A tree based approach for multi-class classification of surgical procedures using structured and unstructured data.

Area of Science:

• Medical Informatics
• Health Services Research
• Computational Medicine

Background:

• Current Procedural Terminology (CPT) code assignment in surgery is complex and manual.
• Existing literature on predicting CPT codes using text features is limited.
• This study addresses the need for improved CPT code prediction accuracy.

Purpose of the Study:

• To enhance the prediction of CPT codes in surgical and other services.
• To develop a novel re-prioritization algorithm for CPT code prediction.
• To leverage informative features and text data for improved accuracy.

Main Methods:

• Utilized Random Forest multi-class classification for initial CPT code prediction.
• Extracted key information including label probabilities and feature importance.
• Developed a novel algorithm to re-prioritize candidate CPT codes based on calculated weights.

Main Results:

• Random Forest achieved 74-76% accuracy for primary CPT prediction.
• The complementary algorithm improved results by an average of 8%.
• Incorporated text features enhanced output quality by 20-35%, outperforming neural networks.

Conclusions:

• A robust framework using a decision tree predictive model was established.
• The model predicts surgical codes more accurately than state-of-the-art deep neural networks.
• This framework can significantly aid surgery billing and scheduling.