A tabular foundation model achieved higher discrimination for 30-day mortality (AUC 0.89; 95% CI 0.89-0.90) and morbidity than logistic regression and XGBoost in emergency surgery patients.
Observational (n=7,281)
Yes
Does a tabular foundation model improve the prediction of 30-day mortality and morbidity in emergency surgery patients compared to logistic regression and XGBoost models?
A tabular foundation model provides superior 30-day mortality and morbidity risk prediction in emergency surgery patients compared to traditional logistic regression and XGBoost models, even when trained on smaller site-specific datasets.
Effect estimate: AUC 0.89 (95% CI 0.89-0.90)
INTRODUCTION Surgical risk is locality-specific, and database infrastructure to support accurate preoperative risk stratification is limited globally. Recently, foundation models for risk prediction trained on large corpus of synthetic data that are ready for domain-specific applications have emerged. We aimed to evaluate the role of a tabular foundation model in widening access to high-accuracy local risk stratification in emergency surgery. METHODS We applied a transformer-based tabular pretrained foundation model with comparison to logistic regression and gradient boosting methods in our institutional data from the American College of Surgeons National Surgical Quality Improvement Program database of patients undergoing emergency surgery. We first compared performance overall, then at individual sites (n = 5), followed by comparison of the tabular prior-data fitted network model trained on site-level data against logistic regression and XGBoost models trained on all available multisite data not used for testing. Outcomes of interest were 30-day mortality and morbidity. RESULTS Among 7,281 emergency surgery patients (4.8% mortality, 30.2% morbidity), tabular prior-data fitted network achieved the highest area under the receiver operating characteristic curve (0.82, 95% confidence interval 0.81-0.83, for morbidity; 0.89, 95% confidence interval 0.89-0.90, for mortality) and under the precision-recall curve (0.68, 95% confidence interval 0.66-0.7, for morbidity; 0.35, 95% confidence interval 0.29-0.39, for mortality), and best calibration (Brier score 0.15 for morbidity and 0.04 for mortality) compared with logistic regression and XGBoost models. The tabular prior-data fitted network's excellent performance persisted in smaller site-specific cohorts. A tabular prior-data fitted network model trained only on a single site's data performed comparable to logistic regression and XGBoost models trained on all available multisite data (P < .4). CONCLUSIONS Access to high-performance surgical risk stratification can be improved through a tabular foundation model. This portable approach offers flexibility to missing data, and strong comparative performance in smaller data sets.
Varghese et al. (Mon,) conducted a observational in Emergency surgery (n=7,281). Tabular prior-data fitted network (foundation model) vs. Logistic regression and XGBoost models was evaluated on 30-day mortality (AUC 0.89, 95% CI 0.89-0.90). A tabular foundation model achieved higher discrimination for 30-day mortality (AUC 0.89; 95% CI 0.89-0.90) and morbidity than logistic regression and XGBoost in emergency surgery patients.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: