Performance and clinical utility of image-based machine learning models for the diagnosis of neglected tropical diseases in low- and middle-income countries: a systematic review

Abstract

Background

Neglected tropical diseases (NTDs) disproportionately affect populations in low- and middle-income countries (LMICs), where diagnostic capacity is often limited. Image-based machine learning (ML) has emerged as a potential tool to support diagnosis, but its clinical utility remains unclear.

Aim

To systematically evaluate the diagnostic performance and clinical utility of image-based machine learning models for NTD diagnosis in LMICs.

Methods

This review was registered on PROSPERO and conducted in accordance with PRISMA guidelines. Searches were performed in PubMed/MEDLINE, Embase, Scopus, Web of Science, and IEEE Xplore from January 2010 up until January 31st, 2026. Peer-reviewed primary studies evaluating image-based ML models for NTD diagnosis in LMICs and reporting diagnostic performance metrics were included. Risk of bias was assessed using QUADAS-2, and findings were synthesised narratively.

Results

Eight studies met the inclusion criteria. Microscopy-based ML models demonstrated consistently high performance, with reported sensitivities and specificities frequently above 90%, particularly for malaria and helminth infections. Clinical image–based models for skin NTDs showed more variable accuracy. External validation and implementation evaluation were inconsistently reported, limiting generalisability and clinical applicability.

Conclusion

Image-based ML models show strong diagnostic potential for NTDs in LMICs, especially in microscopy-supported workflows. However, translation into routine practice is constrained by limited dataset representativeness, inadequate external validation, and insufficient attention to operational feasibility and explainability. Future research must prioritise implementation-oriented evaluation to realise public health impact.

Clinical trial number

Not applicable.

PROSPERO

CRD420261339435.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-13784-8.

Introduction

Neglected tropical diseases (NTDs) represent a group of communicable diseases that disproportionately affect impoverished populations in low- and middle-income countries (LMICs), contributing to significant morbidity, disability, and mortality worldwide [1]. Despite global initiatives such as the World Health Organization’s 2030 NTD roadmap, which estimates that over 1 billion people remain at risk, many NTDs persist due to inadequate diagnostic infrastructure, limited access to laboratory services, and scarce clinical expertise in endemic regions [2, 3]. Traditional diagnostic methods, including microscopy and serological assays, are often resource-intensive, operator-dependent, and slow, thereby constraining timely disease detection and effective treatment delivery [1]. In parallel, advances in artificial intelligence (AI) and machine learning (ML), especially convolutional neural networks (CNNs) for image analysis, have demonstrated remarkable accuracy in medical imaging tasks such as malaria parasite detection and helminth egg classification, with reported sensitivities and specificities exceeding 90% in some controlled studies [4]. Machine learning encompasses computational models capable of learning patterns from data to make predictions or classifications, where performance is typically quantified by metrics such as accuracy, sensitivity, specificity, and F-score [5]. Image-based ML models, particularly deep learning architectures, have further shown potential to detect fine visual patterns in clinical and microscopic images that are challenging for human observers [3].

Despite these promising developments, the translation of image-based ML diagnostics from research prototypes to clinical and public health practice in LMICs remains poorly understood [3]. Existing reviews on ML applications in tropical disease diagnosis have largely focused on algorithmic performance for diseases such as dengue or malaria using non-imaging data or general ML applications rather than real-world image-based systems [6, 7]. Moreover, much of the available literature on AI for NTD diagnosis has been conducted in controlled laboratory environments or high-income settings, with limited evaluation in endemic LMIC contexts where conditions such as variable lighting, limited image standardization, and heterogeneous patient populations may affect performance [1]. For example, a study applying CNN models to skin NTDs reported correct prediction rates above 70%, but highlighted the influence of dataset size and representativeness on diagnostic performance [3]. Similarly, automated helminth egg detection systems have achieved high precision (~ 95%) and specificity (≈ 98%), underscoring technical feasibility but leaving open questions about clinical utility and operational readiness [4]. Importantly, evidence synthesis on how these systems perform across disease categories, imaging modalities, and varied endemic settings remains fragmented, thus impeding clear guidance for clinical deployment and policy adoption.

The rationale for the current systematic review stems from this critical gap: while individual studies suggest that image-based ML models can enhance diagnostic throughput and accuracy in resource-limited environments, there has been no comprehensive, comparative synthesis of performance metrics and clinical utility across the spectrum of NTDs and LMIC implementation contexts. This review aims to rigorously evaluate the diagnostic accuracy, robustness, and translational potential of image-based machine learning models for NTD diagnosis in LMICs, addressing methodological variability, context-specific challenges, and evidence gaps to inform future research and policy directions. The primary aim of this research is, therefore, to systematically assess and synthesize the existing literature on the performance and clinical utility of image-based ML diagnostic models for NTDs in low- and middle-income countries.

Methods

Protocol registration and reporting standards

This systematic review was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO) and available at https://www.crd.york.ac.uk/PROSPERO/view/CRD420261339435. The review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines to ensure transparency, reproducibility, and methodological rigour [8]. All stages of the review process including search strategy development, study selection, data extraction, risk of bias assessment, and synthesis were predefined in the registered protocol to minimise selective reporting and analytic bias.

