Next-generation risk stratification: leveraging TabPFN to Predict MACE in chest pain patients after acute myocardial infarction rule-out in the emergency department
T Tsao, I Strebel, J Boeddinghaus, P Lopez-Ayala, G Hure, A Champetier, L Koechlin, P Bima, L G Hemkens, F Mahfoud, C MuellerAbstract
Background
The ESC 0/1-hour algorithm effectively facilitates the rapid rule-out of acute myocardial infarction (AMI), yet the subsequent identification of unstable angina (UA) remains a diagnostic bottleneck. Current pathways often rely on resource-heavy, radiation-intensive imaging, exacerbating Emergency Department (ED) overcrowding. While our validated No Objective Testing Rule novel (NOTR novel) regression model, based on the original No Objective Testing Rule (NOTR), offers a robust solution, the capacity for State-of-the-Art (SOTA) Machine Learning (ML) to capture high-dimensional, non-linear clinical interactions remains under-utilised.
Methods
In a secondary analysis of a large-scale, multicentre prospective international cohort (n=3,395), we included patients with suspected ACS in whom AMI was ruled out via the ESC 0/1h-algorithm. The primary outcome was 30-day adjudicated Major Adverse Cardiac Events (MACE), comprising AMI, UA, and cardiovascular death, including index events. We benchmarked the NOTR novel against two sophisticated architectures: Extreme Gradient Boosting (XGBoost), a high-performance gradient-boosting framework, and Tabular Prior-Data Fitted Network (TabPFN), a revolutionary Transformer-based tabular foundation model designed for in-context learning. All models utilised identical predictors: original NOTR variables plus sex, 1h-high-sensitivity cardiac troponin T (hs-cTnT), absolute change of hs-cTnT (1h–0h), ST-segment depression, and exercise-induced chest pain. Findings were validated in an international cohort (n=704) and a national cohort (n=2,423).
Results
In this first application of a tabular foundation model for ED risk stratification, TabPFN achieved high safety for the primary 30-day endpoint, demonstrating a sensitivity of 99.3% and NPV of 99.9% in the national cohort (n=2,423), and 100% sensitivity and NPV in the international cohort (n=704). While discriminative performance (AUC 0.81–0.85) was comparable to established XGBoost and regression models, TabPFN uniquely leveraged in-context learning to provide the highest diagnostic efficacy. Specifically, TabPFN achieved a rule-out proportion of 47.5%, exceeding the 45.9% and 41.4% observed for the regression baseline and XGBoost, respectively. These findings demonstrate the feasibility and potential of integrating Transformer-based architectures into clinical decision support systems.
Conclusion(s)
This study represents a successful implementation of TabPFN for risk-stratifying chest pain patients after AMI rule-out. These results mark a significant step toward a new paradigm in AI-driven triage; further refinement of such foundation models holds the potential to surpass traditional machine learning benchmarks and fundamentally transform emergency cardiac evaluation.TabPFN vs. ML: Safety and Efficacy in Triage