CAMI-DM: Development and validation of a multi-algorithm model for in-hospital mortality risk prediction in diabetic patients with acute myocardial infarction - The China acute myocardial infarction registry.

Patients with diabetes constitute a substantial proportion of those with acute myocardial infarction (AMI) and exhibit distinct pathophysiological characteristics. However, existing guideline-recommended traditional and generic risk prediction models show limited performance in this specific population. Based on the China Acute Myocardial Infarction (CAMI) registry, 6,091 diabetic patients with AMI were enrolled and randomly divided into training and test sets (8:2). A comprehensive set of 62 multidimensional candidate features was extracted, and three feature selection strategies were applied to develop 12 machine learning models across six algorithm categories. All models underwent hyperparameter tuning via five-fold cross-validation in the training set, with the optimal combination selected according to the area under the receiver operating characteristic curve (AUROC). Two post-hoc ensemble strategies-stacking and probability averaging-were then employed to explore various combinations of top-performing models from different algorithm categories. Across six algorithm categories, the predictive models developed using 11 features selected by Elastic Net had the optimal performance. After evaluating various fusion strategies, the ensembled GLM + TabNet model was ultimately selected as the CAMI-DM model 2.0, achieving an AUROC of 0.875 in the test set. To balance the predictive performance and model simplicity, the CAMI-DM model 1.0 adopted a linear framework and was developed using five features from consensus feature selection Strategy, with an AUROC of 0.821 in the test set. Comparative analyses revealed that the CAMI-DM 2.0 outperformed CAMI-DM 1.0 in terms of discrimination, accuracy, calibration, and clinical net benefit. Furthermore, CAMI-DM 1.0, with its simplified structure, still outperformed the GRACE score in the overall predictive performance and generalizability. This study focused on the specific population of AMI patients with diabetes, and for the first time developed two dedicated models to predict in-hospital mortality risk.