Machine Learning Forecasts African Malaria Risk but Shows Geographic Transfer Limits

A comprehensive retrospective investigation across 47 African countries demonstrated that machine-learning models trained on integrated climatic, demographic, and socioeconomic indicators can accurately classify future national malaria risk, although their predictive performance declines noticeably when applied to entirely unrepresented geographic regions. Covering the quarter-century period from 2000 through 2024, the study integrated continental disease burden metrics with routine environmental observations and World Bank development indicators to build and rigorously validate predictive systems. Across the entire 25-year timeline, the historical dataset accounted for a cumulative total of 5.63 billion clinical malaria cases and 17.78 million malaria-related deaths. While estimated deaths fell from an early peak of 842,280 in 2003 to 652,613 in 2024, total clinical cases increased over the same span from 204.3 million to 244.7 million, with an intervening continental high of 245.4 million cases recorded in 2020. This persistent disconnect between declining mortality and rising absolute infection counts highlights ongoing challenges for continental public health interventions.

To evaluate whether computational frameworks could anticipate disease shifts, researchers assembled a harmonized country-year dataset encompassing 1,056 observations drawn from 44 African countries where climate, epidemiological, and socioeconomic series aligned continuously from 2001 to 2024. The epidemiological outcomes were established using country-level clinical case and mortality figures derived from the Malaria Atlas Project rather than crude national surveillance notifications, standardizing incidence as cases per 1,000 population and mortality as deaths per 100,000 population. Climate inputs were obtained from the NASA Prediction of Worldwide Energy Resources platform, capturing annual averages and annual ranges for air temperature, precipitation, relative humidity, and surface wind speed. Concurrently, broader population, infrastructure, and economic dimensions were assembled from the World Bank World Development Indicators database. Three countries—Cabo Verde, Comoros, and Sao Tome and Principe—were not represented in the final machine-learning modeling set due to incomplete environmental linkages across the study span, leaving 44 sovereign jurisdictions for statistical learning.

To avoid data leakage and prevent optimistic performance overestimation, the investigation implemented a strict chronological partition between model formulation and final testing. The initial 880 country-year observations spanning 2001 through 2020 formed the exclusive development dataset, while the final 176 country-year observations from 2021 to 2024 were withheld as an independent, out-of-time evaluation cohort. The binary target for classification was constructed around the national median malaria incidence, which was computed strictly within historical training folds and finalized at 220.864 cases per 1,000 population across the complete development baseline. candidate predictors that presented missing rates exceeding 50 percent during the development phase were systematically excluded under a prespecified missingness rule, which removed the national poverty headcount indicator and the Gini index. Across the remaining 28 variables, missing values were low—led by domestic health expenditure per capita at 4.20 percent—and all imputation medians as well as feature scaling transformations were calculated solely from training observations before being projected onto future test sets.

Model selection utilized four expanding-window temporal folds in which candidate hyperparameter sets for six distinct supervised algorithms were evaluated on future validation blocks using the area under the receiver operating characteristic curve as the selection criterion. The evaluated model architectures comprised traditional baseline logistic regression, a single decision tree, Random Forest, a radial-basis support vector machine, a gradient boosting machine, and extreme gradient boosting. Across the leakage-free expanding validation blocks, the Random Forest model attained the highest mean temporal cross-validation area under the curve of 0.968, whereas the baseline logistic regression and decision tree configurations yielded lower values of 0.930 and 0.876, respectively. After hyperparameter optimization identified fixed configurations—such as an mtry of 3 with 500 trees for Random Forest, and a cost parameter of 32 for the radial support vector machine—the full 2001–2020 cohort was refitted once and locked before any scoring on the withheld test observations commenced.

When evaluated against the independent 176 out-of-time test observations between 2021 and 2024, the four principal nonlinear machine-learning architectures showed high and closely matched discriminatory ability. The support vector machine achieved the largest point estimate for the area under the receiver operating characteristic curve at 0.955, accompanied by an accuracy of 0.886, a specificity of 0.935, and a precision-recall area under the curve of 0.929. The Random Forest model closely followed with a receiver operating characteristic area of 0.950, securing the highest point estimates for balanced accuracy at 0.902, overall accuracy at 0.903, sensitivity at 0.897, and an F1-score of 0.878. The gradient boosting machine reached an area under the curve of 0.949 with 0.892 accuracy, while extreme gradient boosting recorded an area under the curve of 0.943 with 0.875 accuracy. In contrast, linear logistic regression and the pruned decision tree exhibited lower discrimination, recording areas under the curve of 0.926 and 0.840, respectively.

