Transformer-Based Architecture for Predicting Surgical Complications from EHR Data

Surgical complications represent a major source of morbidity and healthcare costs, highlighting the need for accurate preoperative risk prediction. Traditional risk calculators often overlook the temporal dynamics embedded in EHRs. In this study, we evaluated STraTS, a Transformer-based architecture, for predicting surgical complications using a real-world EHR dataset of 54,395 procedures, of which 1,881 (3.46%) resulted in a complication. The model leverages a self-attention mechanism to capture complex temporal relationships within sparse clinical data. STraTS achieved an AUROC of 0.882 and an AUPRC of 0.406, and demonstrated consistent performance across patient subcohorts stratified by age and sex. These results indicate that Transformer-based models can effectively leverage longitudinal EHR data to generate individualized perioperative risk assessments.