Explainable Machine Learning for Predicting Global Malaria Incidence From 2000 to 2022 With Climatic and Extreme Weather Conditions Across 78 Endemic Countries: A Retrospective Ecological Study

ABSTRACT

Background and Aims

Malaria remains a major global health concern, causing substantial morbidity and mortality, especially in tropical and subtropical regions with limited healthcare resources. Climatic conditions and extreme weather events are increasingly recognized as critical determinants of malaria transmission, yet their relative contributions remain poorly quantified. This study applied machine learning (ML) and explainable AI (XAI) to identify the key climatic and extreme weather determinants of malaria, assess their relative importance, and predict incidence.

Methods

Malaria incidence was analyzed across 78 endemic countries (2000–2022) using climatic and extreme weather variables from the World Health Organization (WHO), the EM‐DAT International Disaster Database, and the Global Data Lab. Random forest (RF) model was trained to predict malaria incidence, with performance assessed using the coefficient of determination (R ²), root mean square error (RMSE), and mean absolute error (MAE). To enhance interpretability and identify influential predictors, explainable AI (XAI) was applied using SHapley Additive exPlanations (SHAP).

Results

Malaria incidence declined steadily across the 78 endemic countries over the study period, remaining highest in Sub‐Saharan Africa. Random forest achieved the best overall predictive performance for malaria incidence (R ² = 0.88, 95% CI: 0.84–0.92; RMSE = 39.21, 95% CI: 35.29–43.13; MAE = 27.33, 95% CI: 23.41–31.25) with gradient boosting performing comparably. SHAP analysis identified the leading predictors associated with malaria incidence as average annual surface temperature and average annual precipitation, followed by storm‐ and flood‐affected population metrics. Climatic variables contributed substantially more to model predictions than disaster‐related indicators.

Conclusion

Random forest integrated with explainable AI provides an effective and interpretable framework for predicting malaria incidence and identifying its key climatic predictors, while quantifying the secondary contribution of extreme weather exposure. These findings can support evidence‐based, climate‐adaptive public health planning for malaria prevention and control.

1. Introduction

Malaria presents a significant and widespread public health challenge globally, particularly in countries with limited healthcare infrastructure and resources [1]. Despite decades of sustained international efforts to control and eliminate the disease, malaria continues to afflict hundreds of millions of people each year, imposing a substantial burden of morbidity, mortality, and socioeconomic loss [2]. The World Health Organization (WHO) reported an estimated 282 million malaria infections and 610,000 malaria‐related deaths across 80 countries in 2024 [1]; the present study analyzed 78 of these, excluding two for which complete climatic or extreme weather data were unavailable over the study period. The disease disproportionately affects children under 5 years of age and pregnant women, particularly across Sub‐Saharan Africa, where approximately 95% of global cases and deaths are concentrated [1]. At the country level, Nigeria, the Democratic Republic of the Congo, Uganda, and Mozambique collectively account for more than half of all global malaria cases, while India and Myanmar bear a considerable burden of Plasmodium vivax malaria in the South Asian region [1].

Malaria transmission is governed by a complex interplay of biological, environmental, socioeconomic, and healthcare‐related determinants [3]. Among these, climatic variables have emerged as particularly influential variables associated with vector ecology and disease dynamics. Temperature modulates the sporogonic development rate of Plasmodium parasites and the survival and biting behavior of Anopheles mosquitoes, with transmission peaking at an intermediate optimum of approximately 25°C and declining at higher and lower temperatures, rather than rising steadily across the 25°C–30°C range assumed in earlier work [4]. Rainfall is a leading predictor of the availability of larval breeding habitats, and its seasonality shapes the temporal patterns of transmission intensity across endemic regions [5]. Beyond average climatic conditions, extreme weather events, including floods and tropical storms, can expand mosquito breeding sites, displace vulnerable populations, disrupt local healthcare delivery, and amplify transmission risk in already‐endemic settings [6]. In Sub‐Saharan Africa, recurrent flood episodes have been associated with measurable increases in malaria incidence in the weeks following inundation [7], while El Niño‐associated climatic anomalies have been linked to malaria epidemics across parts of Latin America [8].

