Adapting transformer-based encoder models for automated billing code assignment of current procedural terminology...

Area of Science:

• Computational Pathology
• Medical Informatics
• Machine Learning Applications

Background:

• Accurate pathology billing relies on correct Current Procedural Terminology (CPT) code assignment (88300-88309).
• Interpretive information in pathology reports presents challenges for consistent CPT code assignment.
• Existing machine learning models struggle with heterogeneous reports and institutional documentation variations.

Purpose of the Study:

• To advance machine learning for pathology CPT code prediction using large, diverse report corpora.
• To assess the adaptability of machine learning models across different institutions.
• To improve the accuracy and efficiency of pathology billing processes.

Main Methods:

• Transformer-based encoder models (SciBERT-Longformer) were sequentially fine-tuned on datasets from two major medical centers.
• Performance was compared against single-stage training and traditional machine learning methods (Naive Bayes, Random Forest, XGBoost).
• SHAP analyses were employed to interpret model predictions and validate clinical relevance.

Main Results:

• The sequentially fine-tuned SciBERT-Longformer achieved a macro-F1 score of 0.8912.
• Sequential fine-tuning demonstrated consistent improvements in F1 scores, particularly for rare CPT codes.
• Model predictions were interpretable, efficient, and aligned with clinical terminology.

Conclusions:

• Sequential fine-tuning on varied pathology report corpora significantly improves CPT code prediction performance.
• Institution-specific transformer models show potential for streamlining billing and reducing administrative burden.
• The developed models offer explainable and accurate solutions for pathology billing challenges.