Because the discriminatory point estimates of the four leading algorithms were closely clustered, researchers applied a country-cluster bootstrap procedure with 2,000 resamples to characterize the statistical uncertainty around test metrics while preserving the grouped four-year structure within each sovereign state. Pairwise difference testing demonstrated that the modest performance margins among the four principal models were not statistically distinguishable. The bootstrap 95 percent confidence interval for the receiver operating characteristic difference between the support vector machine and Random Forest spanned from -0.024 to 0.032, centered at a point difference of 0.004. Similarly, the margin between Random Forest and the gradient boosting machine yielded an interval spanning -0.013 to 0.018 around a difference of 0.002, while the comparison between the support vector machine and extreme gradient boosting produced an interval from -0.015 to 0.044. Because every pairwise bootstrap interval crossed zero, the evidence indicated that no single principal algorithm established definitive superiority over the others.

An inspection of the underlying classification confusion matrices on the independent test period illustrated how the leading architectures distributed operational errors across the 68 high-risk and 108 low-risk national observations. The Random Forest model correctly identified 61 of the 68 high-risk country-years while misclassifying only 7 high-risk periods as low risk, in addition to correctly identifying 98 low-risk instances. The support vector machine correctly detected 55 high-risk observations and 101 low-risk observations, securing the highest number of correct low-risk determinations among all evaluated systems. Gradient boosting correctly categorized 59 high-risk and 98 low-risk events, extreme gradient boosting identified 57 high-risk and 97 low-risk events, and logistic regression correctly designated 51 high-risk and 100 low-risk observations. These performance profiles confirm that while high-risk detection remained strong across models, individual algorithms favored marginally different balances between false-positive alerts and missed high-risk outbreaks.

Beyond binary threshold discrimination, model calibration checks revealed that raw algorithmic probability estimates departed from ideal statistical alignment. Regressing observed outcomes on the logit of model-predicted probabilities yielded an intercept of -0.389 and a slope of 1.258 for Random Forest, alongside an overall Brier score of 0.090, representing the lowest overall mean squared probability error among the group. The support vector machine produced a calibration intercept of 0.521 and a slope of 0.614, while gradient boosting yielded an intercept of -0.290 and a slope of 0.741, and extreme gradient boosting recorded an intercept of -0.272 and a slope of 0.597. Calibration plots revealed systemic departures from the ideal 45-degree line across probability deciles, demonstrating that strong relative ranking of high- and low-risk environments does not automatically yield perfectly calibrated absolute risk probabilities without post-hoc recalibration.

The study demonstrated marked performance degradation when models were subjected to internal grouped-country cross-validation, where five deterministic folds withheld entire sovereign territories from training rather than temporal blocks. Under this geographic separation, pooled areas under the curve dropped to 0.874 for Random Forest, 0.870 for extreme gradient boosting, 0.843 for gradient boosting, and 0.839 for the support vector machine, with logistic regression yielding 0.814. Pooled classification accuracy likewise decreased to 0.784 for both Random Forest and extreme gradient boosting, accompanied by elevated Brier scores of 0.145 and 0.168. Furthermore, fold-specific performance showed substantial geographic volatility, with the Random Forest area under the curve swinging from 0.770 in difficult country groups to 0.945 in others. Across the 44 test countries, the support vector machine achieved four-of-four correct yearly classifications in 36 countries, whereas Random Forest, gradient boosting, and extreme gradient boosting each produced perfect four-year records in 34 countries.