Traditional epidemiological approaches to studying climate–malaria relationships have relied largely on Poisson and negative binomial regression, time‐series analysis, and GIS‐based spatial mapping [9, 10]. These methods have generated important insights, including rainfall‐lag models and temperature‐threshold frameworks that link climatic conditions to transmission risk [11]. However, their capacity to capture nonlinear, high‐dimensional interactions among simultaneously acting determinants remains fundamentally limited; direct comparisons in malaria settings have shown that ensemble machine‐learning models substantially outperform linear and count‐regression baselines in predictive accuracy [12]. Large‐scale geospatial efforts have mapped malaria burden across many countries [2, 12], but these have focused primarily on environmental suitability and intervention coverage, and have typically emphasized average climatic conditions. Consequently, the dynamic and disruptive effects of extreme weather events and the human populations they affect remain comparatively underexplored within unified, interpretable, multi‐country modeling frameworks.

The integration of machine learning (ML) and explainable artificial intelligence (XAI) provides a flexible analytical framework to address these limitations, enabling the simultaneous incorporation of heterogeneous datasets and the discovery of complex, nonlinear relationships that conventional regression approaches may fail to capture [13]. ML models such as regularized regression, K‐nearest neighbors, decision trees, random forest, and gradient boosting have shown strong predictive performance for malaria and other climate‐sensitive diseases [13, 14]. To interpret these models, we apply SHapley Additive exPlanations (SHAP), which, unlike random forest's built‐in importance, provides a consistent measure of both the magnitude and the direction of each predictor's effect at the global and individual levels [15]. Although previous multi‐country studies have examined associations between climate variables and malaria incidence, few have integrated both climatic and extreme weather indicators, including disaster frequency, severity, and affected‐population metrics within a unified, long‐term ML and XAI framework spanning multiple countries.

This study, therefore, investigates the association of climatic variables and extreme weather indicators on malaria incidence across 78 endemic countries from 2000 to 2022, while comparing the predictive performance of multiple statistical and machine learning models. Specifically, the objectives are to (i) identify the most influential climatic and extreme weather determinants shaping malaria patterns across diverse geographic and epidemiological contexts; (ii) assess the contribution of floods, storms, and affected‐population metrics to malaria transmission dynamics.

2. Materials and Methods

2.1. Study Scope and Data Collection

This study examines malaria incidence across 78 countries from 2000 to 2022, spanning diverse climatic regions. It evaluates the association of temperature, precipitation, and relative humidity, together with extreme weather events, particularly floods and storms, while incorporating affected‐population metrics, with the aim of identifying spatial and temporal patterns in malaria burden and generating evidence to inform targeted, climate‐adaptive public health interventions.

Malaria, climatic, and extreme weather data were collected from the World Health Organization (WHO) [16], the EMDAT International Disaster Database [17], and the Global Data Lab [18], encompassing 78 countries over the period 2000–2022. The dataset integrates malaria indicators with key environmental drivers, including extreme weather events (floods, storms, droughts, and heatwaves), the populations they affected, and core climatic variables such as temperature, precipitation, and relative humidity (Table 1).

Table 1.

Summary of variables, codes, and data sources for the malaria incidence analysis across 78 countries (2000–2022).

