Comment on "Development and validation of an interpretable machine learning model for early risk prediction of acute myocardial infarction".

Abstract

Cui et al. (2026) proposed an interpretable machine learning model for early prediction of acute myocardial infarction (AMI) with the potential of XGBoost and SHAP based interpretation for a transparent cardiovascular triage process. Several methodological aspects may limit the model's applicability and generalizability in a clinical setting, though. First, applicability-domain filtering might have led to the exclusion of patients with extreme laboratory profiles, as these are generally the highest-acuity patients and may be the ones to benefit most from the use of decision-support systems. Second, the initial model may be of high dimensionality and can lead to feature selection bias after the reduction performed by the SHAP-guided features.Second, after high dimensionality model, feature selection bias may arise and there is high probability of overfitting after performing SHAP guided feature reduction. Third, variables with high rates of missingness could be imputed with values that would lead to reduced variance and biased relationships in the clinical data, thereby potentially leading to overly confident predictions. Lastly, the use of discharge diagnosis as the reference standard could lead to label noise due to changes in diagnosis, administrative coding or inter-physician diagnosis variation. Relevant validation, missing data treatments and outcome definitions will improve the translatability of clinical studies in the future.