Systematic review of machine learning approaches for predicting sickle cell crisis and mortality risk at the climate-health nexus

Abstract

Background

Sickle cell anemia (SCA) is a severe genetic blood disorder characterized by recurrent vaso-occlusive crises and increased mortality, with the greatest burden occurring in low- and middle-income countries. Climatic and environmental conditions, including temperature variability, humidity, rainfall, air pollution, and seasonal changes, have been associated with disease exacerbation. However, the extent to which these factors have been incorporated into predictive models remains unclear. This study systematically reviews the application of machine learning (ML) models for predicting SCA crises and mortality in relation to climate and environmental factors.

Methodology

The PRISMA guidelines were used, and 34 peer-reviewed studies published between 2005 and 2026 were analyzed to identify the climate variables, ML approaches employed, and predictive performance. The reviewed studies applied a range of ML techniques, including artificial neural networks, random forests, support vector machines, decision trees, logistic regression, and deep learning models. Temperature, humidity, rainfall, wind speed, air quality indicators, and seasonal patterns were the most frequently examined environmental variables.

Results

The findings indicate that most existing models rely predominantly on clinical and demographic data, with limited integration of climate information and inadequate representation of high-burden regions, especially Sub-Saharan Africa. Studies incorporating environmental variables reported improved predictive performance and highlighted the potential of climate-informed early warning systems for SCA management.

Conclusion

The review recommends development of interdisciplinary, climate-aware ML frameworks, expansion of longitudinal environmental datasets, and increased research in underrepresented regions to support climate-resilient and patient-centered SCA care.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03796-4.

Introduction

As nations continue to make intensive efforts in achieving sustainable development goals (SDGs), sickle cell anemia disease (SCA) is a hereditary blood syndrome which deforms the red blood cell [1] and is a life-threatening disease with high-risk crises and mortality in the African region, especially Nigeria [2]. It causes abnormal haemoglobin formation, leading to distortion of the red blood cell into a crescent-shaped obstacle [3], which obstructs normal blood flow in the vessels and thereby triggers painful crises and complications [4]. The disease results from abnormalities in the haemoglobin gene and is characterized by recurrent episodes of severe pain, vaso-occlusive crises, increased susceptibility to infections, stroke, organ damage, and anemia [5].

Sickle cell anemia disease is the most common severe hereditary disease [6], inherited in an autosomal recessive pattern and detected from birth [7]. It is a global health burden [8], with Africa and Nigeria in particular bearing the highest burden of crises and mortality [9]. About < 90% of children with SCA are born in Sub-Saharan Africa and India, the region of highest prevalence [3]. Further studies by Abubakar et al. [10] show that there is a high tendency for suicide to be committed by patients with SCA due to chronic pain, depression, and stigma/ discrimination. This burden affects low- and middle-income countries that have limited access to healthcare [11]. Despite advances in medical intervention such as blood transfusion, bone marrow transplant, and various supportive therapies [3], SCA crises remain the leading cause of death and reduce quality of life in affected individuals [12]. In addition to genetic causes, environmental stressors are also considered important triggers of crises [13, 14]. Climate change has emerged as one of the most pressing global health challenges, with profound implications for human health. For populations living in fragile environments, mainly in high-altitude and tribal regions, fluctuations in temperature, humidity, and rainfall create conditions that directly influence disease burden [15].

In addition, climatic stressors, which are referred to as environmental and societal pressures, are usually caused by climate change (e.g. high temperatures, droughts, floods, wildfires, and sea-level rise) [16, 17]. It intensifies hypoxic states in the body, reduces oxygen availability in the blood and disrupts red blood cell stability. At high altitudes, reduced barometric pressure lowers the partial pressure of oxygen, producing hypoxia [18]. Anand and Mishra [19] show that when combined with seasonal drops in temperature and increased rainfall, oxygen depletion becomes more severe. Hypoxia triggers a cascade of cellular responses, including altered gene expression and biochemical reactions, that aggravate sickling of red blood cells. The result is recurrent vaso-occlusive crises, rapid red cell destruction and worsening anemia. Prolonged exposure to such conditions may contribute to increased mortality among individuals with SCA [8].

Evidence by Brindha et al. [8] also suggests that long-term climatic variation influences the frequency and severity of SCA crises. Increased rainfall and reduced temperatures have been linked with higher rates of painful episodes. These findings support the idea that climate variability is not merely an environmental concern but a direct health trigger that can heighten genetic disorders such as SCA [20]. Understanding this connection is essential for developing predictive tools and targeted interventions to reduce crises and mortality in vulnerable populations.

However, evidence from retrospective hospital studies also highlights other environmental factors that may play a more critical role [21]. Jones at al. [21] reported that cold weather and rainfall in tropical regions contribute significantly to the frequency of painful episodes. Moreso, analysis showed that increased admissions were significantly associated with high wind speed and low humidity, but did not reveal any clear relationship with temperature, rainfall or barometric pressure. The strongest observed effect was for the combined factor of maximum wind speed to humidity ratio, where admissions were markedly higher in both the lowest and highest quartiles. These findings suggest that skin cooling due to high wind and low humidity may be a major driver of acute SCA crises.

Role of machine learning at the climate-health nexus

Climate variability and environmental degradation are increasingly recognized as important contributors to adverse health outcomes across Africa, mainly in regions with limited healthcare infrastructure and high disease vulnerability [14, 22]. Rising temperatures, irregular rainfall patterns, humidity fluctuations and extreme weather events have been associated with increased disease burden, hospitalization, and mortality. In patients with sickle cell disease (SCA), environmental stressors may contribute to vaso-occlusive crises through dehydration, hypoxia, infection susceptibility, and physiological stress responses [8, 21]. Recent advances in machine learning have improved prediction of SCA severity, complications, and treatment outcomes using clinical and genomic data [23]. However, many existing predictive models do not incorporate climatic and environmental variables, limiting their ability to capture the broader climate-health interactions influencing disease progression and mortality risk. Integrating environmental exposure data with clinical parameters may improve predictive accuracy and support early intervention strategies for climate-sensitive conditions such as SCA.

Related works

Recent advances in machine learning (ML) have significantly improved the diagnosis and classification of sickle cell disease (SCA). For instance, Alharbi et al. [23] reviewed ML techniques applied to SCA and reported that models such as Random Forest, Support Vector Machine, and Artificial Neural Networks achieve high accuracy in predicting disease severity and clinical outcomes. However, the study was limited to clinical and biomedical parameters without integrating environmental or climatic variables that may influence disease progression.

Similarly, Okon et al. [24] applied artificial intelligence algorithms to predict the likelihood of sickle cell crises using patient-related data. Their findings indicated that ML models can effectively support early prediction of crises. Nevertheless, the study was constrained by its reliance on hospital-based datasets, which limited generalizability and excluded external environmental factors such as temperature and humidity. In a more clinical-focused approach, Ussher et al. [25] investigated hematological biomarkers and ML models for classifying SCA severity. The study showed that ML can support severity stratification with improved diagnosis performance. However, the work did not consider external environmental drivers of disease variability, which limits its applicability in climate-sensitive prediction scenarios.

Jinad et al. [26] developed a supervised machine learning framework for predicting malaria incidence in Damaturu, Nigeria, using rainfall and temperature as climatic predictors. The study substantiated the effectiveness of climate-informed machine learning for modelling a climate-sensitive infectious disease; however, its scope was restricted to malaria rather than SCD. Nonetheless, the authors underscored the imperative of integrating climatic variables into predictive health frameworks more broadly.

From a broader healthcare perspective, Machado et al. [27] conducted a systematic review of machine learning applications in SCA research. The study highlighted that most existing models focus on predictions of complications using clinical datasets, while very few incorporate multi-domain data sources. A key limitation identified was the absence of climate-health integration in predictive modeling frameworks. In addition, Boafo, [22] emphasized that climate change significantly exacerbates health risks in vulnerable African populations through rising temperatures, rainfall variability, and environmental degradation. Although the study did not focus specifically on SCA or machine learning. Jinad et al. [26] developed a supervised machine learning framework for predicting malaria incidence in Damaturu, Nigeria, using rainfall and temperature as climatic predictors. The research substantiated the effectiveness of climate-informed machine learning for modelling a climate-sensitive infectious disease. However, its scope was limited to malaria rather than SCA. Nonetheless, the authors underscored the imperative of integrating climate variables into health prediction systems. Recent studies have extended the application of computational and AI-assisted methods to a range of biomedical problems. Wang et al. [28] investigated inflammatory responses in pulmonary endothelial cells, whereas Pu et al. [29] examined molecular mechanisms underlying vascular remodeling in aortic dissection. Zhong et al. [30] and Li et al. [31] explored ferroptosis-related mechanisms in osteolysis and glioma, respectively, and Liu et al. [32] reported clinical outcomes of pembrolizumab-induced autoimmune haemolytic anemia. Most of these studies relied on experimental models or small retrospective cohorts, limiting their clinical generalizability. Their primary focus was disease-specific molecular mechanisms and therapeutic interventions rather than the integration of climatic and environmental factors into predictive healthcare models.

