CardioMeta Improves Multi-Task Cardiometabolic Disease Prediction

Cardiometabolic diseases, including diabetes, hypertension, and cardiovascular disease, frequently co-occur and are major drivers of preventable morbidity. Existing machine learning approaches for chronic disease prediction often prioritize discrimination on single datasets, frequently overlooking critical aspects like label leakage, calibration, temporal robustness, and external transportability. This paper introduces CardioMeta, a multi-task framework designed for the joint prediction of these three conditions across diverse data sources: population survey data (NHANES) and Electronic Health Record (EHR) data (MIMIC-IV). CardioMeta employs a shared cardiometabolic encoder combined with disease-specific gated heads and post-hoc probability calibration. A key design principle is the reduction of circular label reconstruction by excluding disease-defining variables from corresponding prediction heads in the primary analysis. In temporal validation, CardioMeta achieved strong performance metrics (macro-AUROC of 0.839, macro-AUPRC of 0.536, macro-F1 of 0.614, and ECE of 0.024), showing modest but consistent improvements over gradient-boosting and neural tabular baselines. While external evaluation on MIMIC-IV showed performance degradation due to domain shift, limited fine-tuning helped recover some performance. The study highlights that the true value of multi-task cardiometabolic modeling lies in reproducible leakage control, calibrated probabilities, and transparent reliability across heterogeneous healthcare data, rather than just inflated accuracy.