Explainable Graph Neural Networks for Multimodal Malaria Risk Prediction Using Malaria Atlas Project (MAP)-Inspired Synthetic Data and Climate Indicators in Sub-Saharan Africa
An explainable multimodal artificial intelligence framework that integrates epidemiological, climatic, geospatial, and demographic information through a modality-specific encoder-fusion architecture is proposed and extended with a spatial Graph Neural Network component designed to capture inter-district dependencies and potential spatial spillover effects.
Abstract
Malaria remains one of the most significant infectious disease burdens in Sub-Saharan Africa, placing sustained pressure on public health systems and malaria-control programmes. Accurate and interpretable identification of high-risk districts is therefore essential for early warning, resource allocation, and targeted intervention planning. This study proposes an explainable multimodal artificial intelligence framework that integrates epidemiological, climatic, geospatial, and demographic information through a modality-specific encoder-fusion architecture. The framework is further extended with a spatial Graph Neural Network component designed to capture inter-district dependencies and potential spatial spillover effects. Using a synthetically generated district-level dataset (n = 1,200), calibrated to reflect the statistical properties of variables commonly used in Malaria Atlas Project (MAP), NASA POWER, CHIRPS, WorldPop, and Sentinel-2 products across five malaria-endemic countries (the Democratic Republic of the Congo, Uganda, Nigeria, Tanzania, and Kenya), the proposed framework achieved a validation R² of 0.7285 for Plasmodium falciparum parasite rate regression. For binary high-risk classification, the framework attained an area under the curve of 0.9740 and an F1-score of 0.9630 on the held-out test dataset. SHapley Additive exPlanations were used to examine the contribution of individual predictors at global and local levels. Incidence rate (67.2%), rainfall (10.2%), and temperature (6.2%) emerged as the leading drivers of predicted risk, while intervention coverage showed a protective contribution. The results should be read as a synthetic proof-of-concept rather than as an operational malaria map. Even so, they show how multimodal learning, graph-based spatial modeling, and explainable AI can be combined into a transparent surveillance pipeline for future validation with real-world malaria data in resource-constrained settings.
Malaria elimination is shaped by complex interactions among climatic, environmental, socioeconomic, demographic, health-system, and intervention-related factors. However most studies examine only subsets of these drivers, limiting understanding of their combined influence on epidemiological risks. In this study, we integrated 25 years of data from 44 African countries on malaria burden and control, climate, environmental and land-use conditions, socioeconomic and demographic characteristics, and health-system capacity within a unified geospatial, explainable machine-learning, and forecasting framework to characterize spatiotemporal patterns of malaria, quantify the relative contributions of key determinants, and generate 10-year Africa-wide and country-specific forecasts of malaria incidence and mortality rates per 1,000 people at risk. We identified and mapped malaria incidence and mortality hotspots using the Getis-Ord Gi* statistic. Our analyses showed that both incidence and mortality burden remained highly heterogeneous across Africa, with persistent hotspots concentrated in the West and Central Africa. The explainable machine-learning model, that achieved high predictive performance (i.e., XGBoost for incidence, holdout R$^2$ = 0.92; Random Forest for mortality, holdout R$^2$ = 0.91), identified lower availability of hospital beds (per 1,000 people), higher mortality rate attributed to unsafe WASH (per 100,000), and lower percentage (%) of people using handwashing facilities as the top three most influential determinants of higher risk of infection across Africa whereas higher mortality rate attributed to unsafe WASH (per 100,000), lower % of people using at least basic sanitation services, and access to electricity (%) were associated with worse mortality outcomes. Forecasting models also demonstrated strong predictive accuracy (Naive persistence and Elastic Net, holdout R$^2$ = 0.98 for incidence and 0.97 for mortality). Assuming current intervention and structural conditions persist, Africa-wide malaria incidence was projected to remain broadly stable, with a modest upward trend by 2035, whereas mortality was projected to decline initially and subsequently remain relatively unchanged. However, substantial country-level heterogeneity showed both emerging transmission hotspots and persistently high-burden countries. this suggests there is a need for sustained control and accelerated elimination efforts. Overall, this study demonstrates that integrating geospatial analysis, explainable machine learning, and forecasting provides a robust framework for understanding malaria dynamics, identifying the key determinants of burden, anticipating future trends, and supporting geographically targeted malaria control across Africa. Beyond malaria, our analyses can be applied as a generalizable approach for infectious disease surveillance, early-warning systems, hotspot detection, resource prioritization, and precision public health using large-scale longitudinal health data.