Existing studies indicate strong progress in applying machine learning to SCA prediction. However, they remain largely limited to clinical and biomedical datasets. Very few studies integrate climatic and environmental variables, despite growing evidence that these factors significantly influence SCA crises and mortality risk. This creates a clear research gap that this study addresses by incorporating climate-health data into machine learning-based predictive modelling.

Research questions

The research questions (RQs) were formulated to guide the structural development:

RQ 1

How have machine learning models been applied to predict and assess the effects of climate factors on sickle cell disease outcomes, including vaso-occlusive crises, hospitalization, and mortality?

RQ 2

How were climate exposures assigned, and which climate and environmental variables were incorporated into machine learning models for sickle cell disease?

RQ 3

What methodological challenges and research gaps limit the integration of climate-health data into machine learning models for sickle cell disease?

RQ 4

Which machine learning models have been most frequently applied to predict sickle cell disease outcomes, and how do their predictive performances compare?

Contributions to the study

This work makes a major contribution to the existing body of knowledge by;

• Evaluating the application of ML in SCA research and identifies key methodological limitations.

• Highlighting challenges such as limited datasets, poor validation practices, and lack of transparency and robustness in models.

• Identifying inequity in research, including the underrepresentation of high-burden regions like Sub-Saharan Africa.

• Showing how climate-aware ML models can support early warning systems and improve healthcare outcomes.

• Establishing a foundation for climate-resilient, data-driven, and equitable SCA care.

• Suggesting new direction for future research.

The remaining sections of the paper include Sect. “Methodology”, which discusses the methodology, and section “Results”, which discusses the results. Section “Machine learning approaches for crisis prediction and mortality risks”, discusses the Machine Learning Approaches for Crisis Prediction and Mortality Risks. Section “Privacy, ethics, and data challenges”. Section “Research gaps and limitations of current studies” discusses research Gaps and Limitations of current studies. Discussions and findings are presented in section “Discussions and findings”. Finally, section “Future Directions” provides the Future Directions and Conclusion in section “Conclusion”.

Methodology

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines to ensure transparent and reproducible reporting. The review protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; Registration No. CRD420251082828). The review process followed predefined procedures for literature searching, study selection, data extraction, and quality assessment. Eligibility criteria, screening procedures, and data extraction methods were established before the literature search to minimize selection bias and improve methodological consistency.

Search strategy

A comprehensive literature search was conducted to identify peer-reviewed studies that applied machine learning, artificial intelligence, or statistical prediction methods to investigate the relationship between climate-related factors and sickle cell disease (SCD) outcomes. The review focused on studies that examined vaso-occlusive crises, hospitalization, mortality, disease severity, or other clinically relevant outcomes using clinical, environmental, or climate-health data. The following electronic databases were searched: PubMed, Scopus, Web of Science, ScienceDirect, and Google Scholar. The search covered studies published between January 2005 and May 2026. The final database search was completed on 30 May 2026.

Search terms combined four main concepts: (i) sickle cell disease, (ii) climate or environmental exposure, (iii) prediction methods, and (iv) clinical outcomes. Controlled vocabulary and database-specific indexing terms were used where available. Boolean operators (AND, OR), truncation symbols, and field tags were adapted to the search requirements of each database. A representative PubMed search strategy was:

(“Sickle Cell Disease” OR “Sickle Cell Anemia”) AND (“Machine Learning” OR “Artificial Intelligence” OR “Predictive Model” OR “Statistical Model”) AND (“Climate” OR “Temperature” OR “Humidity” OR “Rainfall” OR “Environmental Exposure”) AND (“Vaso-Occlusive Crisis” OR “Hospitalization” OR “Mortality” OR “Disease Severity”). The databases searched, publication period, and the number of records retrieved before duplicate removal is presented in Table 1. The full search strategy is presented in Appendix A.

Table 1.

Databases included in the literature search

Database	Search Approach	Publication Period	Records Retrieved	Date
PubMed	Database-specific Boolean search	2005–2026	1,950	5-April-2026
ScienceDirect	Adapted Boolean search	2005–2026	5,927	15-April-2026
Scopus	TITLE-ABS-KEY search	2005–2026	23	25-May-2026
Web of Science	Topic (TS) search	2005–2026	7	30-May-2026
Google Scholar	Simplified keyword search	2005–2026	10	7-Febraury-2026
Total			7,917

Inclusion and exclusion criteria

The retrieved documents were examined to identify relevant articles. Articles considered were those that focused on weather-related mortality, sickle cell disease, environmental health, climate change, healthcare analytics, predictive modeling, epidemiology, global health, and crises. Only articles reporting primary and original research were included. Duplicates from various databases were eliminated. Exclusion criteria included: non-English language publications, studies using non-human subjects, studies not involving individuals with SCA, without exposure to climatic factors, simulation with no clinical relevance, and non-peer-reviewed articles (Abstracts, reviews, editorials, or opinion pieces). Data extracted from the selected articles were entered in Microsoft Excel 2016 for statistical analysis. The variables registered for each article were author, year of publication, journal title, design, and area of study.

Eligibility criteria were established before the literature search to ensure a consistent study selection process. Studies were included if they met the following criteria:

• Published in English between 2005 and 2026.

• Reported primary empirical research involving human participants with sickle cell disease or sickle cell anemia.

• Applied machine learning, artificial intelligence, or statistical prediction methods to investigate clinical outcomes such as vaso-occlusive crises, hospitalization, disease severity, morbidity, or mortality.

• Examined the influence of climate, environmental, or meteorological factors, or integrated climate-health data with clinical information.

Reported quantitative findings and described the analytical methods used.

Regression-based epidemiological studies were retained when they examined the relationship between climate exposure and sickle cell disease outcomes. Although these studies did not employ machine learning algorithms, they provided important evidence on climate-health relationships and informed the interpretation of predictive modelling studies. Studies were excluded if they:

• were review articles, editorials, conference abstracts, commentaries, theses, or opinion papers;

• involved animal or laboratory-only experiments without clinical relevance;

• did not include individuals with sickle cell disease;

• did not investigate climate, environmental, or weather-related factors;

• did not report sufficient methodological details or quantitative outcomes; or.

• were not peer-reviewed or were published in languages other than English.

Duplicate records identified across databases were removed before screening. Information extracted from the eligible studies included publication details, study design, sample characteristics, climate variables, analytical methods, outcome measures, and key findings.

Data extraction

Analytical categorization of the included studies. The included literature was categorized into two analytical categories. Category A included studies related to machine learning and artificial intelligence that used predictive modeling in relation to climate-related Sickle Cell Disease outcomes. Category B included epidemiological and regression-based studies that studied the relationships between climatic or environmental exposure and Sickle Cell Disease outcomes. This second Category was kept due to the fact that epidemiological evidence can help understand the relationship between exposure and outcomes and provide complementary information to prediction literature where there is less evidence from ML.

Geographic distribution of included studies

The included studies were conducted across multiple regions, including Africa, North America, Europe, South America, and the Caribbean as shown in Fig. 1. North America contributed the largest number of studies, followed by Africa and Europe. Fewer studies originated from South America and the Caribbean, highlighting the limited geographic coverage of climate-related sickle cell disease research despite the high disease burden in many low- and middle-income countries. This uneven distribution suggests that current evidence may not fully capture regional differences in climate exposure, healthcare access, and disease patterns.

Fig. 1.

Geographic distribution of included studies in the review

Data extraction and synthesis methods

A structured data extraction form was developed before the review to ensure consistency across the included studies. The extracted information comprised bibliographic details (authors, year, country, and publication source), study design, sample characteristics, sickle cell genotype, study setting, machine learning or statistical prediction methods, climate and environmental variables, data sources, clinical outcomes, model evaluation metrics, and interpretability features where reported. Information on study limitations was also extracted. Table 2 summarizes the data extraction framework. Data were synthesized using a narrative approach because of the variation in study designs, datasets, outcome measures, and analytical methods. The included studies were organized according to the research questions and compared based on their study characteristics, climate exposure variables, analytical approaches, clinical outcomes, predictive performance, and methodological limitations. The synthesis also examined patterns in climate exposure assignment, model validation strategies, and commonly reported performance metrics to identify methodological trends and research gaps. A quantitative meta-analysis was not performed because of the heterogeneity of the included studies.

Table 2.

