Explainable Uncertainty-Aware Convolutional Recurrent Neural Network for Irregular Medical Time Series.

Area of Science:

• * Medical informatics and machine learning.
• * Time-series analysis in healthcare.
• * Clinical decision support systems.

Background:

• * Irregular medical time-series data presents challenges for accurate diagnosis prediction due to varying intervals between patient records.
• * Existing methods often regularize data without accounting for the inherent uncertainty, potentially impacting prediction reliability.
• * Dynamic changes in illness severity necessitate robust methods for analyzing patient health trajectories.

Purpose of the Study:

• * To develop a novel deep learning model, the Uncertainty-Aware Convolutional Recurrent Neural Network (UA-CRNN), for improved diagnosis prediction from irregular medical time-series data.
• * To incorporate uncertainty information into the analysis of medical time series to enhance risk prediction accuracy.
• * To propose an explainable version of the model (eUA-CRNN) for transparent and reliable predictions.

Main Methods:

• * Development of a hierarchical uncertainty-aware decomposition layer (UADL) to adaptively decompose time series into reliable subseries.
• * Implementation of an Uncertainty-Aware Convolutional Recurrent Neural Network (UA-CRNN) that integrates subseries-level uncertainty.
• * Introduction of an Explainable UA-CRNN (eUA-CRNN) utilizing specialized filters and an uncertainty-aware attention module for interpretable predictions.

Main Results:

• * The proposed UA-CRNN and eUA-CRNN methods demonstrate superior performance in diagnosis prediction compared to existing state-of-the-art approaches.
• * Incorporating uncertainty information at the subseries level effectively handles varying time intervals in medical data.
• * The eUA-CRNN provides explainable predictions by learning attention weights from uncertainty data.

Conclusions:

• * The UA-CRNN offers a significant advancement in predicting patient risk by effectively utilizing irregular medical time-series data.
• * Accounting for data uncertainty is crucial for improving the accuracy and reliability of diagnostic predictions.
• * The explainable nature of eUA-CRNN enhances trust and interpretability in clinical decision support.