Classification on Heart Rate Zone using SVM with Random Search: A Comparative Study with RF and XGBoost

Abstract

Real-time classification of heart rate (HR) zones during weight training supports targeted adjustments to training intensity. This study introduces a machine learning framework that employs Support Vector Machine (SVM) as the primary classifier and uses Random Forest (RF) and XGBoost for comparison. HR data were acquired from various wearable heart rate sensor, then subjected to noise filtering, outlier removal, and normalization before being labeled into four zones: recovery, aerobic, anaerobic, and maximum effort. Random Search was applied to tune hyperparameters including the SVM kernel scale, RF tree count and depth, and XGBoost learning rate using cross-validated performance as the selection criterion. On a held-out test set, SVM achieved 98 % accuracy with 97 % precision, 98 % recall, and a 98 % F₁-score; RF reached 95 % accuracy with 95 % precision, 94 % recall, and a 95 % F₁-score; XGBoost recorded 94 % accuracy with 93 % precision, 94 % recall, and a 93 % F₁-score. Analysis of confusion matrices revealed that SVM’s decision boundary yielded fewer misclassifications between adjacent zones, while RF and XGBoost displayed distinct error patterns. These results highlight SVM’s capacity for clear separation in the feature space and establish performance baselines for integration into wearable HR-monitoring applications.