Development and validation of deep continual learning model to sequentially learn multiple clinical prediction tasks...

Area of Science:

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

Background:

• Intensive care unit (ICU) patients present complex conditions requiring monitoring for multiple risks.
• Existing models struggle with continual learning for multiple clinical prediction tasks without catastrophic forgetting.
• This study addresses the need for effective continual learning (CL) models in ICU settings.

Purpose of the Study:

• To propose and evaluate three deep continual learning (CL) models for ICU patients.
• To assess the models' ability to perform multiple clinical prediction tasks sequentially.
• To mitigate catastrophic forgetting in deep learning models for critical care.

Main Methods:

• Utilized three public ICU databases (MIMIC-III, MIMIC-IV, eICU-CRD).
• Developed three CL models (CL_1, CL_2, CL_3) to sequentially learn eight prediction tasks.
• Compared CL models against baseline CL, single-task (ST), and multi-task (MT) models, evaluating performance and memory using backward transfer (BWT).

Main Results:

• The proposed CL models demonstrated effective mitigation of catastrophic forgetting.
• Performance was robust across different training orders, comparable to ST and MT models.
• CL_2 and CL_3 models showed potential for improved performance on current tasks by leveraging information from previously learned tasks.
• Outperformed baseline CL models in most experimental scenarios.

Conclusions:

• The developed CL models show significant promise for sequential learning of multiple clinical prediction tasks in ICU patients.
• CL_2 and CL_3 models exhibit the capability to enhance new task learning by utilizing prior task information.
• Further validation with diverse datasets and tasks is recommended to confirm the CL models' generalizability and effectiveness.