Category	Variable	Code	Source
Spatial variables	Location	Location	78 countries
	Period	Period	2000–2022
Malaria indicator	Estimated malaria incidence (per 1000 population at risk)	Malaria incidence	WHO [1]
Extreme weather factors	Flood: annual average affected area (km2/year)	x1	EM‐DAT [17]
	Storm: annual average wind speed (kph/year)	x2	EM‐DAT [17]
	Drought: annual average affected area (km2/year)	x3	EM‐DAT [17]
	Heatwave: annual occurrence (0 = no, 1 = yes)	x4	EM‐DAT [17]
	Flood: total population affected per year	x5	EM‐DAT [17]
	Storm: total population affected per year	x6	EM‐DAT [17]
	Drought: total population affected per year	x7	EM‐DAT [17]
	Heatwave: total population affected per year	x8	EM‐DAT [17]
Climate factors	Average annual surface temperature (°C)	x9	GlobalData Lab [18]
	Average annual precipitation (mm)	x10	GlobalData Lab [18]
	Relative humidity (%)	x11	GlobalData Lab [18]

2.2. Ethics Statement

This study utilized publicly available, aggregated, de‐identified secondary data from the World Health Organization (WHO), the EM‐DAT International Disaster Database, and the Global Data Lab. No human subjects were directly involved, and no identifiable personal or patient‐level data were accessed. Therefore, ethical approval from an institutional review board was not required for this secondary data analysis. All data sources are openly accessible and cited appropriately.

2.3. Data Preprocessing and Analysis

Missingness was assessed per variable. Overall, missing values were limited (<5%) but unevenly distributed: climatic variables (temperature, precipitation) were near‐complete, whereas some EM‐DAT extreme weather indicators showed higher missingness (Table S1). Importantly, country‐years with no recorded flood, storm, drought, or heatwave were treated as genuine zero‐event observations (coded 0) rather than as missing, to avoid conflating non‐occurrence with data loss; only true random gaps in continuous variables were imputed using K‐nearest neighbors (KNN,k = 5) [19], with linear interpolation to preserve temporal continuity. As EM‐DAT provides systematic global disaster coverage, absent entries were interpreted as non‐occurrence rather than missing data. Continuous variables were then normalized using Z‐score transformation [20]. Multicollinearity was assessed using the Variance Inflation Factor (VIF) [21]; a standard threshold of 10 was adopted to reduce redundancy among correlated climatic predictors, and relative humidity (x11) far exceeded it (VIF = 20.31) and was excluded, while all retained variables had VIF < 10 (Tables S1–S3). Outliers were detected using Mahalanobis distances computed with values exceeding the 95% confidence threshold flagged as extreme [22]. Data processing and analysis were conducted in Python (v3.13) [23], using scikit‐learn [24] for KNN imputation and modeling [19], together with Matplotlib [25], Seaborn [26], and GeoPandas [27] for visualization and spatial analysis (Figure 1).

Figure 1.

Study framework illustrating the analytical pipeline: (1) Data collection from WHO, EM‐DAT, and Global Data Lab; (2) preprocessing including missingness assessment, KNN imputation, Z‐score normalization, and VIF‐based variable selection; (3) model training and hyperparameter tuning for five machine learning models; and (4) SHAP‐based explainability analysis applied to the best‐performing model.

2.4. Model Selection and Evaluation

This study combined machine learning (ML) modeling with XAI. Five models a generalized linear model with elastic‐net regularization (GLM), K‐nearest neighbors (KNN), a decision tree (DT), random forest (RF), and a gradient boosting machine (GBM) were implemented and compared, with GLM serving as the baseline (Figure 1). These models were selected to capture both linear relationships and complex nonlinear interactions among climatic and extreme weather variables [28]. GLM provided a robust statistical framework for linear associations but was limited in handling nonlinear dependencies. KNN offered an intuitive, instance‐based approach for detecting local patterns, although its performance was sensitive to the choice of k and to feature scaling. Decision trees captured hierarchical, nonlinear relationships but were prone to overfitting on complex data. Random forest improved predictive stability and reduced overfitting through ensemble averaging, at some cost to interpretability. Gradient boosting sequentially minimized prediction error and modeled intricate feature interactions, making it well‐suited to environmentally driven malaria dynamics.

