Comprehensive Summary
Zhang et al. systematically reviewed 32 studies to evaluate the predictive performance of machine-learning (ML) models for in-hospital mortality after acute myocardial infarction. Due to heterogeneous modeling techniques and pronounced class imbalance, the authors performed subgroup meta-analyses on balanced versus imbalanced datasets. In external validation sets, models using balanced data achieved a pooled C-index of 0.83 (95% CI: 0.795–0.866), sensitivity of 0.81 (95% CI: 0.79–0.84), and specificity of 0.82 (95% CI: 0.78–0.86). However, models trained on imbalanced data produced a similar pooled C-index of 0.815 (95% CI: 0.789–0.842) but demonstrated lower sensitivity at 0.66 (95% CI: 0.60–0.72) with preserved specificity of 0.84 (95% CI: 0.83–0.85). Logistic regression, support vector machine, and random forest classifiers all showed consistently high discrimination, but methodological heterogeneity and insufficient direct comparisons to conventional clinical scores limited broader conclusions for this study.
Outcomes and Implications
These findings suggested that machine learning models can achieve acceptable discrimination for predicting in-hospital mortality following acute myocardial infarction, particularly when class imbalance is mitigated. However, diminished sensitivity in imbalanced real-world datasets and lack of head-to-head comparisons with established risk tools highlight remaining challenges for this technology. Standardized model development, external validation, and direct comparison with conventional clinical risk scores are necessary before considering clinical adoption.