Data extraction framework

Category	Details Extracted
Study Information	Authors, publication year, country, journal or conference
Study Design	Experimental, observational, cross-sectional, cohort, case-control, or retrospective study
Study Population	Sample size, participant characteristics, SCD genotype, study setting
Climate and Environmental Data	Climate variables, data source, spatial resolution, temporal resolution, exposure assignment method
Analytical Method	Machine learning, artificial intelligence, or statistical prediction methods; model architecture where applicable
Clinical Outcomes	Vaso-occlusive crisis, hospitalization, mortality, disease severity, complications, or other reported outcomes
Data Sources	Clinical records, meteorological databases, public repositories, wearable devices, or other data sources
Model Evaluation	Accuracy, sensitivity, specificity, precision, recall, F1-score, AUC, RMSE, MAE, calibration measures, or other reported metrics
Interpretability	Feature importance, SHAP, attention mechanisms, or other explainability methods where applicable
Key Findings	Principal findings related to climate-health prediction and SCD outcomes
Limitations	Methodological limitations reported by the authors

Quality assessment of selected studies

The methodological quality of all included studies was evaluated using the Newcastle-Ottawa Scale (NOS), which is specifically designed for assessing observational research. The NOS assesses studies across three major domains: Selection (4 points), Comparability (2 points), and Exposure (3 points), allowing each study to receive a total quality of up to 9 points. Based on these scores, studies were classified into the following categories: high quality (7–9 points), moderate quality (4–6 points), or low quality (0–3 points). Quality assessment was conducted by two independent reviewers to minimize subjectivity and ensure consistency. Any discrepancies in scoring were resolved through discussion and, if necessary, consultation with the third reviewer. A comprehensive summary of the quality ratings was presented in tabular form and integrated into the interpretation of the overall findings to strengthen transparency and contextualize the robustness of the evidence base.

The Newcastle-Ottawa Scale (NOS) has been developed mainly for methodological evaluation of observational studies in epidemiology and does not assess all characteristics associated with the development of machine learning models and their predictive power. In this systematic review, the Newcastle-Ottawa Scale (NOS) was used for the methodological evaluation of the observational components of selected studies, namely, the selection of participants, comparability, and exposure/ outcome assessment. Methodological components of the ML models were assessed separately in the process of data extraction, including the type of the model, its inputs, validation, evaluation, and interpretability.

Results

Study selection process

All records retrieved from the database searches were exported to EndNote for reference management and duplicate removal. The remaining studies were screened using a three-stage selection process based on the PRISMA 2020 guidelines. In the first stage, two reviewers independently screened the titles and abstracts to identify potentially relevant studies using the predefined eligibility criteria. Studies that did not satisfy the inclusion criteria were excluded at this stage. In the second stage, the full texts of the remaining articles were assessed independently by the same reviewers to confirm their eligibility. Each study was evaluated against the predefined inclusion and exclusion criteria. Reasons for exclusion were recorded for studies that did not meet the eligibility requirements. Any disagreement during title, abstract, or full-text screening was resolved through discussion until consensus was reached. Where necessary, a third reviewer was consulted to reach a final decision. Formal inter-rater agreement statistics were not calculated.

In addition, using the database search, 7917 publications were found, after the removal of 6450 studies and duplicate studies using each database’s content-type filter. After screening the titles and abstracts, 1,392 of the 1,467 remaining articles were eliminated, leaving only 75 pertinent articles. After a last round of screening, 34 out of the 75 papers were determined to be eligible for inclusion and were added to the review, and 41 papers were discarded due to inconsistency and non-alignment. Using the inclusion and exclusion criteria mentioned above, two authors independently reviewed the titles and abstracts of the identified articles to decide which ones qualified for this systematic review. The reviewers came to a consensus after exchanging minor points of contention. The study selection screening process is shown in Fig. 2.

Fig. 2.

Selection screening process

Characteristics of included studies(Map + Characteristics table)

A total of 34 primary studies met the inclusion criteria and were included in the qualitative synthesis. The studies were published between January 2005 and May 2026, that were conducted across multiple geographical regions (Fig. 1). Most studies originated from North America, followed by Africa and Europe. The included studies comprised observational epidemiological investigations, retrospective hospital-based analyses, registry studies, and machine learning prediction studies. Most investigations focused on vaso-occlusive crises, whereas a smaller number examined mortality, acute chest syndrome, disease severity, or other clinical outcomes. The machine learning studies evaluated algorithms such as Random Forest, Support Vector Machine, Artificial Neural Network, Convolutional Neural Network, and hybrid deep learning models for diagnosis, severity classification, or crisis prediction. The detailed characteristics of included 34 studies examining climate and environmental factors in relation to sickle cell disease outcomes is presented in Appendix B.

Environmental exposure assignment methods

The included studies employed different approaches to assign environmental exposures associated with sickle cell disease outcomes. Most studies obtained meteorological variables, including ambient temperature, relative humidity, rainfall, wind speed, and atmospheric pressure, from national meteorological agencies or local weather stations located near the study sites as presented in Table 3. A smaller number of studies relied on publicly available climate databases or regional environmental monitoring systems. Most investigations linked daily or short-term weather conditions with vaso-occlusive crises, whereas only a few evaluated cumulative or seasonal exposures. Reporting of spatial resolution and exposure lag periods was inconsistent across studies, limiting direct comparison of exposure assessment methods and contributing to methodological heterogeneity.

Table 3.

Environmental exposure assignment methods used across the included studies

Study	Exposure Source	Climate Variables	Spatial Resolution	Exposure Lag
Jones et al. [21]	Local weather station	Temperature, humidity, wind speed	City	Same day
Dessap et al. [33]	Meteorological station	Temperature, humidity	Regional	0–2 days
Parriault et al. [2]	National meteorological service	Temperature, rainfall, humidity	Regional	0–7 days
Desai et al. [13]	National weather database	Temperature, seasonal variation	National	Not reported
Asnani et al. [34]	Environmental monitoring records	Socio-environmental exposures	Community	Not reported

Outcome definitions across studies

Considerable variation existed in the clinical outcomes evaluated across the included studies. Most investigations examined vaso-occlusive crises (VOC), although definitions differed according to study setting. Several studies defined VOC as hospital admission for pain episodes, whereas others considered emergency department visits or clinically confirmed pain crises. Additional outcomes included acute chest syndrome, disease severity, mortality, and overall healthcare utilization. This variability in outcome definitions contributed to differences in reported associations and limited direct comparison between studies is presented in Table 4.

Table 4.

Outcome definitions reported by the included studies

Study	Primary Outcome	Outcome Definition
Jones et al. [21]	VOC	Hospital admission
Dessap et al. [33]	VOC	Emergency hospital admission
Parriault et al. [2]	VOC	Hospital-treated pain episode
Desai et al. [13]	VOC / ACS	Hospital admission and acute chest syndrome
Abubakar et al. [10]	Mortality	Death among hospitalized SCD patients
Okon et al. [24]	Crisis prediction	Predicted probability of VOC

Quality assessment of included studies

The methodological quality of the included studies was assessed using the Newcastle–Ottawa Scale (NOS). Most studies were classified as high quality (scores 7–9), whereas a smaller proportion achieved moderate quality (scores 5–6), as represented in Fig. 3. No study was rated as low quality. Lower-quality studies generally lacked detailed descriptions of exposure assessment methods, adjustment for potential confounding factors, or clear definitions of outcomes. Nevertheless, the principal findings were consistent across methodological quality categories, suggesting that the overall conclusions of this review were not substantially influenced by study quality.

Fig. 3.

Sample distribution of methodological quality among the included studies based on the NOS

The climate-health nexus in sickle cell anemia

Pathophysiology of climate triggers in sickle cell crises

People living with SCA appear to be highly sensitive to changes in their environment, and Vaso-occlusive Crises (VOC) are often triggered by shifts in weather and climate. Studies suggest that fluctuations in temperature and humidity may disrupt normal blood flow, hydration, and oxygen delivery, potentially setting the stage for sickling events [35]. Cold weather is thought to constrict blood vessels and limit oxygen supply, excessive heat may lead to dehydration and thicker blood, both of which increase the risk of a crisis. Changes in humidity add another layer of stress: dry air accelerates fluid loss, whereas very humid conditions make it harder for the body to release excess heat [33]. These factors help explain why people with SCA are especially sensitive to even small changes in climate and weather.