Model performance was evaluated using the coefficient of determination (R ²), root mean square error (RMSE), and mean absolute error (MAE) to assess predictive accuracy, reliability, and generalizability. Following model selection, SHAP were applied to the best‐performing model to interpret its predictions and quantify the contribution of each climatic and extreme weather variable to malaria incidence (Figure 1).

2.5. Cross‐Validation and Hyperparameter Tuning

The dataset was divided into a training set (2000–2019) and a test set (2020–2022), with 80% of countries used for model training and the remaining 20% reserved for testing to assess geographic transportability. The data partitioning strategy is presented in Table S3. A fivefold cross‐validation was applied exclusively within the training set for hyperparameter optimization (Tables S4, S5). To prevent data leakage during cross‐validation, Z‐score normalization, and KNN imputation parameters were estimated within each training fold and applied to the corresponding validation fold. VIF‐based variable selection was performed once on the full training set prior to model fitting. For random forest, the optimal configuration was: n_estimators = 300, max_depth = 10, min_samples_split = 5, and min_samples_leaf = 2 (random_state = 42). For gradient boosting: n_estimators = 200, learning_rate = 0.1, and max_depth = 3 (random_state = 42). All models were evaluated using R ², RMSE, and MAE. The 95% confidence intervals for the test‐set performance metrics were estimated using non‐parametric bootstrap resampling of the test‐set observations, with the corresponding percentile intervals reported in Table S11.

3. Results

3.1. Temporal Trends in Malaria Incidence (2000–2022)

Across the study period, average malaria incidence declined substantially and steadily, falling from 124.29 per 1000 population in 2000 to 71.03 in 2022 an overall reduction of approximately 43% (Table S6; Figure S1). The decline was steepest during the early 2000s (124.29 in 2000 to 105.79 in 2005) and continued more gradually through the 2010s, reaching 80.20 by 2015. Thereafter the downward trend slowed, with minor year‐to‐year fluctuations a temporary rise to 82.76 in 2017 and a slight uptick from 74.57 in 2019 to 75.34 in 2020 before continuing to its lowest value of 71.03 in 2022. This trajectory is consistent with the broader global decline in malaria incidence over the period, although the pace of reduction was uneven across years and, as shown below, across countries (Figure 2).

Figure 2.

(A) Global distribution of average malaria incidence (2000–2022); (B) spatial patterns of flood‐ and storm‐affected populations and associated malaria burden; (C) top 10 countries by year‐wise average malaria incidence; (D) year‐wise global average malaria incidence and the number of affected countries.

Analysis of extreme weather exposure provides additional context. The most populous countries China, India, the Philippines, Bangladesh, and Pakistan, accounted for the largest cumulative numbers of people affected by floods and storms between 2000 and 2022, with China alone exceeding 1.14 billion affected individuals, followed by India (411 million), the Philippines (167 million), and Bangladesh (128 million) (Figure 2; Table S7). Notably, these countries are not among the highest malaria‐burden settings, indicating that large absolute extreme weather exposure does not translate directly into high malaria incidence. This apparent decoupling likely reflects the confounding association of population size, economic development, and health‐system capacity, and is examined further in Section Conclusion, 4.

3.2. Country‐Level Variation in Malaria Incidence

Average malaria incidence varied markedly across the 78 countries. Burkina Faso exhibited the highest burden (485.6 per 1000 population), separated by ~98 points from the next tier—the Central African Republic (387.23), Mali (386.71), Mozambique (386.07), and the Democratic Republic of the Congo (382.77), which themselves spanned a narrow ~4‐point range (Tables S8–S10; Figure S2). These WHO‐derived values are central estimates subject to the uncertainty intervals inherent to modeled malaria surveillance and should be interpreted accordingly rather than as exact quantities. The distribution, therefore, comprises a tightly clustered group of high‐burden countries alongside a single, pronounced positive outlier. Because such an outlier could disproportionately shape global feature attributions, we repeated the SHAP analysis with Burkina Faso excluded; the ordering and relative magnitude of the leading predictors were essentially unchanged, (full dataset: temperature 40.75%, precipitation 35.34%, storm‐affected population 13.36%; after Burkina Faso exclusion: temperature 40.8%, precipitation 35.2%, storm‐affected population 13.4%), confirming that the model's predictive associations with climatic and extreme weather variables are robust to this influential observation and not an artifact of a single high‐incidence country.

