A Curriculum Batching Strategy for Automatic ICD Coding with Deep Multi-Label Classification Models.

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Medicine

Background:

• Deep multi-label classification models are crucial for automatic International Classification of Diseases (ICD) coding.
• Minibatch gradient descent (MBGD) training for these models faces challenges due to large ICD label sets and imbalanced data distributions.
• This data distribution inconsistency leads to overfitting and reduced model generalization.

Purpose of the Study:

• To propose a novel curriculum batching strategy to enhance the performance and generalization of deep learning models for automatic ICD coding.
• To address the overfitting issue caused by data distribution mismatches in MBGD-based training.
• To improve the models' ability to learn from long-tailed label distributions.

Main Methods:

• Developed a curriculum batching strategy involving offline generation of three batch sets using predefined sampling algorithms.
• These batch sets represent uniform, shuffled, and original training data distributions, creating a learning progression from simple to complex tasks.
• Implemented and evaluated this strategy with three deep multi-label classification models for automatic ICD coding.

Main Results:

• The proposed curriculum batching strategy significantly improved the performance of all investigated deep multi-label classification models.
• Models trained with the curriculum batching strategy demonstrated reduced overfitting and enhanced ability to learn long-tailed label information.
• Performance gains surpassed those achieved by a state-of-the-art label set reconstruction model.

Conclusions:

• The curriculum batching strategy is a simple yet effective method for improving MBGD-based training in automatic ICD coding.
• This approach enhances model generalization and mitigates overfitting, particularly in the presence of large, imbalanced ICD label sets.
• The strategy offers a promising direction for advancing automated clinical coding systems.