Learning from undercoded clinical records for automated International Classification of Diseases (ICD) coding

Area of Science:

• Medical Informatics
• Machine Learning
• Clinical Data Analysis

Background:

• Clinical records often have incomplete International Classification of Diseases (ICD) code annotations, leading to noise in training data.
• This annotation noise can mislead deep neural networks used for automatic coding algorithms.
• Undercoded clinical records present a significant challenge for developing accurate automated coding systems.

Purpose of the Study:

• To develop an unbiased objective for training automatic coding algorithms using partially annotated clinical records.
• To address the challenges of annotation noise and data imbalance in learning from undercoded clinical records.
• To improve the reliability and accuracy of deep learning models for clinical record coding.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care III (MIMIC-III) dataset.
• Employed positive-unlabeled learning for unbiased loss estimation, mitigating misleading training signals.
• Implemented a reweighting mechanism to address sample imbalance and integrated supervision from the Medical Concept Annotation Toolkit.

Main Results:

• Positive-unlabeled learning with reweighting demonstrated superior performance compared to baseline methods across various missing label ratios.
• The integration of supervision from the annotation tool further enhanced model performance.
• Benchmarking confirmed the effectiveness of the proposed approach in handling noisy and imbalanced clinical data.

Conclusions:

• Unbiased loss estimation and reweighting are crucial for effective learning from undercoded clinical records.
• The combination of positive-unlabeled learning, reweighting, and annotation tool supervision offers a promising solution.
• This approach effectively tackles annotation noise and data imbalance, advancing automatic clinical coding.