Comparing Cost-Sensitive LightGBM and TabPFN for Imbalanced Coronary Heart Disease Risk Prediction

Coronary heart disease (CHD) constitutes a catastrophic non-communicable disease burden that operates as a non-natural disaster for health systems and directly threatening Sustainable Development Goal 3 (Target 3.4) on premature mortality; yet ten-year CHD risk prediction remains difficult because clinical cohorts are strongly class-imbalanced and most prior evaluations report point estimates without statistical or decision-analytic support. This study compares a fully tuned, cost-sensitive LightGBM pipeline with a zero-tuning tabular foundation model (TabPFN) on the imbalanced Framingham cohort of 4,240 records with a 15.19% event rate, under a common leakage-free protocol with validation-based threshold optimisation and an eight-scenario ablation, and evaluates the primary comparison with the DeLong test, bootstrap confidence intervals, a five-seed robustness check, calibration, and decision-curve analysis. TabPFN achieved higher discrimination, with an area under the receiver operating characteristic curve of 0.713 compared with 0.677 for the tuned ensemble (DeLong p = 0.0037), and better calibration, with a Brier score of 0.120 compared with 0.199. It also attained the best imbalance-aware detection, an F1-score of 0.380 and a Matthews correlation coefficient of 0.245, and remained superior under all five seeds, averaging 0.737 compared with 0.706. Threshold optimisation raised minority-class recall from 0.031 to at least 0.47 without resampling, and ten-predictor curation helped rather than harmed the foundation model. Shapley-based attribution yielded a transparent 0–100 score dominated by age, smoking, and blood pressure. A zero-tuning foundation model thus provides statistically significant, better-calibrated predictions than a tuned cost-sensitive ensemble, offering a scalable screening aid toward SDG 3 (Target 3.4) in resource-limited settings.