Malaria remains a significant public health burden in tropical and subtropical regions, where the efficient identification and prediction of risk areas remain challenging. Conventional field surveys used to map Anopheles breeding sites are costly, time consuming and often incomplete. Therefore, there is a pressing need for a geospatially integrated surveillance framework that can accurately map malaria risk and forecast future risk dynamics to support targeted control efforts. A geospatial hybrid modelling framework was developed by integrating multi-source remote sensing, malaria, and climate datasets. A Random Forest model was employed to determine the relative importance of the input variables, which were then weighted and selected for inclusion in a deep-learning architecture. The predictive model combines a 3D Convolutional Neural Network (3DCNN) to capture spatial patterns with a Long Short-Term Memory (LSTM) network to learn temporal dynamics. The model was trained against a baseline mean squared error (MSE) of 0.1 representing a naïve mean predictor. To improve spatial realism of the final risk maps, a Cellular Automata (CA) model was incorporated using a 3 × 3 Moore neighborhood structure, with parameters calibrated at γ = 0.293 and β = 43.9 to enhance the spatial propagation of risk across neighboring cells. The framework successfully mapped malaria risk in Dar es Salaam with a Spearman rank correlation of 0.92, identifying Kigamboni South, Tundwi, and Msongola as high-risk areas. The hybrid 3DCNN-LSTM model reduced the training by 97.3% and validation losses by 90.2% respectively from the baseline. Projections to 2060 indicate a steadily spatial increase in malaria risk, with a cumulative slope of 0.077 risk units over the 35-year horizon (an annual slope of 0.002), demonstrating the model's ability to estimate future risk.
Edmund Steven Kanjagaile, D. Deus, A. Msusa· International Journal of Hea...· 0 citations
The review points out four persistent gaps: the lack of attention-augmented RNNs for malaria forecasting in Southern Africa, limited integration of health-facility infrastructure features with climate predictors, inadequate handling of missing data in African satellite-derived climate series, and the absence of operational dashboards that make model outputs usable for district health officers without statistical training.
Sophia Tembure, W. Manjoro· Computer Science and Informa...· 0 citations
Understanding dengue dynamics from a spatiotemporal perspective has become increasingly relevant in recent years, as it allows the identification of factors influencing disease transmission, including climatic conditions, human behavior, and the distribution of Aedes aegypti, the primary vector responsible for dengue virus transmission in tropical and subtropical regions. This knowledge is essential for supporting public health decision-making and reducing the impact of dengue on vulnerable populations. This study proposes a spatio-temporal deep learning framework based on the U-Net++ architecture to generate high-resolution dengue risk maps in Colombia. The model integrates climatic, environmental, demographic, and socioeconomic information derived from satellite and census sources. Two spatial approaches were evaluated: a high-dimensional geography (HDG) approach at the national scale and a low-dimensional geography (LDG) approach applied to five departments with high dengue incidence. The model integrating climatic, social, and environmental information achieved the best performance in both approaches. In the HDG configuration, the best model (M9) reached a test mIoU of 0.6646 (validation mIoU = 0.7361). In the LDG configuration, the corresponding model achieved an average test mIoU of approximately 0.73 across departments. These results highlight the potential of integrating heterogeneous climatic, environmental, and socioeconomic data within a spatiotemporal deep learning framework to characterize dengue risk patterns and support high-resolution surveillance.
