KIT-LSTM: Knowledge-guided Time-aware LSTM for Continuous Clinical Risk Prediction

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHR) offer vast temporal data for AI-driven risk prediction.
• Existing AI models struggle with the asynchronous and irregular nature of real-world EHR data.
• Accurate and timely patient risk prediction is crucial for clinical decision-making.

Purpose of the Study:

• To propose a novel deep learning approach, Knowledge-guided Time-aware LSTM (KIT-LSTM), for continuous mortality risk prediction.
• To address the limitations of current methods in handling complex EHR data structures.
• To enhance the interpretability of AI models in healthcare.

Main Methods:

• Developed KIT-LSTM, an extension of Long Short-Term Memory (LSTM) networks.
• Incorporated two time-aware gates and a knowledge-aware gate to model EHR data.
• Validated the approach on real-world patient data for acute kidney injury with dialysis (AKI-D).

Main Results:

• KIT-LSTM demonstrated superior performance compared to state-of-the-art methods in predicting patient risk trajectories.
• The model achieved improved accuracy in continuous mortality predictions.
• Enhanced model interpretability was observed, aiding clinical understanding.

Conclusions:

• KIT-LSTM effectively models complex, asynchronous EHR data for precise risk prediction.
• The proposed method offers a significant advancement for timely clinical decision support.
• This approach holds promise for improving patient outcomes in critical care settings.