify trade-offs between health and socio-economic costs; support real-time decision-making on standard hardware; explicitly model compliance and delays.	Inclusion of dynamic behavioural adaptation and economic feedback loops; rigorous external validation against post-implementation data; explicit modelling of equity impacts across vulnerable populations.	[127,128,178,180,181]

3.3. Molecular and genomic frontiers: AI in viral evolution and pathogen discovery

3.3.1. Tracing evolution and host-pathogen interactions

AI is transforming our understanding of viral evolution by surpassing traditional approaches to viral characterisation, which have relied heavily on sequence-homology tools such as BLAST and HMMER and are limited in their ability to detect highly divergent viral genomes or to predict the functional consequences of mutation [190]. To overcome these computational limitations, AI-driven approaches learn complex patterns directly from sequence and structural data [39,190,191]. In this context, AlphaFold and ESMFold predict three-dimensional protein structures with atomic precision, enabling rapid identification of conserved vulnerabilities across viral families without prior structural data [39]. The application of machine learning-based protein structure prediction to the Flaviviridae family enabled detailed mapping of viral glycoprotein structures that outperformed traditional sequence-based approaches and revealed more than 100 previously unrecognised glycoproteins. Therefore, machine learning models utilising interactome data accurately predict host-pathogen protein interactions, offering insights into disease pathogenesis, post-infection complications, and therapeutic targets [39,192].

This structural insight is crucial for understanding how mutations affect viral fitness and immune evasion [192,193]. Graph neural networks (GNNs) [190,193] trace how single mutations ripple throughout biological networks, clarifying systemic consequences [193]. This proves essential for understanding tissue tropism, the preferential infection of specific cell types or organs by viruses, which emerges from complex interactions between viral surface proteins and host cell receptors, entry factors, and intracellular machinery [194,195]. Additionally, AI models trained on these virus-host interaction data could predict which cell types a novel or mutated virus is likely to infect, informing clinical management and public health responses [194,196]. Moreover, explainable AI techniques (e.g., SHAP (see Fig. 4), CAM) applied to multi-omics data illuminate the internal logic of these models, addressing the “black box” challenge and providing reliable mechanistic insights [190,193,195,197].

Fig. 4.

AI Architectures in Global Health: From Surveillance to Diagnostics. (A) Population-Level AI Surveillance: Global surveillance is being transformed by the integration of open-source data, global news reports [170,224], and wearable devices such as smart rings and watches [170,259]. Early-warning systems use natural language processing (NLP) and Bayesian algorithms to probabilistically evaluate these data streams [170,225]. This identifies localised anomalies days to weeks before official recognition, facilitating Early Outbreak Detection mapped on a global scale [170,224]. Systems such as HealthMap and EpiWatch provide vital early-warning capabilities for emerging threats [39,159,170,225]. (B) Patient-Level Multimodal Diagnostics: In clinical settings, Machine Learning models and deep learning architectures like CNNs, ResNet50, & XGBoost process heterogeneous datasets [[260], [261], [262], [263], [264], [265]]. The integration of multimodal data includes Chest Imaging, Clinical Tabular Data, and Meteorological/Environmental Data [39,[266], [267], [268]]. This significantly enhances Rapid and Precise Clinical Diagnosis compared to unimodal approaches. The deployment of these tools in high-stakes environments requires Multimodal Integration and Explainable AI (e.g., SHAP) to interpret complex algorithms [269]. This fosters Clinician Trust and ensures a Rapid Public Health Response in resource-constrained or novel outbreak scenarios [121]. Designs and templates were created with BioRender, and licenses were obtained for publication.

3.3.2. Next-generation deep learning architectures

To reliably transform raw, multi-site metagenomic next-generation sequencing (mNGS) streams into actionable outbreak signals, genomic architectures require an end-to-end processing pipeline. Agnostic pathogen detection relies on unbiased de novo assembly and alignment pipelines [198], paired with statistical background-decontamination algorithms to filter out host and environmental sequencing noise [199,200]. Extracted viral genomes are subsequently processed through protein Language Models (pLMs) and phylogenetic frameworks to infer mutational fitness landscapes and escape trajectories [201]. Finally, these molecular fitness metrics are integrated into multi-modal AI models alongside spatial mobility telemetry and epidemiological surveillance data, generating unified spatio-temporal outbreak forecasts [202].

Building upon these structural and evolutionary predictions, AI also revolutionises the direct detection of pathogens in complex biological matrices. mNGS offers an unbiased approach to pathogen detection [50,190,197] (see Fig. 4), and its integration with AI accelerates outbreak responses by rapidly uncovering novel viruses and transmission chains [197]. Thus, Convolutional Neural Networks (CNNs) [190,[203], [204], [205]] excel at detecting localised sequence motifs, short, recurring patterns in DNA or protein sequences that often correspond to functional elements like transcription factor binding sites or protein domains [193]. By leveraging these advanced neural networks, tools like ViraMiner, a pioneering CNN-based method for identifying viral genomes in human metagenomic samples, significantly outperform conventional machine learning methods [204] and can detect highly divergent viruses among sequences labelled as “unknown” by standard alignment approaches, potentially uncovering infectious aetiologies of diseases with previously unidentified causes [204,206]. Moreover, DeePaC (Deep Learning Approach to Pathogenicity Classification) substantially outperforms traditional alignment-based and taxonomic classification methods in identifying highly divergent or novel pathogens [203,205]. Likewise, Recurrent Neural Networks (RNNs) excel at analysing time-series genomic data to track viral evolution and detect long-range functional dependencies [190,193].

Transformer architectures employ self-attention mechanisms to detect long-range dependencies globally [190,191,193,197]. For instance, LucaProt, a transformer-based tool, identified new potential RNA virus species in meta-transcriptomes from diverse ecological systems based on RNA-dependent RNA polymerase sequences, demonstrating superior performance compared to traditional bioinformatics approaches [39]. Furthermore, pLMs have revolutionised viral genomics by learning the “grammar” of protein sequences [207,208]. Models such as ESM-2 (Evolutionary Scale Modelling), when combined with graph neural networks (e.g., R-DELF), achieve high predictive accuracy for viral evolution and support the assessment of vaccine suitability [209].

The capacity of AI to uncover hidden transmission chains through anomaly detection represents a critical advance for outbreak investigation and control [192,210]. Traditional phylogenetic analysis reconstructs transmission networks by comparing viral genome sequences and inferring ancestral relationships, but it requires substantial sequence divergence to resolve individual transmission events and can be confounded by within-host diversity and sampling biases [192]. AI-powered anomaly detection identifies unusual patterns in metagenomic data, such as unexpected viral sequences, atypical mutation patterns, or novel recombinant genomes, that may indicate unrecognised transmission chains, zoonotic spillover events, or cryptic circulation in unsampled populations. During outbreak investigations, these methods can identify superspreading events, detect imported cases from external sources, and reconstruct transmission networks with greater resolution than traditional approaches. This represents what has been named the “AI-first” paradigm for outbreak preparedness [192].

