Improving case duration accuracy of orthopedic surgery using bidirectional encoder representations from Transformers...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Data Analysis

Background:

• Operating room inefficiencies arise from discrepancies between scheduled and actual surgical times.
• Accurate surgical duration prediction is crucial for optimizing scheduling and resource allocation.

Purpose of the Study:

• To demonstrate a proof-of-concept for predicting surgical case duration using natural language processing (NLP) and machine learning.
• To interpret radiology reports for patients undergoing radius fracture repair to estimate case duration.

Main Methods:

• Compared logistic regression, random forest, and feedforward neural networks with and without NLP (bag-of-words).
• Utilized feedforward neural networks with ClinicalBERT, a transformer model pre-trained on clinical notes, for NLP.
• Evaluated 201 cases using 70% training and 30% test sets, with 10-fold cross-validation for root mean squared error (RMSE) calculation.

Main Results:

• The feedforward neural network with ClinicalBERT achieved the lowest average RMSE (25.6 min), significantly outperforming the baseline model (39.3 min).
• The percentage of accurately predicted surgical cases (within 15% of actual duration) increased from 26.8% to 58.9% using ClinicalBERT.
• Statistical significance (P < 0.001) was observed for all primary comparisons.

Conclusions:

• NLP and machine learning successfully extract features from unstructured clinical data (radiology reports).
• This approach significantly improves the accuracy of surgical case duration prediction.
• The findings support the potential for enhanced operating room efficiency through AI-driven scheduling.