Cardiovascular disease remains one of the leading causes of mortality worldwide. The application of Machine Learning (ML) provides a promising approach for the early detection of heart attack risk, yet the predictive modeling process is frequently hindered by imbalanced data and the presence of irrelevant features. This study evaluates the performance of eight classification algorithms in predicting heart attack risk using a dataset of 8,763 patient records. To address the bias caused by the majority class dominance (64.1% healthy patients), the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training dataset. The evaluation was conducted comparatively across three data dimensionality scenarios, that is Baseline Data (all features), the top five features derived from Random Forest Feature Importance, and a subset of five features selected via Recursive Feature Elimination (RFE). The results demonstrated that the Gradient Boosting Machine (GBM) algorithm achieved the highest functional accuracy of 60.4% under the Baseline scenario. The most critical finding of this experiment was the empirical evidence of the Accuracy Paradox in the RFE feature scenario. Although linear models such as Logistic Regression and probabilistic models like Naive Bayes achieved the highest “accuracy” at 64.3%, further evaluation revealed that their Precision and Recall scores for detecting heart attack risk were exactly 0.00. These models completely failed to identify the positive class, merely predicting the majority class for all test samples. In conclusion, demographic and lifestyle features independently possess only moderate predictive power (~60%), and relying solely on Accuracy as an evaluation metric in clinical predictive modeling can lead to highly misleading conclusions.