Search strategy

A comprehensive literature search was conducted across PubMed/MEDLINE, Embase, Scopus, Web of Science, and IEEE Xplore. The search strategy combined controlled vocabulary terms (MeSH) and free-text keywords related to neglected tropical diseases, machine learning approaches, image-based diagnostic techniques, and low- and middle-income country contexts. Boolean operators (AND/OR), truncation symbols, and field tags were applied to optimise retrieval across databases. The search covered from January 2010 to 31 January 2026, which represented the final search date for this review. Reference lists of included studies were also manually screened to identify additional eligible records.

The search strategy combined controlled vocabulary terms (Medical Subject Headings [MeSH]) and free-text keywords related to NTDs, machine learning, artificial intelligence, image-based diagnostics, and LMIC settings. Boolean operators (“AND”, “OR”) were applied to refine and expand the search, while truncation and phrase searching were used where appropriate (Table 1). In addition, the reference lists of all included studies and relevant review articles were manually screened to identify further eligible publications not captured in the database search. The final search strategy was iteratively refined in consultation with methodological guidance for systematic reviews of diagnostic technologies [9]. The full database-specific search strategies, including field tags, Boolean operators, and limits, are provided in Supplementary file- Appendix 1 (Table A1-A5).

Table 1.

Search strategy and key terms used across databases

Search Concept	MeSH Terms and Free-Text Terms	Boolean Structure	Databases Searched
Neglected Tropical Diseases	“Neglected Tropical Diseases“[MeSH] OR NTD*[tiab] OR “tropical infection*“[tiab] OR leishmaniasis[MeSH] OR schistosomiasis[MeSH] OR onchocerciasis[MeSH] OR filariasis[MeSH] OR “soil-transmitted helminth*“[tiab] OR “helminth infection*“[tiab]	OR	PubMed/MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore
Artificial Intelligence / Machine Learning	“Machine Learning“[MeSH] OR “Artificial Intelligence“[MeSH] OR “deep learning“[tiab] OR “neural network*“[tiab] OR “convolutional neural network*“[tiab] OR CNN[tiab] OR “computer vision“[tiab] OR “image classification“[tiab]	OR	All Databases
Image-Based Diagnosis	“Diagnostic Imaging“[MeSH] OR image*[tiab] OR photograph*[tiab] OR microscopy[tiab] OR dermoscopy[tiab] OR “skin image*“[tiab] OR “microscopic image*“[tiab] OR “digital microscopy“[tiab]	OR	All Databases
Diagnostic Performance	sensitivity[tiab] OR specificity[tiab] OR accuracy[tiab] OR AUC[tiab] OR “diagnostic accuracy“[tiab] OR “classification performance“[tiab] OR “performance metric*“[tiab]	OR	All Databases
LMIC Context	“Developing Countries“[MeSH] OR “low- and middle-income countr*“[tiab] OR LMIC*[tiab] OR Africa[tiab] OR Asia[tiab] OR “Latin America“[tiab]	OR	All Databases
Final Combined Search Strategy	(“Neglected Tropical Diseases“[MeSH] OR NTD* OR leishmaniasis OR schistosomiasis OR onchocerciasis OR filariasis OR “soil-transmitted helminths”) AND (“Machine Learning“[MeSH] OR “Artificial Intelligence” OR “deep learning” OR “neural network*” OR “computer vision”) AND (“Diagnostic Imaging“[MeSH] OR image* OR photograph* OR microscopy OR dermoscopy OR “skin image*” OR “microscopic image*”) AND (“Developing Countries“[MeSH] OR LMIC* OR “low- and middle-income countr*” OR Africa OR Asia OR “Latin America”)	AND	All Databases

Eligibility criteria

The eligibility criteria were formulated using the Population–Index test–Comparator–Outcome–Study design (PICOS) framework, which is widely recommended for diagnostic accuracy reviews [10].

Inclusion criteria were as follows:

(i) studies involving human populations in LMICs as defined by the World Bank; (ii) studies evaluating image-based machine learning or deep learning models for the diagnosis of NTDs using clinical photographs or microscopy images; (iii) studies reporting at least one diagnostic performance metric (e.g. sensitivity, specificity, accuracy, AUC); (iv) peer-reviewed original research articles; and (v) studies published in the English language. Although malaria is not formally classified as a neglected tropical disease by the World Health Organization, it was included as a high-burden comparator infectious disease due to its shared epidemiological context, reliance on microscopy-based diagnosis, and relevance to low- and middle-income country diagnostic workflows. This inclusion enabled broader assessment of image-based machine learning performance across similar diagnostic paradigms.

Exclusion criteria included:

(i) studies using non-image-based data only (e.g. clinical variables, genomics); (ii) simulation-only or algorithm-development papers without diagnostic evaluation; (iii) conference abstracts, editorials, commentaries, and reviews; (iv) animal-only studies; and (v) studies conducted exclusively in high-income countries without relevance to LMIC implementation.

Study selection

All identified records were imported into Zotero reference manager for de-duplication before being exported to Microsoft Excel for screening. Two independent reviewers conducted title and abstract screening, followed by full-text assessment of potentially eligible studies against predefined inclusion criteria. Data extraction and risk-of-bias assessment were also performed independently by two reviewers. Disagreements at any stage were resolved through discussion, and where consensus could not be reached, a third reviewer adjudicated. The study selection process and reasons for exclusion were documented using a PRISMA flow diagram [8].