Beyond their computational capabilities, these architectures are evolving from purely theoretical discovery tools into practical clinical decision-support systems, enabling frontline practitioners to use CNNs and transformer-based tools and models to resolve previously “unknown” viral aetiologies in metagenomic samples, and thereby obtain critical evidence to initiate or adjust antimicrobial therapy and implement isolation protocols several days before traditional culture-based testing yields a definitive result [190,197,204,[211], [212], [213], [214], [215], [216]]. Notably, AI anomaly detection identifies unusual metagenomic patterns to uncover cryptic transmission chains or zoonotic spillovers [197], and the Taxon-aware Compositional Inference Network (TCINet) refines its predictions by enforcing compositional constraints, propagating evidence across taxonomic hierarchies, and calibrating confidence using entropy and variance-based metrics [217]. This approach addresses the well-documented tendency of deep learning models toward overconfident predictions [217].

Despite these transformative capabilities, AI-driven metagenomic analysis faces substantial challenges, including heavy computational burdens that widen the “AI divide” for LMICs, as well as dataset biases that disproportionately favour well-characterised viruses from high-income settings [190,197,210]. Most importantly, despite theoretical promise, a significant challenge persists between AI research and clinical deployment. AI innovations predominantly remained confined to academic journals. Addressing these barriers will require global collaboration to standardise data collection and sharing across institutions and countries, alongside the development of integrated architectures that combine machine learning with causal inference to enhance both robustness and interpretability, thereby catalysing the translation of AI innovations into real-world infectious disease control [210]. Thus, forward-looking strategies to address these challenges include federated learning [197] and explainable AI techniques to enhance biological interpretability and enable researchers to understand why models make specific predictions [197]. In this scenario, hybrid frameworks that combine AI pattern recognition with classical phylogenetics and functional annotation emerge to leverage the complementary strengths of both approaches, enhancing scalability and maintaining interpretability [190]. All these combinations and complementary approaches constitute the core of the “Silicon Shield”.

3.3.3. Predicting immune escape

The capacity to forecast viral evolution and predict immune escape mutations before they emerge clinically represents one of the most transformative applications of AI in pandemic preparedness, enabling proactive rather than reactive vaccine and therapeutic development [191,218].

An Evolutionary model of Variant Effect (EVE) has been developed to quantify the viral escape potential of mutations at scale (EVEscape) [39,218]. EVEscape is a computational, AI-driven framework designed to forecast viral escape [218], representing a landmark achievement that integrates deep learning on historical sequence data with biophysical measurements [39,218]. Although EVEscape was trained exclusively on viral sequences available before 2020, that is, before the emergence of SARS-CoV-2, it has demonstrated accuracy comparable to high-throughput experimental antibody-escape scans in anticipating SARS-CoV-2 variation [39,218]. The EVEscape model could generate predictions immediately upon identification of a novel pathogen, before surveillance sequencing accumulates, experimental antibody panels are developed, or three-dimensional structures of antibody-antigen complexes become available [218]. EVEscape's generalisability to other viruses, including influenza, HIV, and understudied viruses with pandemic potential such as Lassa and Nipah, has demonstrated that the framework captures fundamental principles of viral evolution rather than pathogen-specific patterns [39,218]. This broad applicability was further supported by the emergence of the SARS-CoV-2 Omicron VOC, which provided a critical window to demonstrate EVEscape's efficacy in performing real-time risk analysis on novel viral lineages [219]. This allows developers to target conserved regions less susceptible to evasion [218].

Complementing this structural approach, viral language models offer a complementary approach to immune escape prediction by treating viral evolution as analogous to natural language, where mutations represent “word” changes that alter immune recognition, and preserving viral infectivity [191,208]. Viral language models (pLMs) trained on influenza haemagglutinin, HIV-1 envelope glycoprotein, and SARS-CoV-2 spike protein sequences can accurately predict structural escape patterns from sequence data alone, without requiring experimental antibody-binding measurements or three-dimensional structural information [208]. Likewise, Machine Learning-guided Antigenic Evolution Prediction (MLAEP) integrates structure modelling, multi-task learning, and genetic algorithms to predict the viral fitness landscape [220]. MLAEP accurately infers variant order along antigenic evolutionary trajectories, with predictions correlating strongly with variant sampling times, demonstrating that the model captures the actual historical path of viral evolution [220]. MLAEP identified novel mutations in immunocompromised COVID-19 patients and in emerging variants such as XBB.1.5, with its predictions subsequently validated in vitro [220].

Predictive models incorporating population-level immunity dynamics represent the next generation of infectious disease forecasting [221,222]. A comprehensive mechanistic model incorporating deep mutational scanning data, antibody pharmacokinetics, and regional genomic surveillance data predicts variant-specific relative numbers of susceptible individuals over time [222]. The model demonstrates that ongoing pandemic dynamics continue to shape variant-specific population immunity, which in turn determines transmission potential and thus variant fitness [222]. Therefore, integration of epidemiological, immunological, and genomic data enables identification of key drivers of viral evolution [219,221]. Models incorporating these shifting drivers identified, with high accuracy (AUROC 0.92–0.97), mutations that subsequently spread (i.e., up to four months before their emergence) across multiple phases of the pandemic [219]. Applied prospectively, these models forecast mutations likely to spread [219].

Similarly, the SARS-Escape Network demonstrates how natural language processing techniques can recognise significant spike protein sequences using sequence data alone [223]. This approach proved applicable beyond SARS-CoV-2 to other viruses, including influenza, monkeypox, and HIV, where sufficient escape mutant data and protein sequence datasets are available [223].

Despite remarkable progress, immune escape prediction faces inherent limitations. Viral evolution is fundamentally stochastic, influenced by random mutation, genetic drift, and unpredictable epidemiological dynamics. Models can identify mutations with high escape potential but cannot predict with certainty which variants will emerge or when, as these outcomes depend on stochastic events and complex ecological factors. Moreover, modelling high-dimensional epistatic fitness landscapes remains computationally challenging.

The integration of AI-driven immune escape prediction into vaccine development requires careful consideration of how to translate probabilistic forecasts into actionable design decisions. Multivalent vaccines targeting multiple predicted escape variants, broadly neutralising antibody cocktails designed against conserved epitopes identified through escape modelling, and adaptive vaccine platforms represent promising strategies. These challenges notwithstanding, AI-driven immune escape prediction represents a fundamental shift from reactive to proactive pandemic preparedness, offering new opportunities to stay ahead of viral evolution rather than perpetually chasing it [191,218,219,[221], [222], [223]].

These data and developments are summarized in Table 3, which outlines applications of AI at the molecular level to elucidate viral evolution and forecast viral variability.

Table 3.

Molecular and genomic frontiers for AI in viral evolution and pathogen discovery.

