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

Operating rooms (ORs) are highly resource-intensive, yet surgical case duration is often estimated using heuristics that are prone to errors. While machine learning models based on structured perioperative data improve accuracy, unstructured clinical text remains underutilized despite containing valuable contextual details.

To benchmark classical and contextual text encoding strategies, combined with structured perioperative data, for predicting surgical case durations.

We retrospectively analyzed 180,370 elective surgical cases from three tertiary care hospitals (2015-2020). Structured variables such as age, sex, BMI, ASA score, and case service were combined with unstructured text features (procedure descriptions), which were encoded using five different methods (label encoding, count vectorization, TF-IDF, ClinicalBERT, Sentence-BERT). We trained a diverse set of machine learning models including linear regression, tree-based ensembles, and neural networks and evaluated predictive accuracy using standard error metrics with cross-validation.

Adding unstructured clinical text to structured perioperative variables improved prediction accuracy across all models. Contextual embeddings consistently outperformed structured-only and traditional text encodings. Sentence-BERT and ClinicalBERT achieved comparable best performance, reducing MAE to approximately 26.4 minutes and SMAPE to 21.6%, with R² of 0.86; neither encoder was statistically superior to the other (p>0.82). Improvements over structured-only baselines were statistically significant (p<0.01), corresponding to up to 16% reduction in prediction error. Traditional encodings (label, count, TF-IDF) provided limited benefit.

Integrating semantically rich clinical text with structured perioperative data substantially improves surgical duration prediction. Our multimodal approach which combines structured and unstructured data with contextual embeddings, directly improves prediction accuracy, which in turn supports more reliable OR scheduling, better resource utilization, and improved patient care. Future work should incorporate additional narrative sources and interpretability techniques to support clinical adoption.