Data extraction

Data extraction was undertaken using a pre-piloted Excel extraction form after references were transferred from Zotero. Extracted variables included: author(s), year of publication, country and setting, NTD type, image modality (clinical photograph or microscopy), machine learning model architecture, dataset size, reference standard, performance metrics (sensitivity, specificity, accuracy, AUC), validation approach, and reported clinical or public health utility. This structured approach facilitated consistency and reduced extraction errors.

In this review, “clinical utility” was operationally defined as the extent to which image-based machine learning diagnostic systems demonstrate potential usefulness in real-world healthcare settings, particularly within low- and middle-income countries (LMICs). Clinical utility was assessed narratively across several domains reported in the included studies, including workflow integration, turnaround time, usability by non-specialist personnel, device and infrastructure requirements, implementation feasibility in resource-limited environments, potential influence on diagnostic decision-making, and broader public health relevance. These domains were extracted where explicitly reported or inferable from study descriptions to distinguish between purely technical diagnostic performance and evidence suggesting potential applicability in clinical practice. Performance metrics were extracted as reported by the original studies; where available, the unit of analysis (image-level or case-level classification) was recorded.

Risk of bias assessment

Risk of bias and applicability concerns were assessed using the QUADAS-2 tool, which is recommended for diagnostic accuracy studies [9]. QUADAS-2 evaluates four domains: patient selection, index test, reference standard, and flow and timing. Each domain was judged as “low”, “high”, or “unclear” risk of bias based on predefined signalling questions. An overall judgement of study quality was derived narratively rather than through numerical scoring, in line with best-practice guidance.

Data synthesis

Given the heterogeneity in NTD types, imaging modalities, machine learning architectures, and performance metrics, a narrative synthesis approach was adopted. Findings were synthesised thematically by disease category, image type, and diagnostic context, with comparative emphasis on performance metrics and reported clinical utility in LMIC settings. This approach enabled contextual interpretation of results and identification of evidence gaps relevant to real-world deployment.

Results

Overview of the search process

Following a comprehensive search across multiple electronic databases, a total of 771 records were initially identified. After the removal of 114 duplicate records, 657 unique studies were screened at the title and abstract level. Of these, 421 records were excluded for irrelevance to the review objectives. 236 full-text articles were subsequently assessed for eligibility, leading to the exclusion of 227 reports due to reasons including lack of focus on neglected tropical diseases, absence of image-based machine learning approaches, review-only designs, or other methodological limitations. A total of eight studies met the final criteria for this review. Figure 1 presents the PRISMA flow diagram detailing the full study selection process.

Fig. 1.

PRISMA flow diagram showing the study selection process [8]

Characteristics of the included studies

The characteristics of the eight studies included in this systematic review are summarised in Table 2. All included studies were primary diagnostic accuracy investigations employing image-based machine learning models to support the diagnosis of neglected tropical diseases in low- and middle-income countries. The studies varied in disease focus, image modality, and model architecture, encompassing both clinical skin photographs and microscopy images for neglected tropical diseases and related high-burden infectious conditions, including malaria, helminth infections, schistosomiasis, and cutaneous leishmaniasis [3, 11–17]. Evidence relating to workflow integration was reported in 5 of 8 studies, implementation feasibility in 6 studies, task-shifting potential in 4 studies, and explainability features in 1 study.

Table 2.

Characteristics of included studies

Author (Year)	Country / Setting	Study Design	Disease / NTD Type	Image Modality	ML Model / Architecture	Sample Size (Images / Slides)	Reference Standard	Performance Metrics (Key Results)	Validation Approach	Clinical Utility / Notes
Yotsu et al. [3]	Multiple endemic LMIC sites (Africa & Asia)	Cross-sectional diagnostic accuracy	Skin-related NTDs (leprosy, Buruli ulcer, yaws)	Clinical skin photographs	Convolutional Neural Network (CNN)	1,340 images	Expert dermatological diagnosis	Accuracy: 71–76%; Sensitivity: up to 74%; Specificity: up to 80%	Internal validation	Demonstrated feasibility of AI-assisted triage for skin NTDs in low-resource settings
Oyibo et al. [11]	Nigeria (endemic setting)	Observational diagnostic study	Soil-transmitted helminths & Schistosoma mansoni	Stool microscopy images	Deep CNN	7,500 + microscopy images	Expert parasitologist microscopy	Accuracy: 94–98%; Sensitivity: 92–96%; Specificity: 95–98%	Train–test split	High diagnostic precision suggests suitability for laboratory automation
Yang et al. [12]	LMIC field settings	Diagnostic accuracy study	Malaria	Smartphone-captured blood smears	Deep CNN	13,000 + images	Expert microscopy	Accuracy: 94%; AUC: 0.98	External validation	Demonstrated feasibility of smartphone-based diagnostics in rural settings
Purwar et al. [13]	India (malaria-endemic)	Observational diagnostic study	Malaria	Blood smear microscopy	Machine learning classifiers	1,200 images	Expert microscopists	Accuracy: 91–95%	Internal validation	Early evidence of ML outperforming manual screening speed
Mohammed et al. [14]	Endemic laboratory datasets	Diagnostic accuracy study	Schistosoma spp.	Microscopy images	Explainable Deep Learning (CNN + XAI)	4,096 images	Expert egg identification	Accuracy: 93%; Sensitivity: 91%; Specificity: 95%	Cross-validation	Added explainability to support clinical trust and adoption
Sadeghi et al. [15]	LMIC-relevant laboratory dataset (endemic regions)	Experimental diagnostic study	Leishmaniasis	Microscopy (parasite smear images)	Convolutional Neural Networks (ResNet, DenseNet, MobileNet)	Not clearly stated (high-volume microscopy dataset)	Expert parasitological annotation	Accuracy: ~98.95%; Sensitivity: ~100%; AUC: ~99%	Train–test split	Demonstrates highly accurate automated detection of Leishmania parasites, supporting diagnostic efficiency in resource-limited laboratory settings
Abdelmula et al. [16]	Endemic LMIC setting (clinical dermatology context)	Experimental diagnostic study	Cutaneous leishmaniasis	Clinical skin lesion images / smear images	Convolutional Neural Network (CNN)	Not clearly stated	Expert-confirmed diagnosis (microscopy/clinical reference)	High classification performance (reported accuracy > 90%)	Internal validation	Supports AI-assisted identification of cutaneous leishmaniasis lesions, enhancing early detection and triage in low-resource clinical settings
Steyve et al. [17]	Multiple endemic LMIC sites (Africa & Asia)	Observational / experimental ML study	Skin-related NTDs (Buruli ulcer, leprosy, yaws)	Clinical skin photographs	Machine learning-based image classification (CNN + feature extraction methods)	Not clearly stated	Expert dermatological diagnosis	Moderate-to-high diagnostic accuracy across conditions (varies by disease)	Internal validation	Demonstrates feasibility of AI-based triage for multiple skin NTDs, highlighting potential for frontline screening and task-shifting in endemic regions

