LEVERAGING DEEP LEARNING FOR REAL-TIME, CONTINUOUS MONITORING AND PREDICTION OF SEPSIS IN ICU PATIENTS USING MULTISENSORY DATA

Sepsis is a life-threatening condition caused by the body's extreme response to infection, which may end fatally, with high death rates, particularly in the critical care unit, associated with multi-organ failure. Physiological data in ICU patients is highly non-stationary and complex, significantly challenging early detection. Standard machine learning models and traditional scoring systems fail to learn spatial and temporal patterns and accurately provide poor early detection performance. To overcome these limitations, we present a new framework called SepsisNet, a deep-learning model for real-time continuous sepsis prediction from multisensory ICU data. Our proposed attention-based CNN BiLSTM model incorporates CNN for spatial feature extraction and BiLSTM networks for temporal sequence modeling and adds an attention mechanism to emphasize the most informative physiological features for classification. On the benchmark dataset MIMIC-III, experimental results show that SepsisNet achieves 98.68% accuracy, surpassing the baseline models, including Logistic Regression, Random Forest, SVM, LSTM, and standard CNN. The ablation study also reinforces the importance of each architectural component. We demonstrate that SepsisNet has the potential to serve as a reliable, interpretable, and computationally efficient sepsis predictor, thereby enabling real-time clinical decision support in ICU settings. This research will help enhance sepsis detection as well as play a significant role in early medical treatment, which is significantly required to reduce fatalities caused by sepsis.