Time-Series-Enhanced Sepsis Mortality Prediction Using Bidirectional LSTM with Temporal Attention and Focal Loss

Abstract

Predicting sepsis mortality in the ICU is a challenging and complex clinical AI problem. Although there is enough data available, the most statistically significant information is hidden within time-based trajectories that most published models ignore. In the earlier research based on SMOTE RF RFE XGBoost pipeline for MIMIC-III sepsis, it is clear that a patient whose lactate level rises from 1.6 to 4.9 mmol/L over six hours has a very different prognosis than one who stays stable at 4.9 throughout, even though both have the same max_lactate feature value. We present a Bidirectional LSTM (BiLSTM) with multi-head self-attention that jointly models the 39 static tabular features and raw hourly measurements of six vital signs heart rate, MAP, $\mathbf{S p O}_{\mathbf{2}}$, respiratory rate, temperature, lactate from MIMIC-III. The model achieves 14.7% relative recall improvement which substantially identifies more patients at risk of dying. The result significantly matters most in a time-sensitive ICU context.