Risk of bias

The methodological quality of the eight included studies was assessed using the QUADAS-2 tool (Table 3). Overall, the risk of bias varied across studies, with microscopy-based investigations generally demonstrating lower risk in the index test and reference standard domains due to the use of expert-validated ground truth. However, concerns were frequently identified in patient selection and flow/timing domains, largely due to non-random sampling, limited reporting of inclusion procedures, and reliance on retrospective or curated datasets. Additionally, several studies lacked external validation, raising applicability concerns (Fig. 2). These methodological limitations may partly explain variations in reported diagnostic performance and highlight constraints in translating these models into real-world LMIC settings (see Table 3).

Table 3.

QUADAS-2 risk of bias assessment of included studies

Study	Patient Selection	Index Test	Reference Standard	Flow & Timing	Overall Risk of Bias	Quality Rating
Yotsu et al. [3]	Unclear – convenience sampling and limited representativeness of skin lesion images	Low – CNN applied independently of reference standard	Low – expert dermatological diagnosis	Unclear – limited reporting on exclusions and data flow	Moderate	Moderate
Oyibo et al. [11]	Low – well-defined laboratory samples from endemic population	Low – model evaluated independently	Low – expert parasitologist microscopy	Low – complete dataset analysed	Low	High
Yang et al. [12]	Low – field-representative smartphone-acquired samples	Low – CNN applied independently	Low – expert microscopy reference	Low – clear validation workflow including external testing	Low	High
Purwar et al. [13]	Unclear – limited description of sampling strategy and dataset construction	Low – ML classifier applied consistently	Low – expert microscopists	Unclear – incomplete reporting of exclusions	Moderate	Moderate
Mohammed et al. [14]	Unclear – secondary dataset use and unclear sampling process	Low – explainable DL model applied independently	Low – expert parasitological identification	Low – cross-validation applied	Low–Moderate	High
Sadeghi et al. [15]	High – curated dataset with limited variability and unclear population sampling	Low – CNN models applied independently	Low – expert-labelled microscopy dataset	High – absence of external validation	High	Moderate
Abdelmula et al. [16]	Unclear – partial reporting of dataset selection and clinical sampling	Low – CNN applied independently	Low – expert-confirmed diagnosis	Unclear – validation process not fully described	Moderate	Moderate
Steyve et al. [17]	High – restricted dataset with potential selection bias across sites	Low – ML models applied consistently	Unclear – reference standard not fully described	High – lack of external validation and unclear flow	High	Moderate

Fig. 2.

QUADAS-2 risk of bias summary of the included studies

Several themes were discussed in this review which include diagnostic performance which is consistently promising, but varies by disease type and image modality. In addition, dataset quality, representativeness, and external validation remain the main threats to clinical utility. Also, operational feasibility in LMICs is emerging, especially for smartphone and low-resource microscopy workflows. Again, explainability and trust-building features are under-reported but critical for safe deployment.

Diagnostic performance

Across the included evidence base, diagnostic performance was generally promising, but it was not uniform; rather, it tracked closely with disease phenotype and, critically, image modality (Table 4). Microscopy-driven models for malaria and helminthiasis tended to report the most consistently high performance because the visual targets (parasites/eggs) are comparatively standardised objects with recurrent morphology, supporting stronger feature learning and more stable decision boundaries [13–17]. In these studies, performance was typically reported using diagnostic accuracy metrics aligned with clinical decision-making (e.g., sensitivity/specificity, AUC), and the strongest results were observed where workflows mirrored routine microscopy (digitised slides or smartphone-captured smears) and the reference standard was robust (expert microscopy [12, 14, 15]. Importantly, microscopy studies also more readily supported clearer error attribution (e.g., missed eggs/parasites vs. artefacts), which can indirectly strengthen the credibility of performance claims in lab-like settings, even when heterogeneity in staining, lighting, and smear quality persists [11, 15].

