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

Introduction: Acute myocardial infarction (AMI) is highly prevalent (3.8% in developed countries), affecting heterogenous populations, and can be influenced by varied factors, including demographics, clinical risk factors, and comorbidities. Identifying distinct AMI patient profiles can aid in understanding the disease and developing personalised treatment strategies. Methods: This study analysed data from UK Biobank participants with an AMI diagnosis. Using unsupervised clustering techniques - UMAP, latent profile analysis, and K-means clustering - distinct and robust patient profiles were identified and associated with co-morbidity prevalence. Next, we trained three supervised machine learning classifiers (Logistic Regression, Random Forest, and XGBoost) to predict profile membership from 28 biochemistry markers. SHAP values were used for post-hoc interpretation of the best-performing model. Finding: Four distinct patient profiles were identified: “CMR-GIRespRenal”, “AG-CMS”, “CM-MultiCardio”, and “PostMeno-CMSurgGI”. Each profile showed unique characteristics in socio-demographics, clinical risk factors (e.g., BMI, age, smoking, alcohol intake, waist and hip circumference) and disease prevalence. The Random Forest classifier outperformed all others, achieving an average weighted AUROC score of 78%. SHAP analysis highlighted key biochemical markers, such as Testosterone, Creatinine, Vitamin D, Urate, and lipid profile markers, as significant predictors of AMI profiles. Conclusion: This study underscores the heterogeneity of AMI patients and the importance of integrating patient profiles with biochemical markers for improved stratification in diagnosis and treatment. These identified profiles can guide personalised treatment strategies, tailoring interventions to the specific needs of each group. Understanding these profiles may also lead to novel therapeutic targets.