An Original Machine Learning Framework for Predicting Sickle Cell Disease Crisis Events Using Longitudinal Electronic Health Record Data from Nigerian Tertiary Hospitals

This dissertation develops an original machine learning framework for predicting vaso-occlusive crisis (VOC) events in sickle cell disease patients using longitudinal electronic health record (EHR) data from Nigerian tertiary hospitals, addressing the absence of predictive decision support tools in high-burden African SCD settings. Existing VOC prediction models were developed in North American populations with access to rich digital data infrastructures that differ substantially from Nigerian EHR systems. EHR data spanning five years were extracted from the SCD clinic information systems at three tertiary hospitals: Lagos University Teaching Hospital, Obafemi Awolowo University Teaching Hospital, and Federal Medical Centre Owerri. Data covered 2,400 patient-years of follow-up for 480 HbSS patients and included full blood counts, reticulocyte counts, hydroxyurea doses, HbF percentages, crisis attendances, hospitalisations, and complication events. Time-series feature engineering extracted trajectory metrics and variability indices from longitudinal haematological data. Multiple machine learning algorithms were compared including gradient-boosted decision trees, long short-term memory recurrent neural networks, and a novel hybrid architecture integrating time-series haematological features with static clinical covariates. The hybrid LSTM-gradient boosting model achieved AUC of 0.82 for 30-day VOC prediction on holdout data, substantially outperforming clinician-rated risk assessment. Platelet variability, RDW trajectory, and reticulocyte count dynamics were the strongest predictors. The original framework is implemented as a clinical decision support prototype with real-time EHR integration capability. Keywords: machine learning, sickle cell disease, vaso-occlusive crisis prediction, electronic health records, Nigeria.