AbstractBackgroundPathology reports serve as an auditable trail of a patient’s clinical narrative containing important free text pertaining to diagnosis, prognosis and specimen processing. Recent works have utilized sophisticated natural language processing (NLP) pipelines which include rule-based or machine learning analytics to uncover patterns from text to inform clinical endpoints and biomarker information. While deep learning methods have come to the forefront of NLP, there have been limited comparisons with the performance of other machine learning methods in extracting key insights for prediction of medical procedure information (Current Procedural Terminology; CPT codes), that informs insurance claims, medical research, and healthcare policy and utilization. Additionally, the utility of combining and ranking information from multiple report subfields as compared to exclusively using the diagnostic field for the prediction of CPT codes and signing pathologist remains unclear.MethodsAfter passing pathology reports through a preprocessing pipeline, we utilized advanced topic modeling techniques such as UMAP and LDA to identify topics with diagnostic relevance in order to characterize a cohort of 93,039 pathology reports at the Dartmouth-Hitchcock Department of Pathology and Laboratory Medicine (DPLM). We separately compared XGBoost, SVM, and BERT methodologies for prediction of 38 different CPT codes using 5-fold cross validation, using both the diagnostic text only as well as text from all subfields. We performed similar analyses for characterizing text from a group of the twenty pathologists with the most pathology report sign-outs. Finally, we interpreted report and cohort level important words using TF-IDF, Shapley Additive Explanations (SHAP), attention, and integrated gradients.ResultsWe identified 10 topics for both the diagnostic-only and all-fields text, which pertained to diagnostic and procedural information respectively. The topics were associated with select CPT codes, pathologists and report clusters. Operating on the diagnostic text alone, XGBoost performed similarly to BERT for prediction of CPT codes. When utilizing all report subfields, XGBoost outperformed BERT for prediction of CPT codes, though XGBoost and BERT performed similarly for prediction of signing pathologist. Both XGBoost and BERT outperformed SVM. Utilizing additional subfields of the pathology report increased prediction accuracy for the CPT code and pathologist classification tasks. Misclassification of pathologist was largely subspecialty related. We identified text that is CPT and pathologist specific.ConclusionsOur approach generated CPT code predictions with an accuracy higher than that reported in previous literature. While diagnostic text is an important information source for NLP pipelines in pathology, additional insights may be extracted from other report subfields. Although deep learning approaches did not outperform XGBoost approaches, they may lend valuable information to pipelines that combine image, text and -omics information. Future resource-saving opportunities exist for utilizing pathology reports to help hospitals detect mis-billing and estimate productivity metrics that pertain to pathologist compensation (RVU’s).
Comparison of machine-learning algorithms for the prediction of current procedural terminology (CPT) codes from pathology reports
