Understanding emotional responses to music using machine learning: evidence from a public dataset

Around several hundred million people globally suffer from severe or negative mental health illnesses (MH). However, individuals report highly variable perceptions of music's impact on their mental health. The purpose of this study was to develop and validate machine learning-based model to predict the effects of music on self-reported mental health (MH). The benchmark dataset is used to analyse the evaluated machine-learning models which has 1,626 individuals across 34 variables. Further this dataset is split into training and testing set, respectively, using state of the art techniques. Of all the machine learning algorithms examined, the Subspace KNN classifier performed the best with a classification accuracy rate of 90.2 percent. Nevertheless, the main importance of the study lies in its methodology, and not in the algorithms themselves, showing how sensitivity analysis shows that the predictors and the design of the study have much more impact on prediction performance than the algorithms used. While there have been studies that have introduced novel architectures for machine learning, this study aims to systematically compare existing supervised learning methods to a relatively under-explored prediction task in which heterogeneous self-reported perceptions of music's effects on mental health are predicted. The key methodological advance is not just comparison of classifiers, but the sensitivity analysis that shows that the choice of the classification algorithm has far less effect on the predictive performance than does the choice of the predictors and the study design. It is necessary to note that the high level of classification accuracy does not mean that listening to music is the only factor predicting the efficacy of therapy. The fact is that a large portion of the prediction ability results from the initial state of the individual's mental health, which is directly related to the outcome variable and thus makes up the predictor-outcome circularity. However, sensitivity analysis showed that the performance of the classifier was significantly affected by the presence of baseline mental health factors such as anxiety, depression, insomnia, and obsessive-compulsive disorder. However, it is important to note that this accuracy is substantially inflated by the inclusion of mental health severity predictors (anxiety, depression, OCD) that conceptually overlap with the outcome variable. A sensitivity analysis excluding these variables yielded a more modest accuracy of 73.2%, indicating that the model's predictive power derives primarily from this circular relationship rather than from music-specific features. In order to determine the contribution of the features related specifically to music, a further sensitivity test was conducted by excluding mental health-related variables. With this, the accuracy of classification was reduced to 73.2%. This suggests that the selection of predictors is more significant to the success of the model than the selection of the algorithm. Thus, results should be taken to reflect prediction of response to music. Moreover, about one-third of participants who reported balanced distributions for each group on all outcome measures indicated that they experienced negative effects of music in terms of their MH, which puts into question whether it is fair to assume that all people with MH disorders derive benefits from exposure to music. Finally, results indicate that exposure to music can help differentiate groups based on the genre of music they listened to before outcome assessments. However, the primary contribution of this study lies more in its empirical nature than in its algorithmic aspect. Unlike previous studies that have introduced novel machine learning techniques for modeling behavioral health, this research assesses existing machine learning methods to predict self-reported perceptions about the effect of music on mental wellbeing. This assessment reveals that predictor selection impacts machine learning models significantly more than the choice of machine learning algorithm.