Explainable deep learning for early sepsis detection from ICU time-series data using XAI techniques

Abstract

Sepsis is one of the most deadly illnesses with a high risk of mortality. Consequently, identifying it at the beginning of illness symptoms is crucial and plays a key role in improving patient outcomes. This study presents a customized solution for the early detection of sepsis with an emphasis on the use of interpretability and explainability techniques, utilizing a range of machine learning approaches and interpretable artificial intelligence methods. The database on which this research study is based has many problems; the main ones being large data gaps and class disparities. Employing robust methods, precise categorizations, and rigorous computations, Approximately 12 diverse models were developed and optimized. With ROC-AUC indicators of 0.9566 and 0.9595 and F1 scores of 0.85 and 0.85 respectively, Bidirectional Long Short-Term Memory (BiLSTM) and Temporal Convolutional Network (TCN) models performed better than conventional models in terms of sepsis prediction. These two approaches have shown remarkable progress in detecting clinical patterns while avoiding false negative results—an essential aspect of the medical field. To assess model performance and offer clear insights into model predictions, interpretation-based techniques were employed. This improved clinical confidence and facilitated well-informed decisions in crucial medical diagnoses.