Comprehensive Summary
This study evaluates the use of GPT-4 with a few-shot learning approach to extract key features from total hip arthroplasty (THA) operative notes. The model was trained on a small number of annotated notes and custom prompts designed by orthopedic experts. It categorized notes for implant fixation type, surgical technology, and surgical approach with high accuracy: 100% for fixation, 98.9% for technology, and 97.5% for approach. The model provided clinical justifications for its classifications, citing original note language and achieving high alignment scores. The study highlights GPT-4's potential to automate the extraction of critical THA surgical data, offering an efficient alternative to manual chart review. The model's ability to function with minimal training examples significantly reduces the time and cost involved in training traditional models. However, limitations include the single-institution dataset, which may affect generalizability, and some misclassifications due to ambiguous language.
Outcomes and Implications
The findings underscore GPT-4's potential to automate the extraction of critical surgical data from unstructured notes, offering a scalable alternative to labor-intensive manual chart reviews. The model's high accuracy in identifying key surgical features suggests it could improve data capture in registries where these features are historically underreported. Its ability to provide clinical justifications for its classifications is crucial for clinical validation and regulatory approval. The system's efficiency with few training examples reduces the resources needed for model training. However, broader validation across different institutions is necessary to ensure generalizability and address potential misclassifications due to language ambiguities.