Myocardial infarction (MI) is still a major cause of sickness and mortality. This study presents a comparative evaluation of Random Forest (RF), Support Vector Machine (SVM), and Neural Network (NN) for predicting MI complications using combined clinical data and electrocardiogram (ECG) changes across multiple leads. 1700 patient records from the Krasnoyarsk Interdistrict Clinical Hospital in Russia, gathered between 1992 and 1995, are included in the dataset. After preprocessing the data to handle missing values, the analysis based on all available attributes shows that RF achieved the highest accuracy at 93.53%. In a second analysis focusing only on ECG-related features isolated using Weka, predictive accuracy remained high across all models, with SVM slightly outperforming the others at 93.24%. Receiver Operating Characteristic (ROC) curve analysis revealed that while all models had high overall accuracy, RF and NN were superior in detecting specific severe complications like third-degree AV block and myocardial rupture, whereas SVM’s performance was more variable. These findings are further contextualized through a methodological comparison with other recent Machine Learning (ML) studies, particularly those using advanced tree-based models, highlighting the relevance of ECG-derived features and future directions for improving clinical assessment.