MalariaXAI: A Synthetic Proof-of-Concept Multimodal Explainable AI Framework for Malaria Risk Prediction Using Clinical, Climatic, and Genomic Data

Abstract

Background and Purpose: Malaria is still the most prevalent public health concern worldwide and as of 2023, an estimated 263 million people were infected and 597 thousand died. Sub-Saharan Africa accounts for an estimated 94% of cases and 95% of deaths. Current AI systems are plagued by monolithic, single-modal architectures, black-box reasoning, and the ongoing absence of a single platform that correlates climatic, genomic, and clinical features with explanations. This work is predicated on the assumption that, in a resource-limited healthcare context, accuracy alone is not sufficient to ensure an AI tool's adoption; practical use also demands interpretability, calibration, computational efficiency, and an explicit representation of the model’s logic. Here, we present MalariaXAI, a multimodal, explainable AI proof-of-concept, synthetically and experimentally developed, to illustrate the potential to correlate and explain successfully in a single multimodal platform.

Methods: A controlled comparative experimental design was adopted following the CRISP-DM framework. A synthetic dataset of 14 features spanning clinical (n=6), climatic (n=5), and genomic (n=3) modalities was constructed using probabilistic generative methods informed by published epidemiological distributions. Outcome labels were generated probabilistically to mitigate target leakage. Three modality-specific models include an XGBoost ClinicalModel, a Random Forest ClimateModel, and a Random Forest GenomicModel, which were trained on stratified 80/20 splits with 5-fold cross-validation and combined via a weighted FusionModel (clinical: 0.45, climate: 0.35, genomic: 0.20). Class imbalance was addressed using SMOTE. Explainability was implemented through SHAP, LIME, and a DoWhy-based structural causal model. Performance was evaluated using AUC-ROC, Accuracy, F1-Score, Recall, and Brier Score.

Results: Across the four models, AUC-ROC ranged from 0.82 to 0.971, accuracy from 80.5% to 96.1%, F1-Score from 0.803 to 0.960, recall from 0.812 to 0.965, and Brier Score from 0.047 to 0.112. The FusionModel achieved the strongest performance across every evaluation metric, including AUC-ROC (0.971), recall (0.965), F1-score (0.960), accuracy (96.1%), and Brier Score (0.047), outperforming each single-modality baseline. Consequently, the value of multimodal integration is better understood through balanced decision-support profiling rather than discrimination maximisation. In the absence of other components, Ablation analysis showed the independent effects of multimodal incorporation, SMOTE balancing, and genomic data in the synthetic scenario. For SHAP, temperature at the fever stage, monthly rainfall, and mutation at Kelch-13 appeared to be the top predictive features across modalities, and their directions were all consistent with prior epidemiological domain knowledge.

Conclusions: This work illustrates the methodological viability of a combined synthetic malaria prediction pipeline using multimodal explainable AI methods. It suggests that multimodal ensembles, combined with structured explainability, can lead to computationally efficient, explainable malaria risk modeling in a controlled synthetic setting. However, it must be emphasised that all results herein are limited to the synthetic setting and cannot be taken as evidence of clinical predictive utility or readiness for deployment. External validation across different cohorts, fairness audit and prospective clinical validation are necessary steps prior to clinical adoption.