Deep reinforcement learning for multi-class imbalanced training: applications in healthcare

Area of Science:

• Machine Learning
• Artificial Intelligence
• Clinical Data Analysis

Background:

• Increasingly complex and imbalanced datasets, especially in clinical settings, pose challenges for traditional machine learning models.
• Rare events in clinical data often lead to severe class imbalance, hindering accurate prediction.
• Existing imbalanced learning methods struggle with extremely imbalanced and multi-class scenarios.

Purpose of the Study:

• To develop a reinforcement learning (RL) framework for training extremely imbalanced datasets.
• To extend the RL framework for effective use in multi-class classification settings.
• To improve classification fairness and minority class prediction in imbalanced clinical data.

Main Methods:

• Implementation of a novel imbalanced classification framework based on reinforcement learning.
• Combination of dueling and double deep Q-learning architectures.
• Formulation of a custom reward function and episode-training procedure tailored for multi-class imbalanced data.

Main Results:

• The proposed RL framework demonstrates superior performance compared to current state-of-the-art imbalanced learning methods.
• Achieved more fair and balanced classification outcomes on real-world clinical case studies.
• Significantly improved the prediction accuracy for minority classes in imbalanced datasets.

Conclusions:

• The developed reinforcement learning framework offers a powerful solution for handling extremely imbalanced clinical data.
• The approach effectively addresses multi-class imbalanced classification challenges.
• This method holds significant potential for improving diagnostic accuracy and patient outcomes in clinical practice.