Foundation models for tabular medical detection: a robust and generalizable framework for clinical decision support

In clinical prediction, tabular data needs to be processed, which is difficult due to small, heterogeneous and imbalanced medical data sets. This work compares the performance of TabPFN V2, a tabular foundation model, across four public medical datasets for oncology, endocrinology, cardiology, and obstetrics. TabPFN V2 was compared against resource-constrained AutoGluon, a stacked ensemble, and logistic regression, with the same splits of stratified five-fold (with the same random seed). All metrics were computed using out-of-fold predictions saved after fitting median imputation in each training fold. The measures of accuracy, precision, recall, F1-score and ROC-AUC were used to assess the predictive performance, and the Brier score, the expected calibration error and reliability diagrams were used to assess the probabilistic reliability. Paired stratified bootstrap resampling (1,000 resamples) quantified uncertainty in pairwise differences. Mean accuracy across the four datasets was 88.38% for TabPFN, 88.13% for the stacked ensemble, 87.02% for resource-constrained AutoGluon, and 84.81% for logistic regression. Compared with AutoGluon, TabPFN had higher mean ROC-AUC by 0.0093 (95% CI 0.0041 to 0.0151). Compared with the stacked ensemble, the ROC-AUC difference was small and compatible with no difference (0.0028; 95% CI -0.0005 to 0.0065), while the stacked ensemble had higher mean F1-score (TabPFN minus stacked ensemble, -0.0150; 95% CI -0.0274 to -0.0028). TabPFN had a lower mean Brier score than all three comparators under the evaluated protocol. TabPFN showed strong predictive discrimination and probability calibration on these four datasets, but the results do not establish general superiority across clinical tasks. The AutoGluon findings should be interpreted as a resource-constrained comparison, and the paired intervals are conditional on the four datasets studied.