Hyperparameter Tuning and its Effects on Cardiac Arrhythmia Prediction

Cardiac arrhythmia is an important indicator for a range of conditions that can cause fatal and irreversible damage on patients. Recognizing cardiac arrhythmia through ECG signals typically requires advanced medical experience, so automatically identifying the presence of arrhythmia using learning-based techniques can bring a major breakthrough for clinical diagnosis. Since most learning algorithms support a wide range of operation points for each set of hyperparameters, a careful tuning of these values must be conducted. In this paper, three classification algorithms are used to train arrhythmia predictors from a set of ECG measurements and other clinical variables: Random Forest, Gradient Boosting, and Support Vector Machines. A two-step hyperparameter tuning was conducted to assess a variety of training configurations, and the best sets of parameters are reported. Experimental results show that a maximum performance gain of 34% is achieved in terms of accuracy and 84% in terms of sensitivity through proper parameter adjustments. It was also observed that tuned parameters can improve the performance compared with the default library sets, with an increase of up to 10% in accuracy and of 31% in sensitivity. Compared with similar references from the literature, our best model (using Random Forest) is able to outperform every other solution in terms of accuracy.