Table 4.

Diagnostic performance by disease type and image modality across some included studies

Study	Disease focus	Image modality	Primary approach (as reported)	ML (as in study)	Reported performance indicators (as presented in study)	Interpretation for this theme
Oyibo et al. [11]	STH + S. mansoni	Stool microscopy images	Deep CNN	Deep CNN	High accuracy with strong sensitivity/specificity (reported)	High performance consistent with standardised targets; supports laboratory automation potential
Sadeghi et al. [15]	Leishmaniasis	Microscopy (parasite smear images)	Deep learning-based classification	CNN (ResNet, DenseNet)	Very high accuracy (~ 98–99%), sensitivity (~ 100%) and AUC (reported)	Highly standardised parasite morphology enables robust and consistent classification performance in microscopy-based workflows
Abdelmula et al. [16]	Cutaneous leishmaniasis	Clinical lesion / smear images	CNN-based classification	CNN	High diagnostic accuracy and sensitivity (reported)	Clinical-image variability introduces context dependence, but performance remains strong for lesion recognition tasks
Steyve et al. [17]	Skin-related NTDs (Buruli ulcer, leprosy, yaws)	Clinical skin photographs	ML-based image classification	CNN + feature-based ML	Moderate-to-high diagnostic performance across conditions (reported)	Multi-disease classification highlights feasibility of AI-assisted triage, though variability reflects heterogeneity in lesion presentation
Purwar et al. [13]	Malaria	Blood-smear microscopy	ML (microscopy classification)	Image classifier	High overall classification accuracy (reported range)	Standardised parasite morphology supports stable classification under microscopy workflows
Yang et al. [12]	Malaria	Smartphone-captured microscopy images	Deep CNN	Deep CNN	High accuracy and AUC (reported)	Demonstrates feasibility under low-cost acquisition, but still within microscopy’s constrained visual domain
Yotsu et al. [3]	Skin-related NTDs (multiple)	Clinical skin photographs	CNN	CNN	Moderate-to-high diagnostic performance with variability by condition (reported)	Lesion heterogeneity + capture variability likely drive performance dispersion across NTD categories

By contrast, clinical photography-based models for skin NTDs showed greater variability, reflecting the inherent complexity of dermatologic presentation in endemic contexts (Table 4). Skin NTD phenotypes may overlap with non-NTD dermatoses, vary by stage and severity, and be influenced by body site, co-infections, scarring, and treatment history factors that introduce within-class heterogeneity and can attenuate apparent performance when models are exposed to broader case-mix [3, 16]. In addition, real-world image acquisition introduces domain shift (lighting, focus, distance, camera characteristics), which is particularly consequential for lesion-based tasks where the discriminatory cues may be subtle and distributed across texture and colour gradients rather than discrete objects [3, 16]. Taken together, the included studies suggest that “high accuracy” should be interpreted as context-dependent: image-based ML appears most reliable where the diagnostic target is visually constrained and standardisable (microscopy), and more fragile where disease expression is clinically heterogeneous and image capture is variable (clinical photos), reinforcing the need to report performance stratified by modality and disease type rather than as a single pooled claim [3, 12–14, 16].

Dataset quality, representativeness, and external validation

Several microscopy-focused studies achieved strong diagnostic metrics, yet the underlying data pipelines often relied on selected images with clearer fields, better staining, or higher parasite density, which can inflate apparent performance relative to routine laboratory practice where artefacts, poor smears and low-density infections are common [11, 13]. Similarly, helminth and schistosomiasis models often depend on well-annotated egg images, but in real-world settings egg counts are low and background debris is high; without explicit reporting of infection intensity distributions and image inclusion rules, it remains uncertain whether models are robust across the diagnostic spectrum most relevant for control and elimination programmes [11, 15]. For skin NTD diagnostics, representativeness is an even sharper threat: lesion appearance varies by stage, body site, comorbidity, treatment history and skin tone, and datasets that do not capture this heterogeneity risk developing classifiers that perform well on “typical” cases while failing on borderline or atypical presentations that drive misdiagnosis in frontline care [16]. Thus, high accuracy reported within narrowly defined datasets cannot be equated with decision reliability in the heterogeneous contexts of LMIC clinical pathways, where screening and triage are most needed [13, 17].

A closely related threat is the inconsistent use of external validation, which limits confidence that models will generalise across settings, devices and endemic ecologies (Table 5). While some studies incorporated cross-site testing or evaluated performance on distinct datasets, others relied largely on internal splits or cross-validation within a single collection, which is vulnerable to leakage of site-specific imaging signatures (camera type, staining protocols, magnification, lighting) into the model’s learned features [12, 13, 15]. Smartphone-enabled malaria diagnostics demonstrate how device and capture variability can drive domain shift: even when algorithms perform well in controlled evaluations, changes in phone model, microscope attachment, focus and illumination can degrade performance, undermining clinical usefulness unless such variation is explicitly represented in training and tested during external validation [12, 14]. The same logic applies to skin-image models, where generalisability across community clinics, outreach campaigns and referral centres cannot be assumed without geographically and demographically diverse test sets [3, 16]. The evidence indicates that dataset representativeness and validation design are not methodological “details” but determinants of whether image-based ML can move from proof-of-concept to safe, scalable diagnostic support in LMIC health systems [11, 15].

