Enhancing Sepsis Detection Using BiLSTM with Attention Mechanism: A Time-Series Deep Learning Framework

Sepsis is a life-threatening condition that requires early and accurate detection to improve patient survival rates. This study proposed encoder-decoder approach with a Bidirectional Long Short-Term Memory (Bi-LSTM) and an attention method to find temporal dependencies from physiological data and enhance feature representation through adaptive weighting. The model consists of three key components among them the first one a Bi-LSTM-based encoder to capture bidirectional temporal dependencies, second an attention mechanism will assign dynamic importance to different time steps, and finally an LSTM-based decoder with dense layers for final classification. For training this model dataset from Kaggle and preprocessed to handling missing values, outlier removal, and addressing severe class imbalance using Random Under-Sampling (RUS). The proposed methodology Bi-LSTM with Attention model achieved an accuracy of 87.1%, a Cohen’s Kappa Score of 0.8818, and a Matthews Correlation Coefficient (MCC) of 0.6227.