Common climate factors

Several climatic factors emerge consistently in the reviewed studies as contributing to SCA complications:

• (i)Temperature: Temperature plays a critical role in shaping sickle cell outcomes, and both extremes can be harmful. During cold weather, narrowing of blood vessels often leads to reduced circulation and oxygen delivery, which is linked to a rise in hospital admissions for vaso-occlusive crises. On the other hand, periods of extreme heat increase fluid loss and disrupt electrolyte balance, creating conditions that promote dehydration and trigger crises [33, 35].
• (ii)Humidity: Low humidity raises hydration risks, whereas high humidity impairs thermoregulation, leading to greater physiological strain [34].
• (iii)Rainfall: Variations in rainfall can influence hydration and the transmission of infections such as malaria, indirectly increasing SCA morbidity [4].
• (iv)Air quality and pollution: Poor air quality whether from dust, smoke, or urban pollutants has been associated with acute chest syndrome and other respiratory complications, further intensifying SCA outcomes (De, A. et al. 2023).
• (v)Seasonal transitions: Sudden weather changes, mainly harmattan-to-rainy season cycles in West Africa, heighten the likelihood of crisis events by combining shifts in temperature, humidity, and air quality [36]. Harmattan is a dry, dusty trade wind that blows from the Sahara Desert across West Africa, typically between November and March (from November to March), characterized by low humidity, cooler temperatures, and reduced air quality and can cause drought.

Our in-depth findings show that climatic conditions act not as isolated triggers but as interacting stressors that exacerbate the underlying pathophysiology of SCA.

Epidemiological evidence linking climate and SCA outcomes

Epidemiological research supports the strong link between climate and health outcomes in SCA. Studies in France found that environmental factors such as temperature and humidity influenced emergency admissions for SCA crises, showing that the climate health nexus extends beyond tropical regions [33]. In West African countries, seasonal patterns, mainly during the harmattan were consistently associated with spikes in hospitalization for VOC. Broader socio-environmental analyses also confirm that fluctuations in temperature and humidity contribute to increased SCA mortality [34]. State some empirical studies or instances or examples of some studies that have provided or experienced Epidemiological evidence linking climate and SCA. Nonetheless, limitations persist across these studies. Some did not assess mortality, others lacked climate variables, and very few used predictive methods such as machine learning. This fragmentation leaves gaps in understanding which climatic exposures carry the greatest risks across different geographies. Addressing these limitations requires more systematic, multi-site studies that integrate both epidemiological and environmental data.

Machine learning approaches for crisis prediction and mortality risks

Overview of ML models applied in SCA studies (Revised)

Machine learning has been widely applied to predict and classify clinical outcomes in sickle cell anemia (SCA). Most of the reviewed studies developed prediction models using clinical records, laboratory biomarkers, medical images, physiological measurements, or wearable sensor data. These studies establish the current baseline for machine learning applications in SCA and have focused primarily on predicting vaso-occlusive crises (VOC), disease severity, hospitalization, and mortality [27]. Supervised learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Gradient Boosting, and Multilayer Perceptron (MLP), were the most frequently reported methods. Deep learning architectures, such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, were commonly applied to medical image analysis and longitudinal clinical data. High predictive performance was reported in several studies; however, these results were largely obtained from models developed using clinical or biomedical data rather than climate-health information. Model performance also varied according to the prediction task, dataset characteristics, and validation strategy adopted by individual studies.

Machine learning has also been used to estimate mortality risk through survival models, including Random Survival Forests and Cox regression, which support patient risk stratification using demographic and clinical variables. Despite these advances, only a limited number of studies incorporated climate or environmental variables such as temperature, humidity, rainfall, or seasonal variation into model development. As a result, the integration of climate-health information into machine learning models for SCA remains limited, leaving an important gap in understanding how environmental exposures influence disease progression and clinical outcomes.

Supervised learning approaches

Supervised learning techniques such as Random Forest (RF), Support Vector Machines (SVM), and Gradient Boosting have strong performance in SCA prediction tasks. Random Forest models are mainly effective in handling noisy, incomplete, and heterogeneous clinical data, which are common in real-world healthcare environments. Their ensemble structure reduces overfitting while capturing complex interactions between clinical features [37]. SVMs, on the other hand, are well-suited for high-dimensional datasets, such as those combining hematological parameters, demographic variables, and patient history. By identifying optimal separating hyperplanes, SVMs have been shown to accurately distinguish SCA complications from other forms of anemia [38]. Studies by Okon et al. [24] further shows that Gradient Boosting and SVM models can successfully predict VOC, enabling early warnings through real-time monitoring systems. Such predictive capabilities have the potential to reduce pain severity, hospital admissions, and mortality. Ensemble approaches that combine multiple supervised models further improve robustness and reliability across different patient populations, which is essential for clinical deployment and physician trust.

Deep learning approaches

Deep learning has significantly advanced image-based diagnosis in SCA, mainly through automated analysis of blood smear microscopy images. Convolutional Neural Networks (CNNs) excel at learning spatial and morphological features of red blood cells, enabling accurate identification of sickled cells [39].

Hybrid CNNLSTM architectures represent an important innovation by combining spatial feature extraction with temporal modeling. In such frameworks, CNNs analyze red blood cell morphology, while Long Short-Term Memory (LSTM) networks capture temporal changes across image sequences or longitudinal data [40]. This dual capability improves diagnostic precision, especially when disease progression is dynamic. Comparative studies show that deeper CNN architectures such as ResNet50 and InceptionV3 outperform shallower models like VGG16 due to their ability to learn richer hierarchical representations [41]. These advances support the integration of deep learning into routine pathology workflows, enabling high-throughput, objective screening in regions where hematopathology expertise is scarce.

Hybrid and multi-scale models

Hybrid and multi-scale models combine the strengths of multiple architectures to extract richer and more clinically meaningful information from biomedical data. Hybrid CNN-LSTM models leverage CNNs for detailed morphological analysis and LSTMs for learning temporal disease patterns, enabling detection of subtle changes that single-model approaches may miss [42]. Multi-scale CNNs use filters of varying receptive fields to capture both fine-grained cellular details and broader contextual patterns. This is mainly important for blood smear images affected by staining variability, magnification differences, and patient-specific morphology. Aggregating features across multiple scales improves model generalizability and stability across datasets [43]. Additionally, explainable hybrid architectures help clinicians understand which cellular features drive predictions, improving interpretability and clinical acceptance.

Feature engineering: clinical, biomarker, and environmental variables

The available information from the selected literature shows that the feature engineering process in the ML model used in SCA involves mainly the use of clinical and biochemical variables. Commonly used features include levels of haemoglobin, reticulocytes, white blood cells, platelets, lactate dehydrogenase (LDH), and bilirubin. They are usually obtained through laboratory equipment such as hematology analysers, high-performance liquid chromatography (HPLC), and ELISA platforms. The features are then used to predict disease outcomes through classification of severity, crisis occurrence, and mortality risks [25].

The systematic review by Machado et al. [27] represents the most extensive evaluation of ML instrument application in SCA available, but its focus remains primarily on clinical and hematologic predictors rather than environmental factors. This review evaluates several types of machine learning tools such as Multilayer Perceptron, Support Vector Machine, Random Forest, Logistic Regression, Long Short-Term Memory, Extreme Learning Machines, Convolutional Neural Networks, and Transfer Learning techniques. In the reviewed articles, the input predictors include mostly hematological indicators (hemoglobin, hematocrit, mean corpuscular volume, reticulocyte count, fetal hemoglobin), vital signs (heart rate, blood pressure, oxygen saturation, respiratory rate), microscopic findings on blood smears, and clinical outcomes derived from electronic medical records. Organ-level indicators like glomerular filtration rate (GFR), liver enzymes such as AST and ALT, creatinine, and oxygen saturation, were used in various monitoring datasets reviewed, and were determined to provide a valid reflection of disease severity and potential complications. Patient-level measures like frequency of pain events, history of crises, compliance with drugs (specifically dosage of hydroxyurea), and frequency of visits to hospitals were important data points. By employing these in LSTM models as time-series data, forecasting of VOC was possible.

These studies made use of wearable physiological monitoring tools as well as institutional electronic health records. In one of the studies, PPG signals gathered using wearable sensors from the sleep periods of children with SCA were used to predict vulnerability to pain crisis events, with the two-level stacked ML framework performing better than other base learners, especially the Multilayer Perceptron. Another study involved an Apple Watch in conjunction with a pain reporting mobile app (Nanbar) that produced wearable-generated heart rate variability and steps taken data. The resulting data were then classified using a Random Forest algorithm with an accuracy of 92% in predicting VOC pain index values.

The findings reported by Machado et al. [27] serve as important guidance when interpreting feature engineering statements within this field. The sample size of several studies cited in the review is rather small, and this might restrict the applicability of their findings. Large discrepancies between the preprocessing methods, labeling stages, and algorithms utilized have been observed, thus hindering any direct comparison between the articles. In addition, no external validation has been performed in most of the studies, raising concerns regarding the generalizability of their findings to different cohorts. Notably, not a single one out of the 34 studies assessed by Machado et al. [27] employed climatological factors as features an important omission that is addressed in this study.

