Machine learning-based prediction of climate-driven sickle cell anemia crises across Nigerian regions using synthetic multimodal data

Sickle cell anemia (SCA) is a genetic hemoglobinopathy that imposes a major health burden in sub-Saharan Africa, particularly in Nigeria. Vaso-occlusive crises (VOCs), the principal acute manifestations of SCA, may be influenced by environmental factors such as temperature, low humidity, harmattan dust, and rainfall. Although associations between VOCs and environmental conditions have been reported, prediction models remain difficult to develop for data-constrained clinical settings. We developed a proof-of-concept VOC prediction framework for six Nigerian ecological regions using a synthetic multimodal health--environment dataset comprising 12,000 patient-month observations. Random Forest, support vector machine (SVM), XGBoost, neural network, and logistic regression models were evaluated. Logistic regression achieved the best overall performance (accuracy: 0.78; F1-score: 0.74; area under the receiver operating characteristic curve [ROC-AUC]: 0.84), followed by Random Forest (0.77, 0.72, and 0.83), SVM (0.78, 0.73, and 0.82), neural network (0.77, 0.72, and 0.82), and XGBoost (0.76, 0.70, and 0.82). Regional analysis showed a north--south gradient in simulated crisis occurrence: the North-East had the highest probability (80.1%), associated with temperature and particulate matter with aerodynamic diameter below 2.5, mu m (PM2.5), whereas the South-South had the lowest probability (10.4%). Temperature, PM2.5, humidity, and previous hospitalization were the most influential predictors. These findings establish the methodological feasibility of a climate-sensitive prediction pipeline that should be externally validated with clinical and meteorological data before use in health-surveillance applications.