Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection, responsible for over 11 million deaths annually worldwide and constituting one of the most prevalent causes of ICU mortality. Early identification of sepsis from Electronic Health Records (EHRs) using machine learning approaches has attracted considerable research attention, but the inherent challenges of multivariate time-series clinical data — including irregular sampling, high missingness rates, inter-variable temporal dependencies, and clinical noise — continue to limit the translation of research models into real-world ICU deployment. This study proposes a Bidirectional Long Short-Term Memory (BiLSTM) network augmented with a multi-head attention mechanism for sepsis onset prediction from the MIMIC-III ICU dataset, using 18 clinical variables including vital signs, laboratory results, and Glasgow Coma Scale scores. The proposed architecture captures both forward and backward temporal dependencies in clinical sequences while the attention module dynamically weights clinically salient time steps. Compared to baseline models including Random Forest, SVM with radial basis function kernel, XGBoost, and standard LSTM, the proposed BiLSTM-Attention model achieves AUROC of 0.995, F1-Score of 0.981, sensitivity of 0.980, and specificity of 0.983 on the held-out test set. Ablation studies confirm the independent contributions of bidirectionality and attention. The model provides a 5.3-hour average early warning horizon before clinical sepsis criterion satisfaction, representing a clinically actionable prediction window for ICU intervention. SHAP-based feature importance analysis identifies lactate trend, heart rate variability, mean arterial pressure, and FiO₂/SpO₂ ratio as the most predictive variables, consistent with established sepsis pathophysiology.
Priya Mehta, Rajesh Kumar Sharma, Ananya Bose, Vikram Nair (Sat,) studied this question.