3.3. Machine Learning Model Performance

Among the evaluated models, Random Forest performed best, with the highest R ² = 0.88 (0.84, 0.92), the lowest RMSE = 39.21(35.29, 43.13), and a low MAE = 27.33 (23.41, 31.25), making it the most reliable model for predicting malaria incidence. Gradient boosting followed closely, with a comparable R ²= 0.86 (0.82, 0.90) and RMSE = 42.21 (38.29, 46.13) and an MAE = 34.33 (30.41, 38.25); given the near‐identical accuracy of the two ensembles, random forest was selected based on its lower average cross‐validation RMSE (41.48 vs. 49.54) and more consistent fold‐level performance (Table S5). GLM with elastic net showed strong predictive performance, but KNN and the decision tree showed comparatively lower predictive performance. Overall, ensemble methods outperformed the simpler models in capturing the complex relationships in the data. The random forest model was subsequently integrated with XAI to enhance interpretability and improve the transparency of the results (Table S11; Figure 3B).

Figure 3.

(A) Model performance metrics on the training data (R ², RMSE, MAE); (B) model performance metrics on the test data (R ², RMSE, MAE); (C) SHAP beeswarm plot for the 10 retained predictors; each point represents one country–year observation, plotted on the x‐axis by its SHAP value (contribution to predicted malaria incidence, in z‐scored units); color indicates the feature value (red = high, blue = low); (D) normalized SHAP contribution (%) for each predictor, calculated as each predictor's mean absolute SHAP value divided by the sum across all 10 predictors.

3.4. Feature Importance in Predicting Malaria Incidence

Feature‐importance analysis was based on mean absolute SHAP values, which quantify each predictor's average contribution to the standardized (z‐scored) model output expressed as the mean deviation it produces from the baseline prediction. On this scale, average annual surface temperature was the most influential predictor (mean |SHAP | = 48.95), followed by average annual precipitation (42.44) and storm‐affected population per year (16.05), indicating that climatic conditions play a dominant role. Extreme weather indicators contributed less: flood‐affected population (9.02), flood annual average affected area (2.18), and drought‐affected population (1.01), while storm wind speed (0.45), drought‐affected area (0.009), heatwave occurrence (0.008), and heatwave‐affected population (0.001) were negligible (Figure 3C; Table S12).

To express relative importance, each predictor's mean absolute SHAP value was normalized by the sum of mean absolute SHAP values across all 10 retained predictors; the resulting contributions sum to 100%, with no variables excluded from the reported list (near‐zero contributors are retained but rounded to ≤0.01%). Temperature had the largest relative SHAP attribution (40.75%), followed by precipitation (35.34%) and storm‐affected population (13.36%), with flood‐affected population (7.51%) and flood‐affected area (1.81%) contributing modestly and the remaining disaster indicators together contributing <1.3%. These results confirm that climatic conditions are the leading predictors of malaria incidence, with extreme weather exposure led by storm‐ and flood‐affected populations acting as a secondary predictive importance (Figure 3D; Table S13).

Beyond magnitude, the SHAP summary plot indicated the direction of each effect. Higher average annual surface temperatures were associated with higher predicted incidence, and higher average annual precipitation with higher predicted incidence; among extreme weather indicators, larger storm‐ and flood‐affected populations were associated with lower predicted incidence. These directional patterns show not only which factors matter most, but how they were associated with model‐predicted malaria incidence (Figure 3).