Extensive sensitivity analyses verified that the predictive capacity of the machine-learning frameworks was not merely an artificial consequence of dichotomizing malaria incidence at the median. Shifting the classification threshold to the 60th percentile resulted in an area under the curve of 0.942, an accuracy of 0.875, and an F1-score of 0.711 for Random Forest, whereas shifting the cut-off to the 75th percentile yielded an apparent accuracy of 0.955 but caused the F1-score to collapse to 0.200 due to severe class imbalance. Complementing the classification tasks, a dedicated Random Forest continuous regression model fitted to log-transformed incidence achieved an R-squared of 0.959 on the independent 2021–2024 test cohort. When back-transformed to the natural incidence scale, the continuous model generated a root mean squared error of 43.69 cases per 1,000 population and a mean absolute error of 31.63 cases per 1,000 population, confirming that a robust underlying predictive signal persisted across continuous metrics.

Interpretability assessments using permutation feature importance and Shapley Additive Explanations revealed that model forecasts relied on a joint combination of environmental dynamics and socioeconomic conditions rather than any single dominant category. In the Random Forest model, permutation importance ranked one-year lagged wind speed as the leading predictor with a score of 0.068, followed by annual wind speed at 0.059, one-year lagged temperature at 0.051, unemployment rate at 0.043, and annual temperature at 0.043. Similarly, global Shapley Additive Explanations for extreme gradient boosting identified annual wind speed as the top predictor with a mean absolute value of 1.115, followed by lagged wind speed at 0.879, life expectancy at 0.614, lagged temperature at 0.611, the age-dependency ratio at 0.607, unemployment rate at 0.511, annual temperature at 0.406, and gross domestic product per capita at 0.359. Monthly precipitation variability, quantified by the monthly coefficient of variation, also ranked prominently among predictive factors.

Spatial mapping of population-standardized burden over the 2000–2024 period underscored vast geographic variation across Africa, explaining why national demographic adjustments are necessary when evaluating transmission intensity. Burkina Faso exhibited the highest mean annual malaria incidence at 468.55 cases per 1,000 population, followed by Benin at 404.31, the Central African Republic at 389.51, Sierra Leone at 378.99, Mozambique at 371.87, the Democratic Republic of the Congo at 364.18, and Nigeria at 363.27. In sharp contrast, northern and southern African nations recorded near-zero or low burdens, including Morocco at 0.00, Algeria at 0.00014, Cape Verde at 0.28, South Africa at 0.68, and Botswana at 1.64 cases per 1,000. Population-standardized annual mortality was highest in Sierra Leone at 186.13 deaths per 100,000 people, followed by Burkina Faso at 177.80, Niger at 155.81, Liberia at 145.19, and Cameroon at 142.73 deaths per 100,000, presenting marked contrasts with raw case counts where larger nations dominated absolute totals.

The authors emphasized several methodological constraints inherent in the macroscopic country-year study design. First, disease burden figures relied on modeled estimates generated by the Malaria Atlas Project rather than direct, unadjusted public health registry entries, meaning that uncertainties within the underlying disease mapping frameworks necessarily propagated into the downstream machine-learning classifiers. Second, national annual aggregations cannot capture localized district outbreaks, elevation gradients, seasonal vector cycles, or subnational healthcare inequalities, nor can country-level ecological correlations be interpreted as direct individual-level causal mechanisms. Third, key operational determinants—including insecticide-treated net usage, indoor residual spraying campaigns, antimalarial drug coverage, vaccine rollouts, vector abundance, and pharmacological or insecticide resistance—were not consistently available across all jurisdictions and could not be included. The researchers concluded that machine-learning models should serve as complementary risk-screening tools alongside conventional public health surveillance rather than standalone autonomous decision systems.

Subject of Research: Machine learning prediction and spatiotemporal analysis of malaria risk across African countries

Article Title: Spatiotemporal patterns and machine-learning prediction of malaria risk across Africa, 2000–2024

Article References: Ofori, S. A., Okaya, J., Addo, A. K., Agongo, G., Dassah, S. D., & Abugri, J. (2026). Spatiotemporal patterns and machine-learning prediction of malaria risk across Africa, 2000–2024. Discover Artificial Intelligence, 6(1), Article 1865. https://doi.org/10.1007/s44163-026-02748-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02748-1

Keywords: malaria risk prediction, machine learning, spatiotemporal patterns, Malaria Atlas Project, African public health, epidemiological modelling, Random Forest classifier, XGBoost algorithm, support vector machine, gradient boosting machine, climate indicators, socioeconomic determinants, temporal cross-validation, model calibration, geographic transferability, Brier score, Plasmodium falciparum, NASA POWER