Authors: Abdullahi Idris, Aminu A. Abdullahi, Jamilu Awwalu, Abdullahi Uwaisu Muhammad
Abstract: The increasing availability of Electronic Health Records (EHRs), ICU monitoring systems and clinical sensor technologies has generated large volumes of temporal healthcare data that require advanced analytical approaches for effective interpretation and prediction. Traditional machine learning and statistical models often face challenges in handling complex temporal dependencies, irregular sampling, missing values and censored survival outcomes in clinical time-series data. This study employed a Hybrid Transformer Framework for Temporal Representation and Longitudinal Risk Prediction in Clinical Time Series synthesizing the relevant studies and clinical decision-making. The framework integrates the Transformer-LSTM architecture with Cox Proportional Hazards (Cox PH), Survival Random Forest (SRF) and XGBoost algorithms. The Transformer component captures long-range temporal dependencies using self-attention mechanisms, while the LSTM network models short-term sequential clinical patterns. Cox PH is applied for interpretable survival analysis, SRF for nonlinear ensemble survival prediction and XGBoost for high-performance risk classification and prediction. The review study utilizes healthcare datasets such as MIMIC-III, MIMIC-IV, elCU and PhysioNet as well as providing suitable comparative approaches against baseline models.
DOI: https://doi.org/10.5281/zenodo.20269570