4. Discussion

In malaria research, many studies have focused solely on climatic conditions, neglecting the dynamic and disruptive effects of extreme weather events and the human populations they affect [5, 29]. In this study, we analyzed both climatic variables and extreme weather indicators, including the populations affected by such events to examine malaria incidence across 78 endemic countries over two decades. Our findings show that extreme weather exposure contributes measurably to malaria incidence, but its influence remains secondary to that of baseline climatic conditions, particularly temperature and precipitation.

Across malaria‐endemic regions, the highest average incidence rates were consistently observed in Sub‐Saharan African countries [2]. The Democratic Republic of the Congo, Mozambique, Mali, the Central African Republic, and Burkina Faso exhibited persistently high burdens, with Burkina Faso showing the highest average incidence [30]. This clustering reflects a combination of favorable climatic conditions for mosquito development, limited healthcare infrastructure, and low adaptive capacity [3]. These baseline conditions, however, do not fully explain inter‐country variation in malaria incidence, motivating a data‐driven examination of the predictive patterns.

The random forest model, interpreted through SHAP, provides a clear hierarchy of these predictors. Average annual surface temperature and precipitation were the two most important predictors by a wide margin, together accounting for roughly three‐quarters of the total SHAP contribution [31, 32]. This is consistent with the well‐established sensitivity of Anopheles survival, biting activity, and the sporogonic development of Plasmodium parasites to ambient temperature, and of larval breeding‐habitat availability to rainfall [5, 33]. Temperature is strongly associated with the pace of parasite development and mosquito longevity, while precipitation is a key predictor of the formation and persistence of breeding sites; their joint prominence reinforces the central role of climatic suitability in defining where and when malaria transmission is most intense [4, 33].

Extreme weather exposure contributed less than the core climatic variables, but was not negligible [7]. Among the disaster‐related indicators, storm‐ and flood‐affected population metrics ranked highest, well above the corresponding hazard‐magnitude measures such as wind speed and affected area. This pattern suggests that, where extreme events influence malaria dynamics, it is the scale of human exposure and the resulting disruption rather than the raw physical intensity of the event that carries predictive weight [34]. Plausible pathways include population displacement, overcrowding in temporary shelters, the breakdown of sanitation and vector‐control programs, and reduced access to diagnosis and treatment, all of which can increase mosquito–human contact and delay care in the weeks following a storm or flood [35].

Drought‐ and heatwave‐related variables showed weak direct importance in the model. Their limited contribution does not necessarily preclude indirect effects we hypothesize, for example, that drought and heat may influence transmission through changes in water‐storage behavior, human mobility, or ecological stress on vector habitats, but these pathways were not tested here and remain speculative, warranting dedicated future investigation. Within this dataset, the association of drought and heatwave indicators was clearly overshadowed by storm‐ and flood‐related exposure and, more strongly, by the core climatic variables.

It is also notable that the countries with the largest absolute extreme weather exposure, China, India, the Philippines, Bangladesh, and Pakistan are not the highest malaria‐burden settings globally [17]. This pattern may partly reflect differences in national income, health‐system capacity, surveillance, and reporting practices. Countries with greater resources may have stronger malaria surveillance and control systems, while differences in disaster reporting and surveillance capacity may contribute to variation in recorded extreme‐weather exposure. These differences may influence the estimated variable‐importance rankings and could partly explain why disaster indicators ranked below the core climatic variables. In malaria‐endemic African regions, where health‐system constraints are often greater, extreme‐weather exposure may show different predictive patterns in relation to malaria incidence [7, 36]. As a predictive, non‐causal model, our framework cannot fully disentangle these effects.

Taken together, the feature‐importance analysis points to a layered understanding of malaria dynamics: climatic suitability captured chiefly by temperature and precipitation is the dominant predictor of baseline transmission potential, while extreme weather–related population exposure, led by storm‐ and flood‐affected populations, modulates that potential, particularly in vulnerable, resource‐limited settings. This framing complements purely climate‐driven models by adding a human‐exposure dimension, while keeping climate as the principal predictor identified by the data.