However, the reviewed article by Elsabagh et al. [44] concentrated on utilizing the principles of artificial intelligence and machine learning in the diagnosis, prognosis, and treatment of Sickle Cell Disease. Therefore, no climate parameters such as temperature, precipitation, humidity, velocity of air movements, or air quality indicators were used in the presented machine learning algorithms. The study authors highlighted the application of clinical, biochemical, imaging, and physiological features to predict outcomes and describe characteristics of the examined health condition.

The researchers identified several technological tools and devices in their paper. These included image-based diagnostic solutions, digital microscopy, retinal imaging methods, mobile apps for health monitoring, wearable gadgets, and biosensors. In turn, the authors applied different types of machine learning and deep learning models, including CNN, RNN, logistic regression, support vector machine, and Random Forest algorithms, to categorize sickle cells, forecast severity of the disease, calculate mortality risk, and detect complications related to organ impairment and vaso-occlusion. Moreover, the experts employed techniques of Explainable Artificial Intelligence, such as SHAP and Grad-CAM, to enhance model interpretation and clinician confidence.

The applications were primarily focused on disease diagnosis automation, improving prediction precision, enabling personalized treatments, and facilitating clinical decision-making processes. Image-based characteristics like red blood cell shape, texture parameters, and geometry attributes were extracted from the blood smear images and utilized in machine learning algorithms to improve their predictive capabilities. Nonetheless, the research pointed out various shortcomings of the methods under review, such as a lack of substantial and balanced datasets, minimal external validation, reliance on historical patient records, diminished population-level generalizability of the models, and poor infrastructure for large-scale implementation of AI technology in health systems in developing countries. Additionally, the lack of climatic factors prevents the models from evaluating the impact of climate change on SCA patients.

The integration of climate-related factors into forecasting models of the consequences of SCA falls under the ambit of a new but analytically different research area in machine learning studies. Sickle cell disease vaso-occlusive events (VOC) lead to high morbidity and costs from frequent visits to the emergency departments and hospitals, with recurrent cases of VOC often leading to various complications such as the acute chest syndrome (ACS) [45]. Although the pathophysiological link between climatic factors and the occurrence of VOC has been known for decades, its numerical analysis was possible only recently using administrative data.

Desai et al. [45] form the most pertinent paper to the current research inquiry. The researchers did not use a machine learning model for classification, they used negative binomial regression modeling to calculate the incidence rate ratios (IRR) of VOC and ACS admissions based on four climatic factors, which included ambient temperature, delta of temperature (difference per day), precipitation, and wind speed. The PHIS database was accessed for identifying SCA cases under the age of 21 years who had visited the emergency department or had been hospitalized for VOC and ACS between 2016 and 2021. The PHIS database acted as the key tool for case detection and outcome assessment, whereas the meteorological parameters were mapped to geographical areas of the hospitals’ location - Midwest, South, Northeast, and West.

The technical approach used in Desai et al. [45] may be described as an analysis of administrative claims that incorporates ecological linkages between climate and health. In the case of each VOC and ACS admissions dataset, the climate at the time of admission in the respective hospital’s area was determined, and a negative binomial regression was run to address the issue of overdispersion when modeling count variables. The method is commonly used in environmental epidemiological studies and ensured sufficient statistical power due to the national sample size of 79,872 admissions.

The outcomes of the analysis of weather variables indicated notable geographic heterogeneity. All three weather factors such as temperature, change in temperature, and wind speed, showed significant associations with VOC occurrence in the Midwest (p < 0.0001 for all) and Northeast (p ≤ 0.0001 for all), precipitation showed the same trend in both climatic zones (p < 0.0001). However, no relationship was found between any of the four studied weather variables and VOC rates in the South (p = 0.110 to p = 0.833), and some inconsistencies were observed in the West.

There are several weaknesses in this instrument and method of analysis relevant to this review. Firstly, there is a significant risk of ecological bias associated with the use of regional average climate parameters rather than more localized or patient-level weather information. Secondly, the inclusion criteria for PHIS data only include pediatric admissions at children’s hospitals, resulting in limited generalizability of results to adults with SCA and to patients outside specialized facilities. Thirdly, the timeframe included in this study (2016 to 2021) encompasses the period of the COVID-19 pandemic, recognized here as a confounding factor for patterns in health utilization. Lastly, and most importantly for this review, the regression method used by the authors is not a machine learning algorithm; consequently, it does not include feature interaction analysis, non-linear thresholds, or validation, which would allow the contribution of weather parameters to be quantified in an ML model.

Machine learning approaches for SCD prediction and climate-health integration

Most of the reviewed studies applied machine learning to predict sickle cell disease outcomes using clinical, laboratory, imaging, or physiological data. These studies provide the baseline for current machine learning applications in SCD. Climate and environmental variables were incorporated in only a few studies, indicating that the integration of climate-health data into predictive models remains limited. For instance, frequently used models included Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), C5.0 Decision Tree, Bagged Trees (TREEBAG), AdaBoost, Gradient Boosting Machine (GBM), Neural Networks (NNET), k-Nearest Neighbour (KNN), Elastic Net (ENET), Logistic Regression (LR), Linear Discriminant Analysis (LDA), and Naïve Bayes (NB).

Ussher et al. [25] evaluated these algorithms using haematological, biochemical, and metabolic biomarkers collected from 481 participants, comprising 356 patients with SCA and 125 healthy controls recruited from Korle-Bu Teaching Hospital, Ghana. The objective was to classify disease severity into mild and severe categories. Data preprocessing included feature scaling, outlier removal using the interquartile range (IQR), Yeo–Johnson transformation, and binary severity labelling. Feature reduction and biomarker selection were performed using Principal Component Analysis (PCA) and Partial Least Squares Discriminant Analysis (PLS-DA). In contrast, Monte Carlo cross-validation and repeated 10-fold cross-validation were used to assess model stability.

Direct bilirubin, total bilirubin, reticulocyte count, hydrogen sulphide (H₂S), and neutrophil count were identified as the most informative biomarkers for severity prediction. The full biomarker model achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.98 with a classification accuracy of 92.4%, indicating excellent discrimination between mild and severe disease. Performance varied across the machine learning models. Ensemble learning methods produced the most consistent results. Random Forest, XGBoost, C5.0 Decision Tree, and TREEBAG achieved AUC-ROC values above 0.80 and AUC-PR values exceeding 0.85. Support Vector Machine recorded the highest median classification accuracy of approximately 80%, with a mean accuracy of 79.9% and a kappa coefficient of 0.598. AdaBoost followed with a mean accuracy of 73.8% and a kappa value of 0.476, whereas C5.0, XGBoost, and TREEBAG achieved mean accuracies of 72.7%, 71.9%, and 70.1%, respectively. Conventional statistical approaches, including Logistic Regression, Linear Discriminant Analysis, Elastic Net, and k-Nearest Neighbour, produced lower predictive performance, with AUC-ROC values ranging from 0.53 to 0.56.

Despite these encouraging results, several methodological limitations were reported. The biomarker panel did not capture all sources of variation observed during principal component analysis, suggesting that additional biological markers could improve prediction. The study relied on data from a single healthcare institution, limiting population diversity and external validity. Clinical information such as medication history, co-existing diseases, blood pressure, and body mass index was incomplete, introducing the possibility of residual confounding. Gender imbalance across study groups was addressed through regression modelling, although some bias may have remained. Disease severity was also simplified into two categories, improving model training but reducing clinical detail. Ethical concerns related to patient privacy and model interpretability remain important considerations before routine clinical deployment.

Other studies reported similar challenges. Limited sample sizes, class imbalance, and inadequate external validation increased the risk of overfitting and reduced model robustness. Collins et al. [46] noted that prediction models evaluated only on their development datasets often produce optimistic performance estimates, especially when sample sizes are small. These findings suggest that strong internal performance does not necessarily translate into reliable clinical use.

Across the reviewed studies, ensemble learning methods such as Random Forest, XGBoost, C5.0, and Bagged Trees consistently achieved better predictive performance than conventional statistical models. Their effectiveness, however, depends on access to representative datasets, rigorous external validation, and clinically meaningful evaluation. Future studies should place greater emphasis on standardized datasets, multimodal data integration, external validation across diverse populations, and clinician-centred evaluation to improve reproducibility and support clinical adoption. The comparative summary of ML approaches for SCA prediction is presented in Table 5 and system flowchart diagram is illustrated in Fig. 4.

Table 5.

Comparative summary of machine learning approaches for sickle cell anemia prediction

Study	Intervention	Data Source	Method	Performance	Key Limitation
Ussher et al. [25]	SCA severity prediction	Korle-Bu Teaching Hospital, Ghana (481 participants)	RF, SVM, XGBoost, C5.0, TREEBAG, AdaBoost	AUC = 0.98; Accuracy = 92.4%	Single-centre dataset; limited biomarkers; no external validation
Salman et al. [37]	Disease risk prediction	Clinical dataset	Random Forest	High prediction accuracy	Computational cost
Machado et al. [27]	Disease progression modelling	Longitudinal clinical records	LSTM/GRU	Improved temporal prediction	Limited interpretability
Deo et al. [40]	Multimodal disease prediction	Clinical and imaging data	CNN–LSTM	Higher predictive accuracy	Increased model complexity
Elsabagh et al. [44]	Multimodal clinical prediction	Clinical, imaging, and environmental data	Multimodal fusion	Improved decision support	Data integration challenges

Fig. 4.

