Using Deep Learning With Few-Shot Learning to Improve Data Capture in Total Hip Arthroplasty Operative Notes

Area of Science:

• Orthopaedic surgery
• Artificial intelligence in medicine
• Clinical data analysis

Background:

• Structured data annotation is crucial for orthopaedic registries.
• Classifying THA implant fixation, technology, and surgical approach is challenging for traditional ML.
• This study explores GPT-4's feasibility for THA note analysis.

Purpose of the Study:

• Evaluate GPT-4's ability to capture and justify key THA operative note elements.
• Assess the accuracy and quality of GPT-4's data extraction.
• Determine GPT-4's potential as an alternative to manual chart review.

Main Methods:

• GPT-4 trained using few-shot learning with THA operative note examples.
• 240 primary THA notes from 38 surgeons were used for testing.
• GPT-4 output compared against manual chart reviews for accuracy and justification quality.

Main Results:

• GPT-4 achieved high accuracy: 100% for fixation, 98.9% for technology, 97.5% for approach.
• Model provided justifications with high readability and diversity.
• Justifications showed strong character-level matches with original notes.

Conclusions:

• GPT-4 with few-shot learning excels at capturing THA fixation, technology, and approach.
• The AI's ability to cite original notes aids validation for clinical data capture.
• GPT-4 offers a promising alternative to manual chart review for orthopaedic registries.