Integrative machine learning model for predicting patient-oriented composite endpoints in acute myocardial infarction: a neuroendocrine biomarker approach

Abstract Background Acute Myocardial Infarction (AMI) represents a significant clinical challenge with diverse outcomes. Predicting patient-oriented composite endpoints (POCE) (all-cause death, any stroke, any AMI, or any revascularization) can significantly enhance post-AMI care. Incorporating clinical data with neuroendocrine biomarkers may offer improved prognostic capabilities. Objective To develop and validate a machine learning model to predict POCE in AMI patients, utilizing clinical parameters and neuroendocrine biomarkers. Methods There was prospective, observational, single-center study. Three approaches have been used to obtain the predictions for POCE event. First, as the dataset was imbalanced, an adaptive synthetic (ADASYN) sampling approach was used to generate synthetic instances, particularly focusing on those that are difficult-to-learn by using a weighted distribution. Then, random forest classifier (RF) was used to build a machine learning model to predict POCE. The model was tuned via a grid-search algorithm for optimal hyperparameters and validated using a 10-fold stratified cross-validation. Finally, the feature importance was determined by Shapley Additives, which measure the average marginal contribution of a feature value across all potential feature combinations. For the comparative purpose, the Gini index, which also shows the feature importance but in terms of mean decrease in impurity, was calculated. Results The study incorporated data from 315 patients, examining 47 variables and identifying 72 instances of POCE. After applying the ADASYN algorithm, the class distribution within the dataset was effectively equalized, facilitating the training of a robust RF model. Upon training with 252 instances, the model distinguished POCE with an accuracy of 83.8%, demonstrating a sensitivity of 80% and a specificity of 86%. During 10-fold cross-validation, the model's accuracy slightly dipped to 75%, with a sensitivity of 56% and specificity of 82%, indicating a balance between generalizability and overfitting. Testing mirrored these results, underscoring the model’s consistent performance. Notably, SHAP value analysis highlighted C-Reactive Protein (CRP) and alanine transaminase (ALT) as the most influential features, underscoring their significance in the context of AMI (Figure 1-2). The predictive power of each feature was further demonstrated by the mean decrease in Gini index (Figure 1-2). Conclusion Our study highlights the feasibility of employing machine learning to predict POCE post-AMI using an integrative dataset of clinical and neuroendocrine markers. The predictive model has the potential to revolutionize post-AMI care by allowing clinicians to identify high-risk patients early, tailor interventions, and allocate resources efficiently. Future research could explore the integration of this model into clinical workflows and its impact on patient outcomes.