Classification of Current Procedural Terminology Codes from Electronic Health Record Data Using Machine Learning.

Area of Science:

• Anesthesiology
• Data Science
• Medical Informatics

Background:

• Accurate anesthesiology procedure codes are vital for quality improvement, research, and reimbursement.
• Advanced data science, including machine learning (ML) and natural language processing (NLP), can develop tools for Current Procedural Terminology (CPT) code classification.

Purpose of the Study:

• To develop and evaluate ML/NLP models for accurate classification of anesthesiology CPT codes.
• To achieve high accuracy in CPT code classification, potentially exceeding current benchmarks.

Main Methods:

• Trained five supervised ML models on over 1.1 million procedures from 16 institutions.
• Refined the top two models (SVM and neural network) on a separate Holdout dataset.
• Utilized actual billing data as the reference standard for accuracy assessment.

Main Results:

• Support vector machine (SVM) and neural network models achieved high accuracy, with SVM reaching 96.8% (top three codes) and 87.9% (single best code).
• High classification accuracy (93.1%-96.4%) was achieved in a significant percentage of cases (47%-62%).
• Procedure text was the most influential feature for model training.

Conclusions:

• ML and NLP techniques can create highly accurate, real-time models for anesthesiology CPT code classification.
• This approach offers potential for performance optimization and cost reduction in healthcare administration.