Early Sepsis Prediction Using Stacked Lstm Architecture With Clinical Time-Series Data

Abstract

The Early Sepsis Detection and Monitoring System improves clinical decision-making and early diagnosis in critical care settings by utilising cutting-edge deep learning techniques, namely a stacked Long Short-Term Memory (LSTM) network. In order to identify patterns suggestive of the start of sepsis, the system analyses time-series clinical data, including vital signs and static patient features. In order to properly capture temporal dependencies, the model analyses a total of 11 important features over 30 timesteps, including age, heart rate, blood pressure, and breathing rate. Up sampling and other data augmentation techniques are used to improve the model's sensitivity to septic cases and overcome the inherent class imbalance in medical datasets. Metrics such as precision, specificity, and F1-score are used to assess the model's performance, which shows up to 98.4% recall and 96.07% accuracy. With the help of Streamlit, a user-friendly interface is created for real-time clinical application, enabling medical professionals to upload patient information in.csv or.psv format and obtain instant sepsis risk estimates. A scalable and successful approach to early sepsis intervention is provided by this combination of LSTM-based temporal modelling, class balancing techniques, and real-world implementation, which may lower mortality and enhance patient outcomes in critical care units.