Towards Reliable Prediction: A Bayesian Continual Learning Approach for Clinical Time-Series Data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Informatics

Background:

• Deep learning models struggle with generalization on clinical time series data.
• Continual learning offers a promising solution by preserving representations while adapting to new data distributions.

Purpose of the Study:

• To propose and evaluate the Continual Bayesian Long Short Term Memory (C-BLSTM) algorithm for domain incremental learning in clinical settings.
• To enhance the generalization capabilities of deep learning models for time series prediction using electronic medical record data.

Main Methods:

• Developed C-BLSTM, a continual learning algorithm integrating architectural pruning, variational inference-based regularization, and coreset replay.
• Evaluated C-BLSTM on public electronic medical record datasets for mortality prediction.
• Applied C-BLSTM to real-world datasets for predicting heart failure readmission risk and type 2 diabetes glycated haemoglobin outcomes.

Main Results:

• C-BLSTM demonstrated superior performance compared to state-of-the-art continual learning methods on mortality prediction tasks.
• The algorithm effectively addressed domain incremental characteristics, including significant marginal and moderate conditional distribution drifts.
• C-BLSTM improved generalization across five diverse real-world scenarios: temporal, site, device, case mix, and ethnicity shifts.

Conclusions:

• C-BLSTM significantly enhances generalization and prediction reliability for clinical time series data.
• The proposed method offers a robust solution for domain incremental learning in healthcare applications.
• C-BLSTM shows promise for improving predictive modeling in dynamic clinical environments.