Explainable time-series machine learning for early sepsis prediction and risk stratification: a multi-dataset computational study

Background: Sepsis is a critical, time-dependent disease, and its early detection is essential to improve patient outcomes. In this work, we propose an integrated computational framework that combines phenotypic data exploration, temporal machine-learning modeling, and dynamic feature relevance to provide a comprehensive toolkit for early warning and risk stratification. The system was evaluated on two complementary datasets: SepsisExp, which includes expert-validated severity labels, and MIMIC-III, a widely used benchmark database for intensive care research. Methods: The framework comprises a phenotypic exploration phase, in which outlier detection and correlation analysis are integrated with time-series \((k)\)-means (ts-KMeans) clustering to map the heterogeneous clinical trajectories of the disease. This foundational analysis informed the development of Long Short-Term Memory (LSTM) models optimized to capture subtle temporal dependencies hours before sepsis onset. To bridge the gap between high-performance computing and bedside utility, model predictions were interpreted through permutation-based feature importance (PermFit), which reveals a dynamic shift in biomarker relevance as clinical events approach. Results: The proposed methodology establishes a robust framework for early sepsis prediction, achieving an in-depth characterization of the phenotypes characteristic of sepsis and stable predictive performance across different time horizons. In addition, the feature-importance analysis reveals a consistent temporal shift in key biomarkers that reflects clinically plausible disease progression. Conclusions: By transforming complex clinical data into interpretable insights, the proposed approach provides a transparent decision-support tool that can support timely clinical evaluation in time-critical settings.