Table 5.

Dataset representativeness and external validation considerations across included studies

Study	Dataset representativeness considerations	Validation approach reported	Implication for clinical utility in LMICs
Yotsu et al. [3]	Clinical skin photos; heterogeneity likely, but representativeness constrained by available labelled cases	Primarily internal evaluation (pilot)	Risk of reduced generalisability across skin tones, lesion stages, and field capture conditions
Purwar et al. [13]	Microscopy images; selection/quality thresholds may bias towards clearer fields	Internal evaluation	May perform less well with low parasitaemia, poor smears, or high artefact backgrounds
Yang et al. [12]	Smartphone-captured smears; device-related variability critical	External testing reported	Feasibility is strong, but performance may vary by phone model, attachment type and capture quality
Oyibo et al. [11]	Stool microscopy images; class balance and infection intensity distributions may affect robustness	Train–test split (reported)	Strong performance, but limited external validation may constrain transfer to other labs/settings
Sadeghi et al. [15]	Microscopy parasite images; dataset likely well-curated with limited variability in acquisition conditions	Train–test split (reported)	Very high performance may reflect controlled dataset conditions; external validation needed to confirm generalisability in field settings
Abdelmula et al. [16]	Clinic-derived lesion and smear images; may under-represent early or atypical presentations	Internal validation (reported)	Potential overestimation of performance outside originating clinical settings due to limited dataset diversity
Steyve et al. [17]	Multi-site clinical skin images; heterogeneity across diseases and settings, but labelling variability may affect consistency	Internal validation (reported)	Supports feasibility across multiple NTDs, but variability in dataset quality and annotation may influence reliability in routine deployment

Operational feasibility

Operational feasibility emerged as an increasingly prominent but still under-examined dimension across the included studies, particularly those leveraging smartphone-enabled microscopy and simplified image acquisition workflows tailored to LMIC constraints. Several microscopy-focused investigations explicitly framed their models as solutions to shortages of skilled microscopists, high diagnostic workloads, and limited laboratory infrastructure, positioning image-based ML as a mechanism for task-shifting and throughput enhancement rather than replacement of clinical expertise [11, 13, 15]. Smartphone-based malaria diagnostics illustrate this trajectory clearly: by combining low-cost microscope attachments with deep learning classifiers, Yang et al. [12] demonstrated that diagnostic performance approaching expert microscopy could be achieved without reliance on high-end laboratory equipment, suggesting feasibility for peripheral health facilities. Similarly, helminth and schistosomiasis studies emphasised automation of egg detection to reduce manual counting time and inter-observer variability, aligning with operational priorities in endemic control programmes where laboratory capacity is stretched [11, 15]. These studies collectively suggest that feasibility is most advanced where ML is embedded within existing diagnostic routines, particularly microscopy, rather than requiring new clinical workflows or specialist skills [13].

However, across the evidence base, operational feasibility was more often implied than empirically evaluated, revealing a critical translational gap. While studies frequently referenced LMIC relevance, few formally assessed implementation metrics such as integration into routine workflows, turnaround time, user training requirements, or maintenance burden under field conditions [3, 12]. For skin NTD diagnostics, feasibility challenges were even more pronounced: although models were proposed as triage or screening tools for community or primary care settings, studies rarely examined how frontline health workers would capture images, interpret outputs, or act on algorithmic recommendations in real time [3]. Even in microscopy-based studies, assumptions of feasibility rested on technical performance rather than prospective evaluation of usability, reliability, or cost-effectiveness at scale [15]. As a result, the distinction between systems that “work technically” and those that “work operationally” remains insufficiently addressed. This gap is particularly consequential in LMIC contexts, where diagnostic tools must function within constrained staffing models, intermittent power supply, variable connectivity, and limited opportunities for retraining. Also, the studies indicate that operational feasibility is emerging, especially for low-resource microscopy and smartphone-based approaches, but that rigorous implementation-focused evaluation is essential before image-based ML can be credibly positioned as a scalable diagnostic support tool in LMIC health systems [11–13].

Explainability and trust-building features

Explainability and trust-building features were sparsely reported across the included studies, yet they emerge as central determinants of whether image-based machine learning can be deployed safely and credibly in LMIC health systems. Only a small subset of the evidence explicitly engaged with XAI approaches, most notably the work by Mohammed et al. [14], which integrated interpretable visual explanations to support Schistosoma egg detection in microscopy images. In this study, explainability mechanisms (e.g. saliency or attention-based visual cues) were framed not merely as technical add-ons but as safeguards that enable clinicians and laboratory staff to interrogate model outputs, identify false positives, and understand failure modes. By contrast, many otherwise high-performing diagnostic models reported in this review provided limited insight into how predictions were generated, leaving uncertainty about whether decisions were driven by biologically meaningful features or by spurious correlations embedded in the training data [3, 15]. This absence of transparency is particularly problematic in LMIC contexts, where diagnostic decisions may be made by non-specialist health workers and where opportunities for independent verification are limited. Without explainability, even accurate models risk being perceived as opaque or unreliable, constraining clinician confidence and uptake regardless of headline performance metrics [14].