Key Theme (Subsection)	Main Concepts and Models	Advantages	Potential Evolution	Cited References
Tracing Evolution and Host-Pathogen Interactions (3.3.1).	Protein-structure prediction (AlphaFold/ESMFold); graph neural networks (GNNs) for interactome mapping; multi-omics integration; tissue-tropism prediction.	Agnostic detection of functional mutations without sequence homology; identification of conserved structural vulnerabilities; prediction of post-infection complications and tropism.	Harmonisation of heterogeneous omics datasets; enhanced model interpretability (saliency mapping/SHAP); validation across divergent viral families and diverse host populations.	[39,190,[192], [193], [194]]
Next-Generation Deep Learning Architectures for Metagenomics (3.3.2).	CNNs (ViraMiner, DeePaC); RNNs/Transformers for sequence analysis; agnostic pathogen detection in raw reads.	Rapid, database-independent discovery of novel viruses; superior sensitivity for divergent sequences; reconstruction of hidden transmission chains.	Reduction of false positives in low-viral-load samples; standardised training datasets for global strain diversity; integration with real-time clinical workflows.	[197,203,204,207]
Predicting Immune Escape (3.3.3).	EVEscape; viral language models (pLMs); MLAEP; population immunity dynamics models; SARS-Escape Network.	Proactive prediction of escape mutations before clinical emergence; generalisability across multiple viral families; integration of structural, genomic, and immunological data.	Incorporation of complex epistatic interactions; real-time updating with global surveillance data; translation of probabilistic forecasts into actionable adaptive vaccine design decisions.	[39,208,218,220,223]

3.4. Predictive tools and diagnostics: AI-assisted early warning systems

3.4.1. Global surveillance and natural language processing

AI-trained algorithms transform global surveillance by parsing open-source data to detect anomalies in real-time, identifying emerging threats days to weeks before official recognition [160,170,224], which could be considered as an “early warning system” (Fig. 4). The implementation of AI in outbreak detection requires (1) NLP of a large volume of multi-dimensional open-source data to detect early warning signals; (2) identifying local and regional patterns in the detected signals; (3) modelling and simulating outbreak behaviours; and (4) rapidly identifying misinformation and disinformation that can cripple pandemic responses [224].

HealthMap exemplifies this success, utilising NLP and Bayesian machine learning to assess web-based reports across multiple languages and extract hyperlocal geographic data [170,225]. HealthMap automatically extracts geographic information, tying multiple reports together and identifying disease clusters that cross-jurisdictional authorities may have missed [170]. For instance, it facilitated early recognition of COVID-19 and previously provided early indications of the emergence of H1N1 in Mexico and the 2019 vaping-associated pulmonary disease outbreak [170]. Similarly, analysis of event-based surveillance systems during the 2018-2019 avian influenza season found that 23-43% of outbreaks were detected before official notification [226]. EpiWatch, developed at the University of New South Wales, represents another sophisticated NLP-based surveillance system that extracts disease, date, and location information [39,224]. EpiWatch proved effective in global mpox surveillance during the 2022 epidemic, and its modular architecture enables rapid adaptation to novel pathogens, a critical capability for responding to Disease X scenarios [39,227] (Fig. 4A).

Complementing NLP data extraction, Bayesian algorithms contribute essential capabilities, enabling probabilistic outbreak detection that accounts for uncertainty and prior knowledge [[228], [229], [230], [231]]. These models offer graduated responses rather than binary alerts, providing higher predictive accuracy and stronger uncertainty quantification when data are sparse or noisy [228,229]. The Bayesian Outbreak Detection Algorithm (BODA) detected influenza outbreaks with a 4-5-day delay and minimal false-positive alarms, outperforming modified CUSUM (Cumulative Sum) algorithms [232]. CUSUM is widely used for early detection and temporal tracking of influenza, COVID-19, and other infectious diseases, monitoring changes in surveillance data trends, such as increasing case numbers or ICU admissions [233]. A critical advantage of Bayesian approaches is their ability to generate forecasts even when local data are scarce or unavailable by incorporating informative priors from related settings [231] (Fig. 4A).

Beyond dedicated epidemiological platforms, broader digital surveillance using social media and search engine data provides population-level signals of disease activity that can precede official case reports by days to weeks [160,234,235] (Fig. 4A). These models show strong correlations (r > 0.8) with official case data [160] and one global real-time disease activity database detected 94% of outbreaks identified by the WHO an average of 43.4 days earlier [236]. Google Trends data have also demonstrated strong correlations with both clinical case data and wastewater viral concentrations for multiple pathogens, including influenza A, respiratory syncytial virus, norovirus, and mpox [237] (Fig. 4A). This approach offers a complementary data stream, which is particularly valuable in areas lacking robust clinical testing infrastructure [237]. However, digital surveillance faces challenges related to data quality and representativeness, including potential confounding from media coverage, demographic biases in internet access and search behaviour, and difficulty distinguishing genuine health signals from media-driven noise [160,238].

Wearable devices and biosensors represent an emerging frontier, enabling detection of infections from subclinical physiological changes [121,170,239] (Fig. 4A). A study using Oura Ring data demonstrated high accuracy (AUC 0.85) and low false-positive rates for population-level fever surveillance during infectious disease outbreaks [239]. Emerging technologies, including electrochemical sensors and wearable biosensors, represent a promising direction for next-generation infection monitoring [70,240,241]. However, adoption remains contingent on addressing challenges related to data privacy, algorithmic bias, and the accuracy of device-based measurements [121]. Despite their transformative potential, AI-powered surveillance systems face substantial barriers [160,224]. Most systems generate large volumes of unfiltered data without curation, and their uptake by public health authorities remains limited [224].

3.4.2. Machine learning in diagnostics

Traditional surveillance provides essential early warnings at the population level, yet individual patient care depends fundamentally on precise diagnostics. In this clinical sphere, the integration of machine learning and deep learning has driven significant progress, with AI systems now matching or surpassing the performance of human experts across various diagnostic modalities [39,121,242]. Deep learning architectures, particularly convolutional neural networks (CNNs), have demonstrated exceptional performance in medical image analysis for the detection of infectious diseases [[243], [244], [245], [246]]. Studies have reported diagnostic accuracies of approximately 96% for COVID-19 using CT scans and up to 99% accuracy for distinguishing COVID-19 from pneumonia on chest X-rays [243,244]. A systematic review further found that ResNet-50 was the most frequently used deep learning architecture for tuberculosis detection, appearing in 11 of 47 studies [245]. This transfer learning approach leverages knowledge encoded in models pretrained on large image datasets (such as ImageNet) and adapts it to medical imaging tasks with limited training data, addressing a critical challenge in emerging outbreak scenarios where labelled datasets are necessarily small [243,246] (Fig. 4B).

Beyond image analysis, gradient boosting algorithms, particularly XGBoost, have emerged as powerful tools for integrating heterogeneous clinical and environmental data [39,247]. XGBoost demonstrated the best predictive performance (R² = 0.901, MAE = 0.047) in modelling relationships between emergency room admissions for cardiorespiratory diseases and environmental factors [247]. The algorithm's capacity to handle missing data, capture non-linear relationships, and provide feature-importance rankings makes it particularly well suited to outbreak settings [247] (Fig. 4B).

