Neural Machine Translation-Based Automated Current Procedural Terminology Classification System Using Procedure Text:...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Documentation Improvement

Background:

• Substantial administrative costs in US healthcare are linked to medical billing errors.
• Deep learning models offer potential for predicting billing codes and reducing overhead.
• Automating coding processes can improve efficiency and accuracy in healthcare administration.

Purpose of the Study:

• To develop an automated anesthesiology Current Procedural Terminology (CPT) prediction system using neural machine translation (NMT).
• To enhance medical billing coding accuracy and reduce administrative costs.
• To compare the NMT system's performance against established machine learning algorithms (SVM, LSTM).

Main Methods:

• Analysis of 2.5 years of operative procedures from Michigan Medicine (Jan 2017-Jun 2019).
• Training and validation of NMT, Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) models.
• Evaluation across three experimental settings using surgical procedure text, preprocessed text, and combined text with preoperative diagnoses.

Main Results:

• The NMT model achieved high top-1 accuracy (81.64%-81.71%) in initial experiments.
• The SVM model demonstrated the highest top-1 accuracy (84.30%) when preoperative diagnoses were included (Experiment 3).
• Inclusion of preoperative diagnoses improved top-1 accuracy across all models, with SVM, LSTM, and NMT showing increases of 3.7%, 3.2%, and 1.3%, respectively.

Conclusions:

• An automated anesthesiology CPT classification system using NMT is feasible and effective.
• The NMT system's performance is comparable to SVM and LSTM models for CPT code prediction.
• Incorporating preoperative diagnoses significantly enhances the accuracy of automated CPT code prediction.