4.1. Limitations

Several limitations should be acknowledged. The analysis used retrospective, country‐level annual data, so national aggregation of climatic variables may mask important sub‐national and seasonal variation. Although model performance was evaluated using a temporal holdout test set (2020–2022), fivefold cross‐validation for hyperparameter tuning used random splitting within the training data rather than temporally blocked folds, which may not fully account for temporal dependence in repeated country–year observations. The analytical framework also did not explicitly model country‐level heterogeneity, temporal autocorrelation, or spatial dependence. Although the 80% country‐level training and 20% test split provided an assessment of geographic transportability, leave‐one‐country‐out or repeated grouped validation could provide a more stringent assessment in future studies.

The 2020–2022 test period overlapped with the COVID‐19 pandemic, potentially affecting malaria surveillance and service delivery. Important determinants, including ITN coverage, IRS, ACT access, health‐system capacity, surveillance quality, and socioeconomic and demographic factors, were not comprehensively included because of data availability and comparability constraints. Their omission may influence the estimated importance of climatic and extreme‐weather predictors and may introduce residual confounding. Finally, SHAP quantifies predictive importance rather than causal effects; therefore, the findings should be interpreted as predictive associations rather than causal relationships. Future studies using higher‐resolution longitudinal data and causally informed approaches could further investigate these relationships.

5. Conclusion

This study characterized the relative contributions of climatic conditions and extreme weather exposure to malaria incidence across 78 endemic countries over two decades. Its novelty lies in jointly modeling both sets of predictors, including disaster frequency, severity, and affected‐population metrics within a single interpretable machine‐learning framework, and using SHAP to show that, at the annual country scale, average temperature and precipitation dominate predicted incidence (together ~76% of importance), while storm‐ and flood‐affected populations contribute meaningfully but secondarily. The strong predictive associations identified here suggest that temperature‐ and rainfall‐based seasonal forecasts may have potential for integration into malaria early‐warning systems in high‐suitability African settings, contingent on validation using intervention data. Importantly, the framework is predictive rather than causal; it identifies and ranks the environmental correlates of malaria but cannot establish causal effects, and its country‐level annual resolution omits sub‐national, seasonal, and intervention‐coverage dynamics. Moving from prediction toward decision‐ready, causal guidance will require higher‐resolution data, explicit intervention variables, and formal causal‐inference methods.

Author Contributions

Md. Abu Bokkor Shiddik: conceptualization, funding acquisition, writing – original draft, investigation, methodology, validation, visualization, writing – review and editing, software, formal analysis, project administration, data curation, supervision, resources. Md. Abdullah Al Fahad Sohan: writing – original draft, visualization, data curation, software, writing – review and editing, formal analysis.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Transparency Statement

The lead and corresponding author (Md. Abu Bokkor Shiddik) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.

Supporting information

Figure S1: Year trend: Average malaria incidence.

Figure S2: Top countries by average malaria incidence (x‐axis represents malaria incidence; y‐axis represents countries).

Table S1: Variance inflation factor (VIF) for predictor variables.

Table S2: After X11 excluded variance inflation factor (VIF) for predictor variables.

Table S3: Model performance comparison including high VIF variable (x11).

Table S4: Hyperparameter tuning of the machine learning models.

Table S5: Cross validation RMSE.

Table S6: Global average malaria incidence year trend, 2000–2022.

Table S7: Country‐wise flood and storm affected populations.

Table S8: Global average malaria incidence country‐wise.

Table S9: Top 10 countries by average malaria incidence, 2000–2022.

Table S10: Country and year‐wise malaria incidence rate.

Table S11: Model performance comparison: Evaluation matrix.

Table S12: Mean SHAP values of predictor variables.

Table S13: Contribution % of mean SHAP values.

Data Availability Statement

Data is available from https://github.com/AbdullahAlFahad01/Research-Publications.