The broader methodological literature underscores why this gap matters for governance and accountability. Evidence show that while not focused on NTDs, it remains influential as a methodological comparator because it demonstrated that high-performing deep learning models can achieve clinician-level accuracy while still functioning as “black boxes” unless interpretability is intentionally designed [11, 15, 17]. In the context of NTD diagnostics, this tension is magnified by ethical and regulatory considerations: algorithmic decisions may influence treatment initiation, referral, or exclusion from care in resource-constrained settings, amplifying the consequences of unrecognised bias or systematic error [3]. Where explainability was discussed, it was explicitly linked to trust calibration helping users understand when to rely on the model and when to override it rather than blind acceptance of algorithmic output [14]. However, across the included studies, there was minimal reporting on uncertainty quantification, error auditing, or post-deployment monitoring, all of which are essential components of safe AI use in public health systems. Consequently, the evidence base suggests that explainability is not yet embedded as a standard reporting or design requirement for image-based ML in NTD diagnosis. Without transparent explanation of model reasoning and limitations, such tools risk remaining experimental artefacts rather than trusted decision-support systems capable of integration into LMIC health governance frameworks [3, 14].

Discussion of findings

This review aimed to synthesise the evidence on the diagnostic performance and clinical utility of image-based ML models for NTD diagnosis in LMICs, with particular attention to how model performance translates (or fails to translate) into deployable decision support. Our thematic pattern aligns strongly with the broader AI-for-parasitology and AI-for-NTDs literature, which consistently argues that algorithmic promise is most credible where the diagnostic target is visually standardised (e.g., parasites/eggs on microscopy) and least stable where the phenotype is clinically heterogeneous (e.g., skin NTDs). For malaria, recent syntheses and reviews converge on the conclusion that deep learning can achieve high diagnostic accuracy in controlled settings, but performance is sensitive to smear preparation, staining, optics, and reader-independent ground truth factors repeatedly highlighted as barriers to field generalizability [18]. Similarly, the pattern we observed strong microscopy performance with variability across acquisition pipelines mirrors the pooled malaria DTA findings reported in a recent systematic review/meta-analysis, which estimates high pooled sensitivity and specificity while cautioning that study designs and validation practices often limit real-world inference [19]. Importantly, the same “performance ceiling” narrative appears in technology-facing reports that emphasise AI’s potential to accelerate diagnosis and surveillance for NTDs, while warning that empirical evidence remains concentrated in a subset of diseases and research-active sites rather than routine LMIC care pathways [20, 21].

Where our findings become most consequential relative to prior reviews is in how dataset quality, representativeness, and external validation delimit clinical utility. The recurrent critique in malaria AI reviews is not that models cannot reach high AUC/accuracy, but that reported performance is frequently derived from curated or lab-optimised datasets that under-represent low-parasitaemia infections, artefacts, and device heterogeneity; these are precisely the conditions that drive diagnostic error and delays in high-burden facilities [19, 22]. This concern maps onto emerging schistosomiasis/helminth work that increasingly tests out-of-distribution (OOD) performance under device shifts and mixed-egg conditions, implicitly acknowledging that “headline accuracy” is fragile without deliberate generalisation testing [14]. Likewise, implementation-oriented parasitology studies argue that digital microscopy plus AI can be operationally feasible in primary care, but they also underscore that feasibility depends on infection intensity, workflow integration, and quality control variables that are rarely standardised across studies [22]. For skin NTDs, the wider literature echoes our theme that heterogeneity in lesion morphology, stage, and context produces uneven diagnostic performance; recent work explicitly investigates how small, diverse training sets in dark-skinned populations influence accuracy, reinforcing that representativeness is both a technical and equity requirement [23]. These observations strengthen the critique advanced in disease-specific systematic reviews (e.g., leprosy AI), which similarly identify limited external validation, inconsistent reference standards, and reporting gaps as barriers to translation [24].

A second area where our synthesis extends prior reviews is the distinction between “works technically” and “works operationally”. Across global health AI discourse, feasibility is often inferred from model metrics rather than demonstrated through prospective workflow studies, cost/maintenance analyses, or human factors evaluation; this critique has been repeated in broader AI-for-diagnosis discussions and is especially salient for LMIC deployment [1, 25]. The WHO NTD roadmap emphasises scale-up, country ownership, and integrated approaches for achieving 2030 targets; however, without operational evidence (turnaround times, training burden, calibration requirements, and referral consequences), AI tools risk remaining research artefacts rather than programme assets [25]. Encouragingly, the schistosomiasis ecosystem has moved toward deployable hardware–software systems (e.g., automated or low-cost digital microscopes) that explicitly address portability and field constraints, but even this literature highlights the need for robust validation across endemic sites and routine workflows [26–29].

Finally, our theme on explainability and trust is highly consistent with international guidance that positions transparency, accountability, and human rights as prerequisites for AI in health yet the empirical diagnostic literature still under-reports uncertainty handling, error auditing, and post-deployment monitoring. WHO’s ethics and governance guidance explicitly calls for safeguards that address bias, accountability, and transparency in design and deployment, which is directly relevant where AI outputs could shift treatment and referral decisions in resource-constrained settings [30]. Evidence on AI governance efforts further reinforces that ethical challenges are not peripheral but structurally linked to data quality, deployment context, and power asymmetries in global health technology [27–29, 31–33]. Leishmaniasis-focused studies further reinforce the robustness of microscopy-based ML approaches, where standardised parasite morphology enables highly consistent classification performance, albeit under controlled dataset conditions [34–38]. In this sense, our findings converge with yet also sharpen the argument made by recent skin-NTD-focused discussions: that the next phase of progress requires moving beyond algorithm development toward contextual validation, governance-ready reporting, and implementation science aligned with LMIC systems.

