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

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Machine Learning for Healthcare

Background:

• Sepsis presents a significant mortality risk, underscoring the need for early detection.
• Timely identification of sepsis is critical for improving patient outcomes and survival rates.
• Existing diagnostic methods may face challenges with data gaps and class disparities.

Purpose of the Study:

• To develop and validate advanced machine learning models for the early detection of sepsis.
• To leverage interpretability and explainability techniques for enhanced clinical decision-making.
• To address challenges in sepsis prediction datasets, such as data gaps and class imbalances.

Main Methods:

• Development and optimization of approximately 12 diverse machine learning models.
• Utilized Bidirectional Long Short-Term Memory (BiLSTM) and Temporal Convolutional Network (TCN) architectures.
• Employed interpretability and explainability techniques to assess model performance and provide insights.

Main Results:

• BiLSTM and TCN models demonstrated superior performance in sepsis prediction compared to conventional models.
• Achieved high accuracy with ROC-AUC scores of 0.9566 and 0.9595, and F1 scores of 0.85 for both models.
• Successfully detected clinical patterns and minimized false negative results, crucial for medical applications.

Conclusions:

• Advanced machine learning models, particularly BiLSTM and TCN, show significant promise for early and accurate sepsis detection.
• Interpretability techniques enhance clinical confidence and support informed decision-making in sepsis diagnosis.
• The developed models offer a robust solution for sepsis prediction, addressing data challenges and improving patient care.