Benchmarking text encoding strategies in multimodal clinical data for surgical case duration prediction

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Operations Management

Background:

• Operating rooms (ORs) are resource-intensive, with surgical case duration often estimated inaccurately.
• Machine learning models using structured data improve predictions, but unstructured clinical text is underutilized.
• Clinical text contains valuable contextual details for enhancing predictive accuracy.

Purpose of the Study:

• To benchmark text encoding strategies for predicting surgical case durations.
• To evaluate the impact of combining structured perioperative data with unstructured clinical text.
• To compare classical and contextual text encoding methods in surgical duration prediction models.

Main Methods:

• Retrospective analysis of 180,370 elective surgical cases (2015-2020).
• Combined structured variables (age, sex, BMI, ASA score) with unstructured text (procedure descriptions).
• Encoded text using label encoding, count vectorization, TF-IDF, ClinicalBERT, and Sentence-BERT; trained diverse machine learning models.

Main Results:

• Integrating clinical text with structured data improved prediction accuracy across all models.
• Contextual embeddings (ClinicalBERT, Sentence-BERT) outperformed traditional methods, reducing Mean Absolute Error (MAE) to ~26.4 minutes and Symmetric Mean Absolute Percentage Error (SMAPE) to 21.6%.
• Improvements over structured-only baselines were statistically significant, reducing prediction error by up to 16%.

Conclusions:

• Semantically rich clinical text integration substantially improves surgical duration prediction accuracy.
• A multimodal approach combining structured data and contextual embeddings enhances OR scheduling and resource utilization.
• Future work should explore additional narrative sources and interpretability for clinical adoption.