Machine Learning in Predictive Healthcare Analytics

Abstract

Predictive healthcare analytics applies machine learning (ML) to electronic health record (EHR) data to forecast clinical deterioration, disease onset, and adverse events before they occur, enabling proactive interventions that improve patient outcomes and reduce healthcare costs. However, clinical deployment of ML models is hindered by dataset shift between development and deployment populations, lack of model interpretability for clinician trust, class imbalance in rare adverse events, and incomplete or irregularly sampled temporal EHR data. This study developed and benchmarked six ML architectures -- logistic regression (LR), gradient-boosted trees (XGBoost), random forest (RF), recurrent neural network (LSTM), temporal convolutional network (TCN), and a proposed Temporal Fusion Transformer (TFT-Health) -- across three clinically validated prediction tasks: 48-hour in-hospital mortality (MIMIC-IV, n = 52,846 ICU stays), 30-day unplanned readmission (eICU, n = 126,418 admissions), and sepsis onset 6 hours before clinical recognition (MIMIC-IV, n = 18,624 sepsis episodes). TFT-Health achieved the highest AUROC for mortality prediction (0.894 +- 0.006), readmission (0.762 +- 0.008), and early sepsis detection (0.882 +- 0.010), significantly outperforming all baselines. The model's interpretable multi-horizon attention mechanism identifies the temporal features (vital signs, laboratory trends, medication timing) driving each prediction, with clinician evaluation confirming that 87.4% of attention-highlighted features were considered diagnostically relevant by intensivists. External validation on an independent European ICU cohort (Amsterdam UMCdb, n = 23,106) confirmed robust generalisability (AUROC 0.878 for mortality, 0.04 degradation from development). These results establish TFT-Health as a state-of-the-art architecture for temporal EHR prediction and provide a reproducible multi-task benchmark for clinical ML research.