Multimodal AI systems that integrate diverse data sources (i.e., text, images, audio, video and/or sensor data) consistently outperform unimodal approaches, with an average improvement of 6.2 percentage points in AUC across medical applications [[248], [249], [250]] (Fig. 4B). Notably, efficient deep learning approaches have been developed for Alzheimer's Disease (AD)/mild cognitive impairment (MCI) classification from scalp electroencephalogram (EEG), weighted toward studies reporting three-class (healthy controls vs MCI vs AD) performance and/or explicit computational efficiency [[251], [252], [253]]. These are examples of the efficient use of deep learning networks for classification and diagnosis from bioelectrical signals and non-invasive physiological data. The Holistic AI in Medicine (HAIM) framework demonstrated that models leveraging all possible input combinations of tabular, time-series, text, and image data outperformed single-source approaches by 6-33% across various healthcare tasks including chest pathology diagnosis, length-of-stay prediction, and mortality prediction [249]. However, HAIM uses data in a task-agnostic manner and lacks explainability. In 2026, a new system called xHAIM (Explainable HAIM) was released. It uses Generative AI to help doctors not only predict health outcomes but also understand exactly how the AI made those decisions using a specific four-step process: (1) automatically identifying task-relevant patient data across modalities; (2) generating comprehensive patient summaries; (3) using these summaries for improved predictive modelling; and (4) providing clinical explanations by linking predictions to patient-specific medical knowledge. Evaluated on the HAIM-MIMIC-MM dataset, xHAIM improves the average AUC from 79.9% to 91.3% across chest pathology and operative tasks. By transforming AI from a “black box” predictor into an explainable decision-support system, xHAIM enables clinicians to trace model outputs back to relevant patient data, thereby bridging AI advances with clinical utility [254].

Environmental and climatic data have become increasingly important for understanding and predicting infectious disease patterns [39,121] (Fig. 4B). Meteorological and transportation data analysed using machine-learning models have predicted DENV incidence in Vietnam with high accuracy, while transportation data incorporated into spatial models have been used to estimate Mpox spread risk during the 2024 outbreak [121]. Although these predictive frameworks facilitate the rapid implementation of containment strategies, their effectiveness is often hindered by persistent obstacles, including limited generalisability, inherent algorithmic bias, and concerns regarding data quality [121]. Translating these technological successes into global health impact, however, presents distinct operational hurdles. AI-powered diagnostics face challenges in resource-limited settings where they could potentially provide the greatest benefit [121,[255], [256], [257], [258]]. The absence of electronic health records and the infrequent use of standardised laboratory information management systems constrain the development of robust AI models [121]. Fragmented IT systems, inconsistent adoption of digital health tools, limited data availability, and difficulties integrating new technologies into existing healthcare infrastructure hinder widespread adoption [121,255,258].

Despite these barriers, implementations such as the Neotree app for neonatal diagnostics and neural network-based image-processing tools for malaria detection have been successfully deployed in hospitals in Malawi and Zimbabwe. These deployments required addressing costs related to staff training, infrastructure, and data hosting, but ultimately enabled reliable diagnostics without the need for specialised local expertise [121]. Furthermore, clinical integration is impeded by algorithmic opacity. A critical challenge for AI diagnostics relates to interpretability and explainability [121]. Many machine learning algorithms function as “black boxes”, hindering clinical acceptance, particularly in high-risk scenarios where incorrect predictions could have profound consequences [121]. Explainable AI techniques, including SHAP (Shapley Additive exPlanations), could enhance transparency and clinician confidence [121] (Fig. 4B).

However, substantial challenges to widespread adoption remain. In a study evaluating 140 radiologists using an AI assistant to interpret chest X-rays, lower-performing radiologists did not consistently derive greater benefit from AI support, and AI errors tended to influence treatment decisions and adversely affect radiologist performance across many of the scenarios tested [121]. These findings highlight significant risks when advanced systems are implemented without robust regulatory oversight and comprehensive technical education for end-users [121].

AI models can rapidly analyse blood test results alongside chest X-ray findings to detect infections or complications, thereby supporting clinicians in making informed decisions about the initiation, adjustment, or cessation of antimicrobial therapy [121]. However, only about 5% of healthcare organizations have successfully deployed AI solutions in routine clinical practice, revealing a substantial gap between the promise of research and real-world implementation [270]. Current evaluations of large language models (LLMs) in healthcare are fragmented and insufficient; robust assessment will require the use of real patient data, explicit quantification of biases, coverage of a broader range of tasks and specialties, and reporting of standardised performance metrics to enable safe and scalable deployment [270].

Table 4 summarises their role in virus surveillance, while Fig. 4 illustrates how they scale from broad outbreak forecasting down to patient-level testing.

Table 4.

AI Predictive Tools for Outbreak Forecasting, integrating genomic surveillance.

Key Theme (Subsection)	Main Concepts and Models	Advantages	Potential Evolution	Cited References
Global Surveillance and NLP (3.4.1).	Natural Language Processing (NLP); Bayesian algorithms; HealthMap; EpiWatch; BODA; Digital footprint analysis (Google Trends/social media).	Real-time anomaly detection; identification of threats weeks before official reports; probabilistic forecasting with sparse data.	Integration with wearable biosensors for subclinical detection; reduction of media-driven noise; improved global uptake by public health authorities.	[39,159,160,170,[224], [225], [226], [227], [228], [229], [230], [231], [232], [233], [234], [235], [236], [237], [238], [239]]
Machine Learning in Diagnostics (3.4.2).	CNNs (ResNet-50); Gradient Boosting (XGBoost); Multimodal AI (HAIM/xHAIM); Transfer Learning.	Diagnostic accuracy matching or exceeding human experts (e.g. 99% in chest X-rays); effective handling of missing or heterogeneous data.	Deployment of Explainable AI (xHAIM) for clinician trust; standardisation of multimodal clinical workflows; adaptation to resource-constrained settings.	[39,121,[224], [225], [226], [227], [228], [229], [230], [231], [232], [233], [234], [235], [236], [237], [238], [239], [240], [241], [242], [243], [244], [245], [246], [247], [248], [249], [250],[254], [255], [256], [257], [258],264,[270], [271], [272]]

3.4.3. Hybrid AI frameworks and deployment safeguards in real-time epidemic surveillance

Beyond purely data-driven diagnostics, a rapidly maturing body of work has developed formal mathematical frameworks for coupling mechanistic epidemiological models with machine learning, moving beyond ad hoc integration toward principled hybrid architectures. A scoping review of 245 studies identified four dominant coupling paradigms, model parameterisation, surrogate modelling, data augmentation, and hybrid forecasting, each addressing distinct limitations of standalone approaches [273]. Among the most promising formalisms are physics-informed neural networks (PINNs), which embed compartmental ODE systems (e.g., SEIR/SEIRD) directly into the neural network loss function, thereby simultaneously assimilating epidemiological theory and observational data during training [274,275]. PINNs have demonstrated superior forecasting accuracy over purely data-driven sequence models (RNNs, LSTMs, Transformers) for COVID-19 case, hospitalisation, and mortality predictions, while also enabling the estimation of time-varying epidemiological parameters such as transmission rates [274,276]. A complementary approach involves neural ordinary differential equations (Neural-ODEs), which replace fixed ODE right-hand sides with learnable neural network functions, allowing the model to discover latent epidemic dynamics directly from data. When combined with Bayesian optimisation, Neural-ODE frameworks have accurately captured multi-wave infection dynamics and predicted epidemic peaks with significant lead time for public health response [277]. Data assimilation techniques, such as ensemble Kalman filtering (EnKF) integrated with machine learning classifiers, offer another formal coupling strategy by continuously correcting mechanistic model states with incoming surveillance data, substantially improving real-time forecasting accuracy relative to traditional compartmental models alone [278]. At the spatial scale, graph neural networks (GNNs) encode geographic regions as graph nodes and mobility or adjacency relationships as edges, learning spatio-temporal transmission patterns that conventional compartmental models cannot capture [279,280]. GNN-based surrogates of mechanistic metapopulation models have achieved up to 28,670-fold computational acceleration while maintaining acceptable forecasting accuracy, enabling real-time scenario evaluation for pandemic decision support [281]. Furthermore, the synergistic integration of machine learning with agent-based models (ABMs) has been systematically reviewed across four operational scenarios: micro-agent situational awareness learning, behavioural intervention, macro-level emulation, and sequential decision-making, with reinforcement learning enabling adaptive policy optimisation within ABM frameworks [282]. Despite these advances, key gaps persist, including the limited incorporation of realistic decision-making considerations, insufficient exploration of socio-behavioural mechanisms, and the need for robust validation against diverse datasets beyond COVID-19 [273].

