An Interpretable Model for Predicting Acute Myocardial Infarction in Distinct Patient Profiles

Area of Science:

• Cardiology
• Biochemistry
• Data Science

Background:

• Acute myocardial infarction (AMI) affects diverse populations with varied risk factors.
• Identifying distinct patient profiles is crucial for personalized treatment strategies.
• Prevalence of AMI is 3.8% in developed countries.

Purpose of the Study:

• To identify distinct patient profiles for acute myocardial infarction (AMI).
• To associate these profiles with co-morbidity prevalence.
• To predict profile membership using biochemical markers and machine learning.

Main Methods:

• Unsupervised clustering (UMAP, latent profile analysis, K-means) on UK Biobank data.
• Supervised machine learning classifiers (Logistic Regression, Random Forest, XGBoost) trained to predict profile membership.
• SHAP values used for interpreting the best-performing model.

Main Results:

• Four distinct AMI patient profiles identified: "CMR-GIRespRenal", "AG-CMS", "CM-MultiCardio", and "PostMeno-CMSurgGI".
• Profiles exhibit unique demographic, clinical risk factor, and disease prevalence characteristics.
• Random Forest model achieved 78% AUROC; Testosterone, Creatinine, Vitamin D, Urate, and lipids were key predictors.

Conclusions:

• AMI patient population is heterogeneous, necessitating profile-based stratification.
• Integrating patient profiles with biochemical markers improves diagnosis and treatment.
• Identified profiles can guide personalized interventions and potentially reveal novel therapeutic targets.