Optimising vitrectomy operation note coding with machine learning

Area of Science:

• Ophthalmology
• Medical Informatics
• Machine Learning

Background:

• Accurate encoding of operation notes is vital for healthcare funding and planning.
• Vitrectomy procedure coding accuracy requires evaluation for improved efficiency.

Purpose of the Study:

• To assess the accuracy of vitrectomy procedural coding.
• To develop machine learning (ML) and natural language processing (NLP) models to aid in coding accuracy.

Main Methods:

• Retrospective analysis of 617 vitrectomy operation notes over 21 months.
• Manual coding based on Medicare Benefits Schedule (MBS), reviewed by consultants.
• Development and testing of XGBoost, random forest, and logistic regression models for classification.

Main Results:

• Manual review identified 66.5% of codes were missed, totaling $736,539.20 in under-reimbursement.
• The XGBoost model achieved 94.6% accuracy in multi-label classification for common procedures.
• XGBoost demonstrated strong performance (AUC 0.87) in identifying notes with multiple missing codes.

Conclusions:

• Machine learning models effectively classify vitrectomy operation note encoding.
• A combined human-ML approach can enhance clinical coding accuracy and financial reimbursement.
• Improved coding accuracy supports better resource allocation and higher quality patient care.