Comparison of Machine-Learning Algorithms for the Prediction of Current Procedural Terminology (CPT) Codes from...

Area of Science:

• Computational pathology
• Natural Language Processing (NLP) in healthcare
• Machine learning for medical informatics

Background:

• Pathology reports are crucial for clinical narratives, including diagnosis and prognosis.
• NLP methods are increasingly used to extract insights from these reports for clinical endpoints and biomarkers.
• Limited comparisons exist for deep learning versus other machine learning methods in predicting medical procedure information for pathology reimbursement.

Purpose of the Study:

• To compare the performance of XGBoost, SVM, and BERT in predicting Current Procedural Terminology (CPT) codes from pathology reports.
• To evaluate the utility of using all report subfields versus only diagnostic text for CPT code and pathologist prediction.
• To identify key report subcomponents influencing prediction accuracy.

Main Methods:

• Utilized advanced topic modeling on 93,039 preprocessed pathology reports.
• Compared XGBoost, SVM, and BERT for predicting primary and ancillary CPT codes using diagnostic text alone and all subfields.
• Analyzed text for predicting signing pathologists and used model explanation techniques to uncover important report subcomponents.

Main Results:

• BERT outperformed XGBoost on diagnostic text alone for primary CPT code prediction.
• XGBoost outperformed BERT when all report subfields were utilized for primary CPT code prediction.
• Incorporating additional subfields significantly increased prediction accuracy for ancillary CPT codes and showed substantial gains for XGBoost with primary CPT codes.

Conclusions:

• The developed approach achieved higher CPT code prediction accuracy than previously reported.
• Information beyond diagnostic text in pathology reports is valuable for accurate predictions.
• Future applications include detecting billing errors, standardizing reports, and estimating pathologist productivity (RVUs).