Interpretable machine learning models for predicting in-hospital and 30 days adverse events in acute coronary syndrome patients in Kuwait

Comprehensive Summary

In this study, researchers use various machine learning models to explore the relationship between complex risk factors and adverse cardiac events following the diagnosis of acute coronary syndrome (ACS). A cohort of 1,976 patients in Kuwait were evaluated for medical history, demographics, clinical treatment, and given a 30-day follow up prior to ACS diagnoses. Outcomes were defined as in-hospital and/or 30 days post-discharge adverse events. Researchers assessed diverse risk factors such as demographics, medical history, symptoms upon hospital admission, clinical treatment, etc. Five algorithms (random forest, gradient boosting, extreme gradient boosting, support vector machine, and logistic regression) were developed using down-sampling to reduce bias towards the majority group (patients who did not experience any adverse events). Models were trained and evaluated with tenfold cross-validation methods. Overall, the random forest algorithm performed best for predicting in-hospital adverse events (AUC = 0.84, Accuracy = 81.29%), while the extreme gradient boosting model performed best for predicting adverse events within the 30-day post-discharge period (AUC = 0.81, Accuracy = 78.01%). Partial dependence plots, developed for model interpretability, showed LVEF values < 40%, furosemide administered in the first 24 hours after catheterization, and heart failure at presentation were the top three most important predictors of in-hospital adverse outcomes. An urgent coronary artery bypass graft, multiple culprit arteries, and an percutaneous coronary intervention were the top three predictors of an adverse event post-discharge. Finally, researchers used Shapley values to make risk predictions for randomly selected patients.

Outcomes and Implications

Major adverse cardiac event prediction is vital in preventing potentially fatal complications associated with ACS. However, modern risk-stratification tools often assume linear relationships between risk factors and are prone to overfitting. Machine learning models do not carry the assumptions traditional regression models hold over linearity and randomness, allowing them to uncover complex relationships between risk factors. Additionally, these researchers further looked into the relationship between specific risk factors themselves, finding certain risk factors were increased in the presence of others. By developing algorithms to not only predict adverse events, but to identify which risk factors contribute most to the model’s prediction, clinicians can use this information to take necessary steps to prevent further risk and complications both in and out of the hospital.