Importantly, several concrete case studies now quantify the incremental gains of mechanistic-machine learning coupling over purely empirical or purely mechanistic baselines in operational, real-time settings. The clearest example is epimodulation, a hybrid approach that embeds fundamental epidemiological principles into pre-existing predictive models: when applied to a panel of empirical and machine learning forecasters (ARIMA, Holt-Winters, gradient-boosting machines, Prophet, and spline models), it improved overall prediction accuracy by an average of 12.3% (range 8.5-18.7%) for COVID-19 hospital admissions and 32.9% (range 24.2-43.7%) for influenza hospital admissions, with substantially larger gains around epidemic peaks (27.9% and 43.8%, respectively), precisely the moments when surge-capacity decisions are most consequential; notably, epimodulation also improved the performance of the COVID-19 Forecast Hub ensemble itself [283]. A second operational example is the real-time forecasting of COVID-19 across Chinese provinces, in which an interpretable machine learning layer fused official case counts, internet search activity, and news media signals with daily outputs from a metapopulation mechanistic model, outperforming a collection of baseline models in 27 of 32 provinces at a 2-day horizon during the earliest outbreak phase, when historical training data were minimal [284]. Notably, the evidence base remains dominated by COVID-19, is largely evaluated using point-error metrics (MAE, MAPE, RMSE) rather than proper probabilistic scoring rules or formal decision-utility criteria, and comparability across studies is limited by heterogeneous baselines and evaluation windows [274,283].

In this context, it is important to consider that deploying streaming machine learning models in real-time epidemiological monitoring without formal guarantees for uncertainty quantification (UQ), dynamic calibration, parameter identifiability, and concept drift mitigation could pose significant risks to operational reliability and clinical trust.

A systematic review of UQ techniques applied to AI in healthcare over the past decade found that Bayesian methods, including Monte Carlo dropout, deep ensembles, and variational inference, were the predominant approach for both machine learning and deep learning models, with applications concentrated in medical imaging and comparatively few studies addressing physiological signals or streaming surveillance data [285]. Objective evaluation of UQ methods for COVID-19 detection from chest X-rays has demonstrated that predictive uncertainty estimates are statistically higher for erroneous than correct predictions, confirming that ensemble-based UQ can effectively flag high-risk outputs for human review [286]. Complementing Bayesian approaches, conformal prediction (CP) provides distribution-free, finite-sample coverage guarantees by producing prediction sets, rather than point estimates, that contain the true label with a user-specified probability; CP has been shown to reduce diagnostic errors from 2% to 0.1% in histopathological cancer grading while flagging 22% of predictions as unreliable for expert adjudication [287]. Importantly, recent work has extended CP to clinical triage under distribution shift, combining importance-weighted conformal sets with cost-aware deferral to maintain coverage guarantees even when the deployed data distribution diverges from training conditions [288]. Moreover, to ensure operational safety and equitable resource allocation in disease forecasting and clinical triage, prediction pipelines must incorporate robust UQ and post-hoc calibration. Distribution-free approaches such as CP provide finite-sample coverage guarantees, generating reliable prediction sets rather than point estimates during high-uncertainty surveillance scenarios [288]. Furthermore, to prevent algorithmic bias across underserved subpopulations, group-conditioned calibration and subgroup conformal prediction must be implemented [289]. These techniques could substantially narrow coverage and error-rate disparities across demographic cohorts, preventing systemic under-triage or resource misallocation in vulnerable populations [290].

Beyond prediction-level uncertainty, the calibration and identifiability of the mechanistic parameters embedded within hybrid AI-epidemiological models require rigorous treatment. Structural identifiability analysis using differential algebra approaches can reveal parameter correlations in compartmental epidemic models (e.g., SEIR) that preclude unique estimation from observed states, and remediation strategies include incorporating additional observations, fixing correlated parameters via prior information, or model reformulation [291]. When likelihood functions are analytically intractable, simulation-based inference methods, including Approximate Bayesian Computation (ABC), Neural Posterior Estimation (NPE), and Preconditioned NPE, offer alternative posterior approximation, with neural methods generally improving posterior fidelity and uncertainty calibration over ABC under constrained simulation budgets, albeit with greater computational cost [292]. Dynamic recalibration via ABC applied to large-scale agent-based COVID-19 models has demonstrated that incorporating the most recent surveillance data in pseudo-real time substantially improves forecast accuracy and reduces predictive uncertainty in later simulation windows [293].

A further critical challenge for streaming machine learning systems is concept drift, the phenomenon whereby the statistical relationship between input features and clinical outcomes shifts over time due to evolving pathogen characteristics, changing diagnostic criteria, or altered clinical protocols [294]. AI models in infectious disease settings are subject to at least three forms of distributional shift: data drift (changes in input feature distributions), concept drift (changes in input–output relationships), and label drift (changes in outcome prevalence), each of which can silently degrade predictive performance if unmonitored [39]. A systematic review of strategies for detecting and mitigating dataset shift in health machine learning identified temporal shift and concept drift as the most commonly addressed types, with model-based monitoring and statistical tests as primary detection strategies and periodic retraining as the predominant correction approach, though no single method has yet demonstrated broad generalisability across clinical use cases [295]. Explainable Machine Learning (SHAP-based feature importance tracking) has been shown to provide a complementary drift detection mechanism: monitoring changes in feature-level SHAP values relative to global importance during the COVID-19 pandemic identified both the timing of distributional shifts and emergent clinical risk factors simultaneously [296]. For operationalising drift remediation, a proactive label-agnostic monitoring pipeline evaluated across 143,049 patients and seven hospitals demonstrated that drift-triggered continual learning, using automated maximum mean discrepancy testing to detect shifts and selectively retraining on recent data, significantly improved mortality prediction performance during the COVID-19 pandemic compared to maintaining a static model [297]. Lifelong machine learning frameworks that simultaneously monitor model calibration and the underlying distribution of disease risk characteristics offer a promising approach for maintaining prediction accuracy through iterative “development-deployment-maintenance-monitoring” life cycles, effectively addressing both gradual and sudden data drifts while providing explanations for performance degradation [298]. Operationally, however, the field still lacks consensus alert thresholds: drift monitoring in deployed clinical AI relies on label-agnostic statistical triggers (e.g., maximum mean discrepancy tests at prespecified significance levels) rather than on validated performance-decay cut-offs, since ground-truth labels arrive too late for timely detection [297,299]. Recommended practice is therefore a tiered monitoring scheme combining continuous data- and model-drift surveillance, root-cause analysis, and predefined correction pathways ranging from recalibration and test-time adaptation to full retraining [299], with thresholds calibrated to the intervention's cost asymmetry rather than to statistical significance alone. In epidemic early-warning settings specifically, anomaly-detection thresholds must be benchmarked against gold-standard reference series, as sensitivity-timeliness trade-offs differ markedly across pathogens and data streams [300].