The system flowchart diagram

Privacy, ethics, and data challenges

The integration of climate and health data for machine learning (ML) modeling presents significant challenges related to data availability, privacy, and ethics. These issues are mainly acute when dealing with a vulnerable population like those with Sickle Cell Anemia (SCA), who are often located in regions with limited data infrastructure and heightened socio-economic vulnerabilities.

Data scarcity and regional bias in SCA climate studies

A fundamental challenge in building robust ML models for the SCA-climate nexus is the severe scarcity of high-quality, granular data. This scarcity is not uniform but is marked by a profound regional bias. High-income countries (HICs) are establishing detailed SCA registries and have access to high-resolution climate data; the countries with the highest SCA burden, primarily in Sub-Saharan Africa (SSA), often lack the necessary digital health infrastructure [47]. This creates a “data desert” for the very populations most affected by climate-sensitive SCA crises.

Consequently, existing models’ risk being trained on unrepresentative data from HICs or specific urban centers within low- and middle-income countries (LMICs), leading to models that fail to generalize to broader SCA populations [27]. For instance, a model trained on data from the United Kingdom, where Jones et al. [21] identified wind speed and humidity as key triggers, may not accurately predict crises in the Sahel region of Africa, where Harmattan dust and seasonal rainfall are predominant factors. This regional bias perpetuates a cycle of neglect, where the populations most in need of predictive healthcare tools are systematically excluded from the research that could generate them.

Data privacy in health and climate datasets

The convergence of sensitive personal health information (PHI) with detailed environmental data raises critical privacy concerns. PHI, governed by regulations like the Health Insurance Portability and Accountability Act (HIPAA) in the U.S. and the General Data Protection Regulation (GDPR) in Europe, is highly protected. However, when PHI is linked with granular climate or location data, even anonymized datasets can be vulnerable to re-identification attacks [48]. For example, a dataset containing a patient’s home coordinates (for hyper local climate exposure) and hospitalization dates for crises could be cross-referenced with public records to re-identify individuals. This risk is amplified in small or close-knit communities, which are common in high-SCA-prevalence regions. Therefore, standard de-identification techniques are often insufficient. Researchers must adopt privacy-by-design principles, leverage techniques such as k-anonymity and differential privacy when creating and sharing linked climate-health datasets to ensure that individual patients cannot be re-identified [49]. Also, a more recent study by Ikegwu et al. [14] implemented federated learning frameworks for healthcare cyber threat detection. It showed that privacy-preserving techniques such as differential privacy and homomorphic encryption allow collaborative model training without exposing sensitive patient records. The study also identified communication cost, model heterogeneity, and deployment in resource-constrained environments as key challenges. These findings support the use of federated learning for secure climate-health prediction systems involving multiple hospitals.

Ethical considerations in predictive models for vulnerable populations

The development and deployment of predictive ML models for SCA crises necessitate a rigorous ethical framework. SCA patients are a clinically and often socio-economically vulnerable group, and models that inaccurately predict their risk could lead to either a false sense of security or unnecessary anxiety and medical intervention [50].

A primary ethical concern is algorithmic bias. If models are trained on biased data, as discussed in “Data scarcity and regional bias in SCA climate studies”, they will produce biased outputs, potentially exacerbating existing health disparities. For instance, a model that under-predicts risk for a specific ethnic subgroup due to their underrepresentation in training data could lead to that group being denied timely interventions. Furthermore, the clinical implementation of such models requires careful thought. Who is responsible for acting on a high-risk prediction? How are patients informed of their risk? There is a danger of placing the burden of crisis prevention solely on the patient (e.g., by advising them to avoid certain weather conditions, which is often impossible) rather than on healthcare systems to provide proactive support [51]. Ethical practice demands that predictive tools are developed transparently, validated across diverse populations, and integrated into clinical care in a way that empowers, rather than blames, the patient.

Research gaps and limitations of current studies

Despite the promising intersection of machine learning, climate science, and SCA research, the field is nascent and characterized by several significant gaps that limit the validity, generalizability, and real-world impact of current studies.

Methodological gaps (sample sizes, validation strategies, reproducibility)

Many existing studies in the broader digital health space suffer from methodological weaknesses that are prevalent in this specific domain. A common issue is the use of small, single-center datasets, which leads to over-fitted models that perform poorly on external data from different geographic or demographic populations [79]. The validation strategies employed are often inadequate; many studies rely solely on simple train-test splits rather than more robust methods like k-fold cross-validation or external validation on completely independent cohorts. Furthermore, a critical gap is the lack of reproducibility. Studies frequently fail to provide full details on their pre-processing pipelines, hyper parameter tuning methods, or the exact definitions of climate variables used. The code and data are rarely made publicly available, preventing other researchers from verifying results or building upon existing work. This lack of transparency and reproducibility hinders scientific progress and the development of reliable, standardized models [52].

Gaps in climate data integration with clinical SCA data

The integration of climate and clinical data remains superficial in many studies. A significant gap is the temporal and spatial misalignment of datasets. Clinical data (e.g., crisis onset) is recorded at a precise time, while climate data is often available as daily averages from weather stations kilometers away from a patient’s location. This fails to capture microclimates and immediate exposures. Future research must leverage high-resolution, gridded climate reanalysis data and, where possible, personal environmental sensors to achieve spatiotemporal alignment with health events [53].

Moreover, most studies focus on a limited set of climate variables like temperature and humidity, overlooking complex compound extremes (e.g., concurrent heat and humidity, known as wet-bulb temperature) and lag effects. The relationship between a climate event and a SCA crisis may not be instantaneous; there could be a lag of several days, which current models often fail to account for systematically. Recent advances in spatiotemporal modelling further support this direction. Guo et al. [54] proposed a spatiotemporal delay correction model that integrates temporal and spatial dependencies to improve prediction accuracy in complex environmental settings. Although developed for geospatial remote sensing, the framework illustrates how high-resolution spatiotemporal modelling can improve climate exposure assignment for climate-health prediction in sickle cell disease.

Lack of explainability and interpretability in ML models

The black-box nature of many high-performing ML models, such as deep neural networks and complex ensemble methods, is a major barrier to their clinical adoption. Clinicians are rightfully hesitant to trust a model’s prediction without understanding the underlying rationale [55]. For a SCA crisis prediction model, it is not enough to know that a crisis is likely; a clinician needs to know why it is due to a sudden drop in temperature, a spike in air pollution, or a combination of patient-specific factors. Though Explainable AI (XAI) techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) exist, their application in SCA-climate studies is rare. The field lacks research that systematically uses XAI to uncover and validate the complex, non-linear interactions between climate stressors and physiological triggers in SCA, which is essential for generating clinically actionable insights and novel scientific hypotheses.

Limited deployment in low-resource and high-burden settings

Perhaps the most critical gap is the chasm between model development and real-world deployment, mainly in low-resource, high-burden regions. The computational resources required to train and run some advanced ML models can be prohibitive in regions with unreliable electricity and internet connectivity. There is a pressing need for the development of lightweight, efficient models that can operate on mobile devices or edge computing systems with minimal infrastructure [56]. Furthermore, successful deployment requires more than just a technically sound model. It necessitates user-centered design involving clinicians, patients, and public health officials to ensure the tool fits within existing clinical workflows and is accessible to the end-user. Studies that pilot the integration of these predictive tools into community health programs or clinical decision-support systems in LMICs are virtually non-existent, representing a significant and urgent direction for future work.

Discussions and findings

This section discusses the findings of the reviewed studies in relation to the research questions. The discussion is organized according to each research question to provide a structured interpretation of the evidence. The reviewed literature shows increasing use of artificial intelligence, machine learning, and predictive analytics in sickle cell disease (SCD). Most studies focused on predicting disease severity, vaso-occlusive crises (VOC), hospitalization, and mortality using clinical, laboratory, imaging, or physiological data. In contrast, only a limited number of studies incorporated climate or environmental information into predictive models. This imbalance indicates that climate-informed machine learning remains an emerging area of research.

Discussion and findings based on research questions

RQ1: How have machine learning models been applied to predict and assess the effects of climate factors on sickle cell disease outcomes, including vaso-occlusive crises, hospitalization, and mortality?

