A next-generation prediction risk model for acute myocardial infarction: Derivation and validation in a multi-centre cohort

Aim Cardiovascular disease remains the leading global cause of death, and the need for accurate, event-specific risk prediction is particularly critical in regions where long-horizon models perform poorly. We developed and internally validated a probabilistic model to estimate 6- and 12-month risk of acute myocardial infarction, with exploratory 5- and 10-year horizons, using routinely collected electronic health record data from an integrated cardiovascular cohort in Colombia. Methods The study followed TRIPOD+AI guidance and analysed 382,589 patients contributing 3.9 million encounters. The modelling strategy combined a calibrated gradient-boosting classifier with an interpretable survival ensemble incorporating Cox regression, random survival forests, and discrete-time hazards. Primary outcomes were prediction accuracy, discrimination, calibration, and concordance with legacy score scales. Results The classifier achieved an AUC of 0.869, while 6- and 12-month survival models reached C-indices of 0.836 and 0.846. Calibration was strong, with predicted vs observed AMI counts nearly identical (O/E = 0.998). Concordance analyses demonstrated only moderate alignment with Framingham and PROCAM, indicating substantial re-ranking at short horizons compared with legacy long-term models. External, label-delayed validation (n = 5,602) showed monotonic risk separation across predefined priority bands. Conclusion This model provides a practical population-health stratification tool for short-term AMI risk, with particular value in resource-constrained settings. Recalibration to local incidence rates is recommended before deployment. Prospective evaluation is warranted to assess real-world clinical and operational impact.