Together, these methodological safeguards, encompassing Bayesian and conformal UQ, structural identifiability verification, simulation-based calibration, and automated drift detection with continual learning, constitute essential prerequisites for the trustworthy deployment of streaming AI systems in real-time infectious disease surveillance and diagnostics.

Finally, preparedness for emerging viruses requires addressing fundamental data constraints. The development of AI diagnostics for emerging pathogens faces the fundamental challenge that training data are necessarily limited during early outbreak phases [246,271,272]. Using techniques such as deep learning and natural language processing (NLP) to extract features from unstructured data represents a powerful strategy for enhancing the performance of diagnostic models when structured data are limited [271]. The research agenda for AI diagnostics in LMICs should be driven by local needs, health system constraints, and disease burden rather than availability of data and funding, with steps taken during development to avoid ethnic, socioeconomic, and gender biases found in some AI applications [272].

These AI architectures bridge the gap between public health and clinical care, representing key actions of the “Silicon Shield”.

4. Discussion: methodological challenges and clinical validation

4.1. Overcoming algorithmic bias

Addressing the opacity, “black box” nature, of AI algorithms and the necessity of independent, external validation to establish true biological security.

The deployment of AI in public health confronts fundamental challenges related to algorithmic bias and model opacity, both of which must be addressed to ensure clinical trustworthiness. Biases frequently arise from algorithms trained predominantly on data from high-income countries, leading to underrepresentation of diverse populations and systematic errors that compromise equity [39,121].

The “black box” nature of deep neural networks poses a distinctive challenge for clinical applications where understanding the rationale behind predictions is essential [39,301,302]. This opacity poses severe risks for clinical accountability, particularly as large language models increasingly provide frontline medical information that may contain unverified advice [39].

Explainable AI (XAI) techniques (e.g., SHAP, LIME) attempt to unveil model reasoning by making AI outputs interpretable to humans, thereby supporting adoption, regulatory compliance, and bias mitigation [301,303]. Consequently, rigorous external validation on independent datasets is considered the gold standard for establishing true generalisability, complementing rather than replacing XAI approaches [301,[303], [304], [305], [306]]. Although critical, this process is frequently flawed; for instance, a systematic review found 87% of machine learning prediction models exhibited a high risk of bias [304,305,307].

The consequences of such biases are particularly profound in epidemiology. Algorithmic bias in infectious disease applications presents concerns, given the geographic and demographic heterogeneity of pathogen epidemiology. Models developed in high-income settings may severely misclassify patients in LMICs, exacerbating diagnostic disparities [121]. Addressing algorithmic bias requires systematic approaches throughout the AI lifecycle, from data collection through deployment and monitoring. This extends beyond mere technical fixes, encompassing strategies such as building diverse datasets that propel biomedical research forward by setting the stage for widespread adoption of AI to tackle complex biomedical challenges beyond human intuition (e.g., the NIH Common Fund's Bridge to Artificial Intelligence (Bridge2AI) programme), thereby mandating fairness assessments and establishing global governance standards [[308], [309], [310]].

Furthermore, the dynamic nature of AI models introduces additional validation challenges related to distributional shifts over time [39]. AI models cannot be considered static, as they are subject to data drift (changes in input data characteristics), concept drift (changes in relationships between input and output data), and label drift (changes in label distribution) [39]. For infectious disease applications, where pathogen evolution, changing diagnostic criteria, and shifting disease prevalence are common, these temporal validation challenges are particularly acute [39].

4.2. Standardisation in trials: scientific rigour and reporting standards

To navigate these complex validation challenges, the rigorous evaluation of AI interventions through randomised controlled trials (RCTs) requires adherence to established methodological standards. The CONSORT-AI and SPIRIT-AI extensions provide the first international standards for clinical trials of AI systems, mandating clear descriptions of human-AI interaction, algorithm inputs/outputs, and error analyses [[311], [312], [313], [314], [315]]. Current adherence remains concerningly low. Systematic reviews show that machine learning RCTs frequently fail to report extended criteria or demographic inclusivity, limiting reproducibility and exacerbating health disparities [316]. Moreover, a systematic review of deep-learning studies found that, despite severe methodological limitations, including lack of blinding, very small comparator groups, and unavailability of code in more than 90% of studies, the vast majority nevertheless claimed AI performance to be comparable to or better than that of clinicians [317].

Addressing these methodological shortcomings requires dedicated appraisal instruments tailored specifically for AI. The APPRAISE-AI tool addresses a critical gap in existing guidelines by providing a quantitative method for evaluating methodological and reporting quality of AI prediction models for clinical decision support [318]. APPRAISE-AI evaluates studies across six domains: clinical relevance, data quality, methodological conduct, robustness of results, reporting quality, and reproducibility, with 24 items and a maximum overall score of 100 points [318]. An application to a systematic review of machine learning for sepsis estimation found that overall scores ranged from 33 (low quality) to 67 (high quality), with most studies demonstrating moderate quality [318]. The five lowest-scoring items included source of data, sample size calculation, bias assessment, error analysis, and transparency, arguably the most important characteristics for determining scientific rigour and generalisability [318]. A review of 46 critical appraisal tools highlighted a frequent mismatch between AI-specific evaluation criteria, particularly algorithmic fairness, and traditional epidemiological appraisal frameworks centred on study design [[318], [319], [320]].

Beyond initial validation, sustained oversight is required to ensure long-term clinical safety and efficacy. The integration of multiple evaluation and monitoring approaches could provide “algorithmovigilance” akin to pharmacovigilance [321,322]. Critically, monitoring standards proposed thus far focus on safety, often failing to specify how it will be determined across different settings and over time [322]. Furthermore, they tend to neglect how long-term clinical effectiveness will be determined across different settings and over time [[323], [324], [325]]. Establishing long-term AI safety will require a shift toward more rigorous evaluation. Healthcare providers must be equipped with the resources to conduct causal-effect studies, such as stepped-wedge trials, to verify that AI tools remain effective and safe throughout their implementation [322]. Scientific rigour must withstand the urgencies of a crisis. During public health emergencies, the immense pressure to deploy innovative tools rapidly must not supersede evidence-based appraisal [311]. Strict adherence to frameworks like SPIRIT-AI and CONSORT-AI is imperative to efficiently identify and commission only the safest, most effective AI interventions for the public [311].

Ultimately, embedding “algorithmovigilance” as a core pillar could be essential to ensuring the sustainable, long-term utility of these systems for public health [322,[326], [327], [328]]. These data and developments are summarized in Table 5, which outlines the challenges AI models must address to achieve independent external validation and to support biological security during viral outbreaks.

Table 5.

Methodological challenges of AI.