(i) Patterns observed in the reviewed studies

The reviewed studies show that machine learning has been widely applied to predict SCD outcomes using clinical and biomedical data. These studies provide the current baseline for machine learning applications in SCD but rarely integrate climate or environmental information. Common prediction tasks included disease severity, vaso-occlusive crises, hospitalization, and mortality.

Machado et al. [27] reported that Logistic Regression, Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks were frequently applied to clinical records, physiological measurements, and wearable sensor data. Okon et al. [24] also reported good predictive performance for Gradient Boosting and SVM models in VOC prediction using patient characteristics and real-time monitoring data. Salman et al. [37] found that Random Forest performed well when analysing heterogeneous clinical datasets because of its ensemble learning strategy. Similarly, Alim et al. [39], Deo et al. [40], and Çınarer [42] applied CNN, CNN-LSTM, and hybrid deep learning models to medical image analysis for SCD diagnosis and disease classification.

Only a small number of studies incorporated climate variables into predictive modelling. Desai et al. [45] examined the effects of ambient temperature, precipitation, wind speed, and temperature variation on VOC and acute chest syndrome (ACS) hospitalizations using negative binomial regression. Although the study did not employ machine learning, it provided evidence that climatic conditions influence important clinical outcomes in SCD. The limited number of climate-informed prediction studies highlights a clear gap between advances in machine learning for SCD and the integration of environmental exposure data into predictive models.

RQ 2: How were climate exposures assigned, and which climate and environmental variables were incorporated into machine learning models for sickle cell disease?

(ii) Climate exposure assignment and environmental indicators

The reviewed studies used several climate variables to examine sickle cell disease outcomes. These variables included temperature, humidity, rainfall, seasonality, wind speed, air quality, and environmental pollution indicators. Most epidemiological studies linked these variables with vaso-occlusive crises (VOC), hospital admissions, or mortality using meteorological records obtained from national weather agencies or publicly available climate databases. In most cases, climate exposure was assigned at the hospital, regional, or city level rather than to individual patients.

Dessap et al. [33] identified temperature and humidity as important predictors of hospital admission during SCA crises. Asnani et al. [34] reported that changes in temperature and humidity were associated with increased mortality among patients with SCA. Brindha et al. [8] found that lower temperature, rainfall, and humidity changes contributed to hypoxia and erythrocyte sickling. In contrast, Jones et al. [21] reported stronger associations with wind speed and low humidity than with temperature or rainfall.

The methods used to assign climate exposure differed across studies. Most studies matched climate records with hospital locations or broad geographic regions. Few studies linked environmental exposure to patient-level locations. Temporal exposure was also inconsistent. Some studies examined same-day weather conditions, whereas others used seasonal averages or longer observation periods. Lag effects between climate exposure and clinical events were rarely evaluated. Compound exposure indices that combined multiple environmental variables were also uncommon.

Machine learning studies showed limited integration of climate data. Machado et al. [27] reported that none of the reviewed prediction models included climate variables as predictors. Most models relied on clinical records, laboratory findings, imaging data, or physiological measurements. Similarly, Elsabagh et al. [44] focused on multimodal clinical and biomarker data without incorporating environmental exposure. These findings indicate that standardized spatiotemporal exposure assignment remains a major gap in climate-health modelling for sickle cell disease.

RQ 3: What methodological challenges and research gaps limit the integration of climate-health data into machine learning models for sickle cell disease?

(iii) Research gaps, methodological drawbacks, and implementation problems:

The reviewed literature suggested several critical research gaps and methodological limitations that hindered the development of reliable prediction models for SCA using climatic and clinical factors. For example, many studies employed small datasets obtained at a single clinical center, which resulted in a lack of generalizability and model overfitting [59]. Moreover, many studies used retrospective clinical datasets with poor external validity, reproducibility, and insufficiently documented preprocessing procedures and modeling approaches [52].

Another essential gap refers to superficial merging of the two datasets. Specifically, the authors of the paper by [53]. noted the mismatch between temporal and geographic scales of patient records and meteorological data that limits precise exposure estimation. Another problem related to this gap is limited attention paid to compound climatic conditions in addition to simple factors like temperature and humidity.

The inability of highly accurate AI models to be explainable and interpretable was another issue highlighted frequently. According to Amann et al. [55], clinicians were hesitant to rely on black box models that lacked a rationale for making decisions. While methods like SHAP and LIME can help make AI models more interpretable, they have not been applied in climate-SCA models. Moreover, infrastructure problems in resource-constrained environments, especially in Sub-Saharan Africa, prevent the practical implementation of advanced AI models [57].

RQ 4: Which machine learning models have been most frequently applied to predict sickle cell disease outcomes, and how do their predictive performances compare?

(iv) Dominant ML architectures and comparative performance:

According to the results obtained in section “Machine learning approaches for SCD prediction and climate-health integration”, supervised ML algorithms continue to be the most dominantly used methodologies in the SCA studies examined. These models comprise logistic regression, SVM, Random Forest, XGBoost, Decision Trees, AdaBoost, GBM, K-Nearest Neighbor (KNN), ENET, NB, and LDA. Ussher et al. [25] employed several ML methods along with hematologic, biochemical, and metabolic biomarkers in 481 people in Ghana to categorize disease severity. This study has employed dimensionality reduction and feature selection methods like PCA and PLS-DA before modeling. Other architectures, such as CNN, ResNet50, InceptionV3, LSTM, CNN-LSTM, GRU, and multimodal hybrid networks, were extensively reported for medical image analysis and time-series forecasting.

Comparative studies have shown that the deeper CNN architectures performed better hierarchical feature extraction and had superior diagnostic performance compared to the shallow architectures like VGG16 [41]. The results highlight that there is a growing trend towards the implementation of multimodal and Explainable Artificial Intelligence (XAI) systems that can integrate multiple sources of data. Recent developments also emphasize knowledge-guided neural networks that combine domain knowledge with deep learning. Shi et al. [58] proposed a knowledge-driven two-stage modulation network that improves feature interaction and semantic representation. Although developed for visual question answering, the architecture illustrates the growing trend toward integrating expert knowledge into deep learning, an approach that could improve interpretability and multimodal learning in climate-health prediction models.

Modeling best practices for climate-informed machine learning in sickle cell disease

The reviewed studies highlight several methodological practices that could improve the development and evaluation of climate-informed machine learning models for sickle cell disease (SCD). First, studies using longitudinal or time-series data should adopt time-aware training and testing strategies instead of random data partitioning. Random data splitting may introduce temporal information leakage and lead to overly optimistic performance estimates, whereas chronological validation provides a more realistic assessment of model performance in clinical practice [59].

Second, external validation should become a routine component of model evaluation. Most studies developed and tested their models using data from a single institution or geographic region, which limits model generalizability. Validation across multiple healthcare facilities and climatic regions would provide stronger evidence of model robustness and clinical applicability [27, 52]. Model evaluation should also extend beyond discrimination measures such as accuracy and the area under the receiver operating characteristic curve (AUC). Calibration measures, including the Brier score and calibration curves, should be reported because they assess how closely predicted probabilities agree with observed outcomes. Well-calibrated models are more appropriate for clinical decision support than models evaluated using discrimination metrics alone [59].

Performance should also be assessed across clinically relevant subgroups. Stratified analyses based on age, sex, SCD genotype, and geographic region can identify differences in predictive performance and reduce the risk of algorithmic bias. Such evaluations are important for developing equitable and clinically reliable decision-support systems [55, 60]. The assignment of climate exposure also requires greater methodological consistency. Future studies should evaluate lagged effects between environmental exposure and clinical outcomes because climatic conditions may influence vaso-occlusive crises several days before hospital presentation.

Greater attention should also be given to compound environmental indicators, including heat index, wet-bulb temperature, particulate matter (PM), dust concentration, and Harmattan-related conditions, rather than relying solely on individual weather variables. Improved exposure modelling would strengthen the integration of climate-health information into predictive machine learning frameworks [45, 53]. Adopting these practices would improve the reproducibility, generalizability, and clinical relevance of climate-informed machine learning models for sickle cell disease and support their translation into routine healthcare settings. In addition, machine learning models account for spatial heterogeneity and sparse observations. For example, Wang et al. [61] addressed these challenges by integrating heterogeneous geospatial information into a unified modelling framework for offshore geological characterization. Recent biomedical studies also show the value of integrating mechanistic knowledge with data-driven modelling. Ma et al. [62] linked molecular mechanisms with computational analysis to explain disease progression. Similar integration of biological mechanisms with climate and clinical variables may improve the interpretability and clinical relevance of machine learning models for sickle cell disease. Similar data-fusion strategies could improve the integration of clinical, environmental, and meteorological datasets in climate-informed sickle cell disease prediction.

