Artificial Intelligence and Predictive Medicine: Predicting Thrombocytopenia in Patients with Atrial Fibrillation

This study aimed to develop and evaluate machine learning models capable of predicting thrombocytopenia in patients with atrial fibrillation, with a focus on African clinical settings where this complication is underexplored despite its clinical importance. Real-world data were obtained from the AFRICA registry, a large multicenter database encompassing diverse patient profiles. Six algorithms were implemented and compared, including Decision Tree, Random Forest, XGBoost, Support Vector Machine, K-Nearest Neighbors, and Multi-Layer Perceptron. Stratified cross-validation was used to ensure robust evaluation based on accuracy, F1-score, AUC-ROC, Log Loss, and Matthews Correlation Coefficient. Model interpretability was enhanced using the SHAP method to identify the most influential predictors. Tree-based models performed best. On cross-validation, XGBoost reached 97.4%, F1 0.92, and AUC-ROC 0.98, and its performance was confirmed on an independent 20% holdout set (accuracy 93.75%, F1 0.857, AUC-ROC 0.9803). SHAP analysis highlighted platelet count, hemoglobin, creatinine, and glycemia as the strongest predictors, alongside clinical factors such as amiodarone therapy, intensive care admission, and depressive symptoms. High-performance, interpretable machine learning models can accurately forecast thrombocytopenia in African patients with atrial fibrillation. These findings provide a solid basis for the development of clinical decision support systems aimed at improving patient management and treatment outcomes. KEYWORDS AFRICA registry, clinical decision, hematology, model interpretability, multicenter cohort, supervised learning