Key Theme (Subsection)	Main Concepts and Models	Advantages	Potential Evolution	Cited References
Overcoming Algorithmic Bias (4.1).	Algorithmic bias; model opacity (“black box”); representative datasets; external validation.	Promotion of equitable healthcare; reduction of systematic errors for underrepresented populations; increased clinical trustworthiness.	Integration of fairness metrics into regulatory approval; mandatory subgroup performance reporting across LMICs.	[39,121,[301], [302], [303], [304], [305], [306], [307], [308], [309], [310]]
Explainability and Interpretability (4.1).	Explainable AI (XAI); SHAP; LIME; feature importance; clinician-AI interface.	Enhanced accountability for high-stakes decisions; unveils model reasoning for clinicians; aids in identifying data artefacts or “shortcut learning".	Development of “human-in-the-loop” systems; standardisation of interpretability reports for clinical auditing.	[39,301,303]
Standardisation in Trials: Scientific Rigour and Reporting Standards (4.2).	SPIRIT-AI; CONSORT-AI; APPRAISE-AI; Cochrane Risk of Bias; External validation.	Systematic identification of reproducibility gaps; transparent appraisal of AI interventions; ensures evidence-based deployment.	Implementation of “algorithmovigilance” (continuous post-deployment monitoring); prospective RCTs with diverse global cohorts.	[[311], [312], [313], [314], [315], [316], [317], [318], [319], [320], [321], [322], [323], [324], [325]]

5. Perspectives: towards an “AI-first” Paradigm and the “Silicon Shield” architecture

Envisioning a globally integrated “Silicon Shield” that maintains continuous, real-time epidemiological surveillance.

The convergence of mathematical modelling and AI has catalysed a paradigm shift towards envisioning a globally integrated “Silicon Shield” capable of continuous, real-time surveillance and rapid response [170,329,330]. Just as the biological immune system maintains constant vigilance against pathogens through distributed sentinel cells, pattern recognition receptors, and coordinated effector responses, a mathematical and AI-based surveillance and decision-support system would integrate distributed data sources, AI-powered pattern recognition, and coordinated public health responses into a unified global defence network [170]. This architecture seamlessly integrates vast, heterogeneous data streams (clinical, genomic, environmental, and mobility) with AI analytics (NLP and deep learning) to generate actionable predictive forecasts and localised decision support [39,129,170,273,329,330].

The necessity for such an integrated system was starkly illuminated when the COVID-19 pandemic exposed critical gaps: inadequate early warning systems, insufficient real-time data on resource needs, and limitations of traditional epidemiological models in capturing complex disease dynamics [331]. AI applications were deployed with varying success, highlighting the need to systematically embed hybrid approaches, combining machine learning's data-mining capabilities with the interpretable frameworks of traditional mechanistic modelling, into public health infrastructure [331,332].

The integration of AI with mechanistic epidemiological modelling represents a particularly promising direction [129,273]. Although traditional compartmental models provide interpretable frameworks grounded in epidemiological theory, embedding this domain knowledge within AI architectures enables hybrid approaches that can potentially combine the strengths of both paradigms [129,273]. Hence, theoretical framework spanning pathogen characterization, epidemiological surveillance, and clinical AI has been proposed [39]. Furthermore, recognising the ecological origins of these threats is paramount. Since 75% of emerging infectious diseases are zoonotic, this system must operate under the One Health paradigm, by utilising AI to enhance capacity to monitor, diagnose, and predict cross-sectoral diseases across these interconnected domains [39,[333], [334], [335], [336]], representing layers of the “Silicon Shield”.

The establishment of dedicated institutional infrastructure represents a critical enabler for this vision [72,337]. The US CDC's Centre for Forecasting and Outbreak Analytics, the WHO Hub for Pandemic and Epidemic Intelligence, and the EU's 2021-2027 surveillance framework exemplify emerging institutional models designed to operationalise this architecture [72,337,338].

However, the ‘Silicon Shield’ faces substantial barriers in transitioning from theoretical frameworks to global implementation. These include incompatible “Information Technology” (IT) systems, complex legal environments for data sharing, limited AI literacy in the healthcare workforce, and chronic underfunding in low- and middle-income countries (LMICs) where most spillovers originate [328,334,[339], [340], [341], [342], [343]]. Addressing these barriers will require coordinated action across multiple domains [272,328,342]. Policymakers must invest in infrastructure and workforce development [328], while public health institutions must prioritise equity in AI design and implementation and promote ethical AI use through clear guidelines aligned with human rights principles [328]. International collaboration is essential for establishing global standards, enabling cross-border data sharing, and ensuring that surveillance capabilities are distributed equitably rather than concentrated in high-income countries [272,342]. Furthermore, the research agenda should be driven by local needs and disease burden, with deliberate steps to avoid biases [272].

In this context, and to transition the “Silicon Shield” from a theoretical framework into an operational global defence system against emerging viruses and associated syndemic crisis, we propose a “Modular Reference Architecture” structured across three functional tiers:

• (i)Data Pipes & Ingestion Layer

Continuous data ingestion is achieved through cloud-native ETL (Extract, Transform, Load) pipelines designed to aggregate heterogeneous data streams in real time. This includes (a) molecular streams via mNGS bioinformatic pipelines [36,344,345]; (b) clinical streams via standardized FHIR (Fast Healthcare Interoperability Resources) APIs extracting anonymized EHR data [346]; and (c) digital intelligence streams leveraging natural language processing (NLP) engines (e.g., EpiWatch, HealthMap) alongside environmental and mobility datasets [170,224,227].

• (ii)Model Pipelines & Workflow Layer

The computational core operates via a multi-stage pipeline:

• (iii)Stage I - Early Signal & Anomaly Detection: Raw streams feed into unsupervised neural networks (e.g., TCINet, ViraMiner) to flag atypical genomic or syndromic anomalies [204,217], before official clinical reporting.
• (iii)Stage II - Evolutionary & Immune Escape Forecasting: Deep learning models (e.g., EVEscape, pLMs) evaluate variant fitness landscapes and predict structural immune escape [221,218,347,348].
• (iii)Stage III - Transmission & Intervention Modelling: Mechanistic epidemiological frameworks (SEIR ODEs) coupled with high-resolution agent-based simulators (e.g., Covasim) [181,349] run dynamic counterfactual scenarios to evaluate targeted NPIs and vaccination strategies under real-world constraints.
• (iii)Decision Interfaces & Public Health Integration

The output tier translates complex model outputs into standardized RESTful APIs feeding interactive, role-specific interfaces. For public health decision-makers, dashboards map hyper-local risk indices, effective reproduction rates (R_t), and supply-chain vulnerabilities. For frontline clinicians, integrated multimodal tools paired with Explainable AI (e.g., xHAIM, SHAP) convert complex computational analyses into reliable, transparent decision support [254,350].

Ultimately, technological deployment must be coupled with societal engagement. AI integration in public health represents a major socio-technical transformation, requiring balanced approaches that embed ethical foresight into design, promote equitable governance, and address global evidence disparities [342]. Without deliberate engagement, AI risks reinforcing existing inequities. Practical steps include embedding AI training in public health education, building multidisciplinary teams, investing in data infrastructure, and ensuring participatory approaches that engage communities in surveillance system design and governance.

Critically, incorporating social determinants of health (SDH) into the “Silicon Shield” architecture must go beyond documenting associations to enable causal interpretations that can reliably inform intervention. A systematic review of infectious disease modelling guidelines found that while 14 of 19 guidelines referenced at least one social determinant, recommendations focused predominantly on routine demographic indicators (age, sex, socioeconomic status), with limited attention to the causal structures linking structural disadvantage to differential disease burden [351]. This gap is consequential: analysing aggregate-level social variables without individual-level causal reasoning risks ecological fallacy, wrongly attributing group-level associations to individual-level effects, and collider bias, whereby conditioning on a common effect of both exposure and outcome (e.g., restricting analyses to hospitalised patients) generates spurious associations between SDH and disease outcomes [[352], [353], [354]].