Daira Velandia, J. Contador, Juan Zamora et al.· Scientific Reports· 0 citations
Background
Surveillance of infectious diseases remains a persistent challenge in many countries across the world. Most of these countries have slow reporting systems, poor infrastructure of data, and a limited capability of predictive analytics. Conventional surveillance methods are reactive, which leads to delayed outbreak containment. The development and assessment of a machine learning-driven early warning system for infectious disease prediction using geospatial big data from South-Western Nigeria, at the level of Local Government Area.
Methods
A longitudinal dataset comprising 152,438 observations (2020-2024) across 30 LGAs in six states was analyzed. The confirmed figures of COVID-19, Cholera, Lassa Fever, Malaria and Ebola, mobility index, temperature, rainfall, population density and location data. The study utilized Logistic Regression (LR), Random Forest (RF), Gradient Boosting Machine (GBM), and ARIMA. We evaluated the performance of our model based on Accuracy, Precision and Recall. A module for the detection of anomalies.
Results
According to the results, Random Forest had the highest predictive performance (AUC = 0.94; F1-score = 0.91) followed by Gradient Boosting (AUC = 0.92; F1-score = 0.89). Moderate performance (AUC = 0.84) by Logistic Regression. ARIMA was able to capture trends over time but considerably underperformed at outbreak spikes (RMSE = 18.7). Mobility index and rainfall were necessary predictors (p < 0.05). Detection framework identified outbreak spikes with 93% sensitivity.
Conclusion
Machine learning early warning systems outperform conventional surveillance systems in developing countries. The fusion of environmental and mobility variables helps in prediction and containment at source. The framework provides a huge scalable and efficient resource model for digital disease surveillance in countries
I. Adewumi, N. Bakare, W. Ajayi et al.· London Journal of Physics· 0 citations
Malaria remains one of the most significant causes of morbidity and mortality in tropical and subtropical regions, and timely diagnosis is essential for effective case management. Microscopy is the traditional parasitological reference standard, while rapid diagnostic tests (RDTs) are widely used, field-deployable alternatives with product- and antigen-dependent sensitivity and specificity; both approaches are constrained by requirements for trained personnel, reagents, or equipment in resource-limited settings. This study develops and evaluates, on a simulated clinical dataset calibrated to published aggregate statistics, an explainable artificial intelligence pipeline for malaria diagnosis prediction from routinely collectable symptoms, vital signs, and haematological indices, using six machine-learning (ML) models: Logistic Regression (LR), Naive Bayes (NB), K-Nearest Neighbours (KNN), Random Forest (RF), Support Vector Classifier (SVC), and Decision Tree (DT). The Synthetic Minority Oversampling Technique (SMOTE) and Random Forest feature selection are embedded within a single leakage-safe pipeline that is refitted in every cross-validation fold. The reported best model is selected on the basis of the cross-validated F1-score rather than held-out test performance, and the test set is used exactly once for confirmatory reporting. Under this design, SVC with RF-selected features was selected (mean cross-validated F1 = 0.740), achieving a test-set accuracy of 0.906 [95% CI 0.852, 0.953], recall of 0.867 [0.667, 1.000], and ROC AUC of 0.959 [0.919, 0.988]. A paired bootstrap test found no statistically significant difference in AUC compared with the runner-up, Logistic Regression (AUC 0.956, p = 0.81). Permutation importance corroborated 8 of the top 10 impurity-based features. Parasite density, the quantity used to determine the parasitological diagnosis, was excluded from the predictor set as a precautionary safeguard against near-total label leakage. Calibration, subgroup recall by age and sex, and a class-weighting comparison are also reported. This framework illustrates, without any claim of clinical validity, how a leakage-safe ML pipeline and SHAP interpretability can be combined and rigorously self-audited; real patient-level data and external validation are required before any clinical inference is drawn.
David Chepkonga, A. Langat, Ebenezer Esenogho et al.· Asian Journal of Research in...· 0 citations