HeartRhythm: ECG-Based Music Preference Classification in Popular Music

Electrocardiogram (ECG) is a promising signal for music psychology research, but it is rare to find a music study applying ECG to machine learning, especially in music preference tasks. This study conducts an ECG-based music preference classification (favored vs. non-favored). The model is evaluated on a MUSEC dataset providing ECG data from 20 humans listening to two types of music: music without (Melody) and with (Song) lyrics. We also investigate the efficiency of physiological and statistical features and compare the model efficiency of ECG with the electroencephalography (EEG) studies listening to the same songs. Our results reveal that the random forest algorithm with statistical features indicates the outstanding F1-scores in Melody and Song
$(82.95 \pm 7.92\%$
and
$84.13 \pm 5.14\%$
, respectively). Moreover, the models can outperform the EEG studies. The result opens the opportunity of using ML with ECG to provide a better experience in appreciating music.