To mitigate these biases, three complementary methodological strategies could be integrated into the modelling pipeline. First, directed acyclic graphs (DAGs) should be constructed a priori to make explicit the hypothesised causal pathways from structural disadvantage (e.g., poverty, housing density, occupational exposure) through intermediate mechanisms (pathogen exposure, susceptibility, healthcare access) to infection outcomes [25,355]. DAGs allow systematic identification of confounders requiring adjustment, mediators that should not be conditioned upon, and colliders whose inadvertent conditioning would introduce bias [353,356]. Second, doubly robust estimation methods, particularly targeted maximum likelihood estimation (TMLE) combined with ensemble machine learning (Super Learner), offer a principled framework for estimating causal effects of SDH from observational surveillance data, providing consistent estimates even when either the outcome or exposure model is misspecified [[357], [358], [359]]. TMLE has been increasingly adopted across epidemiological disciplines precisely because it accommodates the complex, high-dimensional confounding structures typical of SDH-disease relationships [358,359]. Third, agent-based models (ABMs) provide a natural platform for counterfactual causal analysis of SDH, as they can encode individual-level social characteristics, simulate interference (where one person's exposure affects another's outcome), and evaluate the population-level consequences of structural interventions that cannot be randomised [360]. Embedding the target trial framework within ABM design further strengthens causal interpretability by making explicit the hypothetical intervention, eligibility criteria, and assignment mechanism [361]. Additionally, multilevel modelling approaches that simultaneously estimate individual- and area-level effects can help disentangle compositional from contextual SDH effects, reducing ecological bias inherent in purely aggregate analyses [362,363]. Hierarchical causal diagrams that encode the multilevel data-generating mechanism have recently been proposed to formally quantify the ecological fallacy within a causal framework, demonstrating that individual-level data are essential for valid causal assessment [363]. A scoping review of 89 studies incorporating social vulnerability into infectious disease models found that while compartmental and agent-based approaches have begun stratifying populations by SDH, holistic integration remains limited, with cultural and behavioural determinants largely excluded and data source transparency insufficient [364]. Together, these DAG-guided, doubly robust, and simulation-based causal inference strategies would enable the “Silicon Shield” to move beyond correlational surveillance toward actionable, equity-informed causal intelligence for pandemic response. We may point out that implementing this framework always requires privacy-preserving data linkage across public health and social service registries.

Finally, the public health community is both uniquely positioned and ethically obligated to shape AI development towards equitable and trustworthy outcomes rather than passively accepting technological trajectories determined by other actors [343].

6. Conclusions

The synthesis presented throughout this review demonstrates that the convergence of mathematical modelling and artificial intelligence has fundamentally transformed the capacity to understand, predict, and respond to viral syndemics, establishing an essential, evidence-driven framework for mitigating the holistic burden of future pandemics [39,129,273]. From the foundational compartmental models that quantify transmission dynamics and intervention effects, through the deep learning architectures that accelerate pathogen discovery and predict immune escape, to the integrated surveillance systems that detect outbreaks before official recognition, this technological convergence provides unprecedented tools for pandemic preparedness and response [39,129,170].

The evidence base supporting AI applications in infectious disease prevention and management has matured substantially, spanning clinical care and patient support, infection prevention and control, epidemiology and surveillance, pathogen characterisation, and drug development [39]. The integration of AI with mechanistic epidemiological models offers transformative potential by combining the data-mining capabilities of machine learning with the interpretable, theory-grounded frameworks of traditional modelling [129,273]. Notwithstanding these advances, several critical hurdles remain. Although AI holds immense potential, global uptake is hampered by algorithmic bias, “black-box” opacity, and weak clinical validation. This challenge is further compounded by the fact that development is heavily concentrated in wealthy nations, risking the marginalisation of vulnerable populations and the creation of tools that are poorly adapted to regions bearing the highest disease burdens [39,121,272,302,334,341].

Addressing these challenges requires sustained commitment: policymakers must establish stringent regulatory frameworks; scientists must prioritise bias mitigation and the development of hybrid models; and healthcare professionals must cultivate computational literacy [39,272,273,328,342,343]. Looking forward, the integration of AI in infectious disease prevention and management could advance care and help achieve broader societal goals, including health equity and health systems sustainability [39]. However, realising this potential requires that AI development be guided by public health values and priorities rather than purely technological or commercial imperatives [328,342,343].

The technological foundations for such a system are increasingly in place: AI algorithms capable of processing vast, heterogeneous data streams; genomic sequencing technologies that enable rapid pathogen characterisation; digital platforms connecting billions of individuals; and computational infrastructure that supports real-time analytics at a global scale [170,330,333]. What remains is the more difficult work of building institutional capacity, establishing robust governance frameworks, ensuring equitable access, and cultivating the human expertise necessary to translate technological capability into sustained public health impact [328,339,340]. The COVID-19 pandemic served both as a devastating demonstration of global vulnerability to viral threats and as an unprecedented catalyst for innovation in surveillance, modelling, and response [72,330,331].

Ultimately, the proposed “Silicon Shield” constitutes not merely a technological advancement but an essential instrument for evidence-driven pandemic preparedness. Its full potential will be realised only through deliberate, sustained, and equitable investment in the infrastructure, institutions, and human capacity necessary to translate computational capability into public health protection. The tools are increasingly available; the imperative now is to deploy them wisely, equitably, and at scale.

CRediT authorship contribution statement

Romina Cabrera-Rodríguez: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Iriome Reyes-Castañeda: Writing – review & editing, Writing – original draft. Iria Lorenzo-Sánchez: Writing – review & editing, Writing – original draft. Agustin Valenzuela-Fernández: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Rodrigo Trujillo-González: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Resources, Methodology, Investigation, Funding acquisition, Formal analysis, Conceptualization.

Ethical approval statement

It is a systematic review of previously published literature and did not involve any new research with human participants or animals.

Availability of data and materials

Data sharing does not apply to this article as no novel data was generated during the current study.

Funding

This research and A.V.-F.’s laboratory are supported by Grant PID2024-155444OB-I00 funded by MICIU/AEI/10.13039/501100011033 and by “ERDF A way of making Europe” (MICIU/AEI, Spain), Grant PID2021-123031OB-I00 funded by MICIU/AEI/10.13039/501100011033 and by “ERDF A way of making Europe” (MICIU/AEI, Spain), UNLL10-3E-783 (ERDF and “Fundación CajaCanarias”) and “SEGAI-ULL”. I.L-S. was founded by Grant PID2021-123031OB-I00 (MICIU/AEI/10.13039/501100011033, Spain, and ‘ERDF A way of making Europe’) and is now founded by “Ayuda formación de investigadores del Gobierno de Canarias, cofinanciada por la Agencia Canaria de Investigación, Innovación y Sociedad de la Información (ACIISI), de la Consejería de Universidades, Ciencia e Innovación y Cultura, y por el Fondo Social Europeo Plus (FSE+), Programa Operativo Integrado de Canarias 2021-2027 (Exp.: FPI2024010073)”.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.