Implications of the review

The findings of this review have important implications for research, clinical practice, and public health policy in LMIC settings. First, the consistently strong diagnostic performance observed for microscopy-based image ML systems indicates that such tools are nearing operational readiness for laboratory support, particularly for malaria and helminth infections, where sensitivities and specificities frequently exceeded 90%. This supports their use as decision-support and workload-reducing tools, rather than autonomous diagnostic replacements, especially in settings with shortages of trained microscopists. Second, the marked variability in performance for clinical image–based models highlight the need for context-aware deployment, whereby algorithm selection, training data, and validation strategies are explicitly matched to disease phenotype, acquisition environment, and intended clinical role (screening vs. diagnosis). Third, the review underscores the necessity of shifting future research priorities away from isolated accuracy optimisation toward external validation, dataset representativeness, and implementation evaluation, including workflow integration and user interaction. Finally, the limited incorporation of explainability mechanisms suggests that governance, trust, and accountability considerations remain underdeveloped, despite their centrality to safe AI deployment. Collectively, these implications argue for a paradigm shift: image-based ML for NTDs should be evaluated not only as a technical innovation but as a health system intervention requiring rigorous validation, ethical oversight, and alignment with LMIC diagnostic pathways.

Limitations of the review

Some limitations acknowledged in this review include that the review was restricted to English-language publications, which may have excluded relevant studies published in other languages from endemic regions. Also, heterogeneity in disease types, image modalities, ML architectures, and reported performance metrics precluded quantitative meta-analysis and limited direct comparability across studies. Third, reliance on published reports introduces the possibility of publication bias, as studies with poor performance may be under-represented. Several studies did not clearly specify whether reported performance metrics were derived from image-level or patient-level classifications, limiting comparability across studies. In addition, many included studies lacked detailed reporting on operational and implementation outcomes, constraining assessment of real-world clinical utility beyond reported diagnostic accuracy.

Conclusion

This systematic review suggests that image-based machine learning approaches show promising diagnostic potential for selected neglected tropical diseases in low- and middle-income countries, particularly in microscopy-based tasks where image features are relatively standardised. Several studies reported high diagnostic performance metrics; however, the strength and generalisability of this evidence remain uncertain due to methodological limitations across the current literature. In particular, many studies relied on curated or limited datasets, reported insufficient external validation, and provided little empirical evidence on how such systems perform within routine clinical or public health workflows. The findings also indicate that diagnostic performance varies considerably depending on disease type, imaging modality, and dataset characteristics. While microscopy-based models often demonstrated relatively consistent accuracy, models based on clinical images of skin NTDs showed greater variability, likely reflecting heterogeneity in lesion presentation, imaging conditions, and population characteristics. Importantly, the review identified limited reporting on operational feasibility, explainability, uncertainty handling, and post-deployment monitoring factors that are essential for safe and trustworthy integration of artificial intelligence tools into health systems. Taken together, the available evidence indicates that image-based machine learning technologies remain largely at the proof-of-concept stage for many NTD diagnostic applications. Future research should prioritise externally validated models, representative datasets from endemic settings, and prospective evaluations of implementation feasibility within real-world healthcare environments. Strengthening these aspects will be critical to determine whether such technologies can meaningfully support NTD diagnostic pathways and contribute to broader disease control and elimination efforts.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AI

Artificial Intelligence

AUC

Area Under the Curve

CDC

Centers for Disease Control and Prevention

CI

Confidence Interval

CNN

Convolutional Neural Network

CNNs

Convolutional Neural Networks

DTA

Diagnostic Test Accuracy

LMICs

Low- and Middle-Income Countries

MeSH

Medical Subject Headings

ML

Machine Learning

NTDs

Neglected Tropical Diseases

PICO

Population, Intervention, Comparator, Outcome

PRISMA

Preferred Reporting Items for Systematic Reviews and Meta-Analyses

PROSPERO

International Prospective Register of Systematic Reviews

QUADAS-2

Quality Assessment of Diagnostic Accuracy Studies-2

RCT

Randomised Controlled Trial

RoB

Risk of Bias

RR

Risk Ratio

STARD

Standards for Reporting of Diagnostic Accuracy Studies

WHO

World Health Organization

XAI

Explainable Artificial Intelligence

Author contributions

DCI conceptualised the study, developed the review protocol, led the literature search, study selection, data extraction, and risk of bias assessment, conducted the narrative synthesis, and drafted the original manuscript. PEA contributed to study screening, data extraction, and critical revision of the manuscript for intellectual content. RCI supported data extraction, quality appraisal using the QUADAS-2 tool, and assisted with synthesis of results. INSD contributed to methodological refinement, interpretation of findings, and critical review of the manuscript. UMC provided expert input on infectious disease context, contributed to interpretation of results, and revised the manuscript critically. COE assisted with data verification, manuscript editing, and formatting in line with journal requirements. All authors read and approved the final manuscript and agree to be accountable for all aspects of the work.

Funding

No funds were received for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.