Future directions

Towards explainable and interpretable ML at the climate health Nexus

Although machine learning (ML) models show strong potential in predicting crises in sickle cell Anemia (SCA) patients, many remain “black boxes” that clinicians struggle to integrate into practice. This lack of transparency reduces trust and limits clinical adoption [24]. Future research should therefore focus on explainable and interpretable ML approaches that clarify how climate factors, patient biomarkers, and clinical variables interact to influence crisis risk. Explainable artificial intelligence (XAI) tools such as SHAP and LIME can reveal feature importance and model reasoning, enabling clinicians to understand which climate and clinical factors contribute most to crisis risk and thereby improving trust and adoption of predictive systems [63].

Federated learning and privacy-preserving approaches for health data

A major limitation of current studies is the fragmented nature of patient data. Clinical records are often restricted to individual hospitals or healthcare regions. This fragmentation limits the development of robust and generalizable machine learning models. Privacy regulations and ethical concerns also constrain data sharing. The problem becomes more complex when sensitive clinical records are linked with environmental and climate data. Federated learning offers a practical solution. It enables multiple institutions to train machine learning models collaboratively without transferring raw patient data [64]. This approach keeps patient records within their original institutions. When combined with privacy-preserving techniques such as differential privacy and homomorphic encryption, federated learning further strengthens data security and protects patient confidentiality [65].

Ikegwu et al. [14] reviewed federated learning frameworks for healthcare cyber threat detection. The review discussed how differential privacy and homomorphic encryption support secure collaborative learning across distributed healthcare systems. The framework allows institutions to develop shared models without exposing sensitive patient information. The study also identified communication overhead, heterogeneous data distributions, and deployment in resource-constrained environments as major implementation challenges. These challenges should also be considered when developing climate-health prediction systems for sickle cell disease. Secure deployment remains an important requirement for collaborative healthcare systems. Li et al. [66] proposed an interaction-aware framework for detecting vulnerabilities in smart contract bytecodes, highlighting the importance of identifying security weaknesses before deployment. Although developed for blockchain systems, the principles of proactive vulnerability detection are relevant for protecting distributed healthcare platforms that support federated learning. Likewise, Li et al. [67] further showed how cross-system analysis can detect complex security threats that are difficult to identify using isolated inspection. Similar approaches may strengthen the security of distributed healthcare infrastructures that exchange climate and clinical information across institutions. Integrating data from multiple healthcare institutions can improve model robustness and population coverage.

Integrating IoT and wearable devices for real-time crisis prediction

Environmental factors can trigger sickle cell disease crises in real time. However, most existing studies rely on retrospective hospital records. Future research should incorporate Internet of Things (IoT) technologies and wearable devices that continuously monitor physiological indicators such as oxygen saturation, hydration status, heart rate, and body temperature. Wearable sensor systems have shown promise for real-time health monitoring and early risk detection [68]. Song et al. [69] showed that continuous physiological monitoring provides valuable insights into stress responses under changing environmental conditions. Similar continuous monitoring strategies could support wearable sensing systems for detecting physiological changes associated with climate-triggered sickle cell crises. Combining physiological measurements with local climate data, including temperature, humidity, rainfall, heat index, and air quality, could improve real-time crisis prediction. Such systems may support early intervention and shift patient care from reactive hospital treatment to proactive home- and community-based management (World Health Organization [70]).

Ikegwu et al. [71] reviewed embedded systems used in mobile health applications and examined their resilience against common cyber threats. The review identified security weaknesses in wearable devices and remote health monitoring platforms. It recommended stronger authentication mechanisms, encryption, and secure system design. These security measures should be incorporated into future climate-aware monitoring systems to ensure reliable, secure, and trustworthy healthcare services for people living with sickle cell disease.

Cross-disciplinary models: climate science, medicine, and AI collaboration

Understanding the relationship between climate variability and health outcomes in SCD requires collaboration across multiple disciplines. Climate scientists can provide high-resolution meteorological and environmental data. Clinicians and epidemiologists contribute knowledge of disease mechanisms, patient characteristics, and clinical outcomes. Data scientists and artificial intelligence researchers can integrate these diverse data sources into predictive models. Such collaboration improves the understanding of both individual-level risk factors and population-level climate exposures [20, 72]. Integrated climate-health models can also strengthen risk assessment and resilience planning for vulnerable populations [73]. Ikegwu et al. [74] reviewed artificial intelligence expert systems for environmental surveillance and disaster management. The review examined the use of machine learning, IoT, remote sensing, and expert systems for environmental monitoring and early warning applications. It also emphasized the integration of environmental intelligence with decision-support systems. A similar approach could strengthen climate-aware healthcare prediction by combining environmental information with clinical data to support timely intervention and informed decision-making for patients with sickle cell disease. Similar interdisciplinary approaches have been adopted in environmental and ecosystem research. For example, Wang et al. [75] combined ecological knowledge with network-based modelling to support climate-smart decision-making. Whereas Guo et al. [76] integrated environmental observations across multiple spatial scales to evaluate ecosystem responses to climate-related stressors. These studies highlight the value of combining heterogeneous environmental datasets, an approach that could strengthen climate-health prediction frameworks for sickle cell disease.

Path to policy translation and clinical decision support

The value of predictive modelling lies in its application to clinical practice and public health policy. Predictive models should be integrated into clinical decision support systems (CDSS) to help clinicians identify periods of increased risk and make timely treatment decisions [77]. Early warning systems could also support hydration campaigns, patient education, and resource planning during heat waves, harmattan seasons, and other extreme weather conditions [4].

At the policy level, climate-informed prediction tools can support more resilient healthcare systems. They can assist healthcare providers in preparing for periods of increased patient demand and help ensure uninterrupted access to essential medications during climate-related disruptions. These measures could reduce disease complications, lower hospital admissions, and improve the quality of life of people living with sickle cell disease [4]. Ikegwu et al. [78] reviewed artificial intelligence-based clinical decision support systems and examined the use of machine learning, deep learning, and expert systems for disease diagnosis, prognosis, and treatment planning. The review also emphasized the role of explainable artificial intelligence and multimodal data integration in improving clinician confidence and supporting informed decision-making. Mu et al. [79] integrated molecular biomarkers with microbiome information to investigate complex biological interactions. This work illustrates the growing importance of multimodal biomedical data integration, which may also enhance climate-informed prediction models for sickle cell disease. Peng et al. [80] combined resting-state and task-based neural networks using a multimodal fusion framework to improve prediction of antidepressant response. Although applied in neuroscience, the study illustrates how integrating heterogeneous biomedical data can improve predictive performance. Similar multimodal strategies could be adopted to combine climate, physiological, laboratory, and clinical information for sickle cell disease prediction. These advances provide a useful foundation for developing climate-informed clinical decision support systems that assist healthcare professionals in predicting and managing sickle cell disease more effectively.

Conclusion

This review shows that climate variability has a profound effect on the health of people living with sickle cell Anemia (SCA) across different regions. Factors such as extreme heat, cold weather, humidity changes, seasonal transitions, and poor air quality were consistently linked with an increase in vaso-occlusive crises and other complications. These environmental triggers amplify the underlying pathophysiology in SCA and contribute to patterns of hospital admissions and, in many cases, higher mortality.

The findings also reveal important gaps. Many studies focus on crises without addressing mortality outcomes, and most rely on descriptive analyses rather than advanced predictive approaches. This restricts the ability to identify the most critical risk factors or to generate actionable forecasts. For research, there is an urgent need for multi-site, interdisciplinary studies that combine clinical data with environmental datasets. Clinicians can also benefit from climate-aware strategies, such as targeted hydration during heat waves and closer monitoring during harmattan or cold spells. At this policy level, building resilient health systems that can withstand climate-related disruptions is vital to ensure uninterrupted access to treatment and support to protect vulnerable populations.

Hence, Machine Learning (ML) offers significant opportunities to bridge these gaps. By combining clinical records with real-time climate data, ML models can provide early warnings of crises, support timely interventions, and guide hospital preparedness. If designed with transparency, privacy safeguards. And collaboration across disciplines, such predictive tools, could become practical solutions that reduce morbidity and mortality. Ultimately, integrating ML into sickle cell care at the climate health nexus has the potential to transform outcomes and improve resilience for those most affected by climate change.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

M.O.K worked on the Introduction. I.C.O worked on the Methodology. F.O.U worked on the Climate-Health Nexus in Sickle Cell Anemia (SCA) worked on Machine Learning Approaches for Crisis Prediction and Mortality Risks. O.A worked on the privacy, ethics, and data challenges, research gaps and limitations of current studies. M.O.K worked on the discussions and findings. M.O.K and F.O.U worked on the future directions and conclusion. Finally, A.C.I worked on conceptualization, drafting of the original manuscript, and was involved in editing and reviewing the manuscript.

Funding

There is no external funding received by the authors.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Generative AI use

AI tools such as Quibot were used only for sentence refinement by the authors. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Competing interests

The authors declare no competing interests.