Prognostic significance of explainable machine learning-defined type 2 myocardial infarction: unmasking a high-risk phenotype in multicenter data
J Y Ren, Z Zhang, H Zhang, Xian Shao, Z K Xue, R X Guo, X Liu, K Chen, T LiuAbstract
Background
Type 2 myocardial infarction (T2MI) is frequently perceived as a secondary clinical entity compared to type 1 myocardial infarction (T1MI), often leading to a conservative "wait-and-watch" approach. However, this clinical assumption may overlook patients with extreme physiological vulnerability.
Purpose
We utilized an explainable machine learning (ML) framework to characterize the distinct phenotypic features of T2MI and evaluate its associated long-term mortality in a large-scale real-world population.
Methods
A multicenter registry of 163,335 patients was screened to identify individuals with acute myocardial injury. After excluding estimated glomerular filtration rate (eGFR) < 15 ml/min/1.73m2, dialysis, or non-ischemic etiologies, a final cohort of 37,251 patients was established. T2MI was defined as acute myocardial injury triggered by supply-demand mismatch without coronary revascularization. An eXtreme Gradient Boosting (XGBoost) algorithm was trained on 17 clinical variables, with model decision-making decoded using SHapley Additive exPlanations (SHAP) values to ensure diagnostic transparency.
Results
The final cohort included 35,197 patients with T1MI and 2,054 patients with T2MI. The ML model achieved a high discriminative performance with a validation area under the curve (AUC) of 0.870 (95% CI 0.820–0.920) (Figure 1A). Hemoglobin emerged as the most critical predictor for T2MI identification (gain: 0.399), followed by log-transformed N-terminal pro-B-type natriuretic peptide (log-NT-proBNP) and serum glucose (Figure 1C).SHAP analysis confirmed a consistent clinical logic where lower hemoglobin levels and increased hemodynamic stress strongly influenced T2MI diagnosis (Figure 2A), with a non-linear risk increase observed when hemoglobin fell below approximately 120 g/L (Figure 2B). Longitudinal follow-up over 3,500 days revealed a profound survival gap, as AI-defined T2MI patients exhibited significantly higher cumulative mortality compared to T1MI patients (log-rank p < 0.0001) (Figure 2C).
Conclusion
Machine learning effectively identifies T2MI by quantifying the complex interplay between anemia and cardiac load. While the calibration slope of 3.34 reflects a systematic overestimation of probabilities, this outcome is a deliberate consequence of a high-sensitivity diagnostic strategy designed to prioritize the detection of a high-risk minority class within clinical screening. The significantly higher mortality observed in the AI-defined T2MI group underscores that these patients represent a vulnerable phenotype currently under-recognized in clinical practice, warranting earlier risk stratification and targeted intervention.Figure 1Figure 2