Automated ICD coding for primary diagnosis via clinically interpretable machine learning.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Coding Systems

Background:

• Automated coding algorithms are expected to enhance coding quality and productivity.
• Studies on primary diagnosis auto-coding are limited within the Chinese healthcare context.
• International Classification of Disease, tenth version (ICD-10) coding requires improvement.

Purpose of the Study:

• To develop and evaluate a machine learning model for automated primary diagnosis ICD-10 coding.
• To address the gap in automated coding research within China.
• To improve the efficiency and accuracy of clinical coding.

Main Methods:

• Utilized 71,709 admissions from Fuwai hospital, covering 168 primary diagnosis ICD-10 codes.
• Employed sequential and sequential grouping features derived from clinical texts.
• Developed and compared models using Light Gradient Boosting Machine (LightGBM) classifier with optimized hyperparameters.
• Applied SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The best model, based on sequential grouping features, achieved 95.2% accuracy and 88.3% macro-averaged F1 score.
• Sequential information and grouping strategies significantly boosted model performance (P < 0.01).
• 91.1% of individual ICD-10 codes reached an F1 score above 70.0%.

Conclusions:

• The developed machine learning model is effective for automated primary diagnosis coding in the Chinese context.
• The model's interpretable results show potential for assisting clinical coders.
• This approach can enhance coding efficiency and quality in Chinese inpatient settings.