The study, published in Frontiers in Psychology, was conducted by Yakun Li of Sichuan Normal University and Binxin Ren of Zhonghe High School Chengdu. They analyzed the Music and Mental Health survey, an online questionnaire archived on Kaggle that has been reused extensively in teaching exercises and published analyses. The file contains 736 complete submissions collected between late August and early November 2022, with respondents aged 10 to 89 and a median age of 21. Recruitment was open and self-selected, with no probability sampling frame and no design weights, and the survey recorded no information on gender, geography, education, income or clinical history. The authors are explicit that the sample describes this archive rather than any defined population.

The predictors comprised 42 variables: how often respondents listen to each of 16 genres from classical to heavy metal, their single favorite genre, their primary streaming service, daily listening hours, the tempo of their preferred genre in beats per minute, and five indicators of musical engagement such as playing an instrument, composing, and seeking out new music. The outcomes were single self-anchored ratings from 0 to 10 for anxiety, depression, insomnia and obsessive–compulsive symptoms, plus a composite averaging the four standardized items. No validated clinical instrument was administered, so the authors deliberately speak of self-reported distress rather than symptoms or disorders, a caution they consider especially important for the obsessive–compulsive item, where everyday usage diverges sharply from the diagnostic construct.

What distinguishes the analysis is its validation protocol. The researchers compared elastic net, support vector machines, random forests and gradient boosting under nested cross-validation, meaning hyperparameter selection happened entirely inside training folds, with ten outer estimates of performance. Every model was benchmarked against trivial baselines: a regressor that predicts the sample mean and a classifier that always predicts the most common answer. Because repeated cross-validation folds share training data, model comparisons used a corrected t-test rather than the ordinary paired test, which understates variance. Finally, each best model was refitted on datasets with randomly shuffled outcome labels, 100 permutations per regression target and 200 for classification, and interpretability analysis was permitted only if the real model beat its own permutation null.

The out-of-sample results were strikingly modest. Random forests, the best-performing family, achieved cross-validated R-squared values of 0.068 for depression, 0.058 for the composite distress index, 0.032 for anxiety, and 0.006 or below for insomnia and obsessive–compulsive ratings. That means the entire 42-variable battery explained at most about seven percent of the variance in any distress measure beyond what the sample average already provides. Permutation tests separated depression, anxiety, the composite and obsessive–compulsive ratings from noise, but insomnia did not clear the null, and the authors report no feature attributions for it precisely because explaining a model that has learned nothing produces plausible narratives about associations that do not exist.

The clearest finding concerns the survey item asking whether music improves, leaves unchanged, or worsens respondents’ mental health. About three-quarters of participants, 542 of 728 with valid responses, endorsed improvement. A classifier that simply predicts improvement for everyone therefore scores 74.5 percent accuracy. The best machine learning model, a random forest, managed 74.7 percent, a difference of 0.2 percentage points that was not statistically significant. Balanced accuracy peaked at 0.551 against a chance floor of 0.500, and the Matthews correlation coefficient, the strictest metric because it demands above-chance performance on both classes simultaneously, reached only 0.135. The tuned support vector machine illustrates the mechanism starkly: hyperparameter search selected a heavily regularized configuration that assigned the improvement label to all 728 respondents in every fold, making it, in operational terms, a constant function that nonetheless attains the accuracy figure often read as proof that psychological state is predictable from playlists.

Feature importance results came with their own warnings. For anxiety, age dominated, roughly an order of magnitude above the next predictor, and for depression, metal exposure ranked first by SHAP values with age second. But permutation importance with bootstrap intervals showed the rankings to be unstable: only metal and hip hop exposure had intervals excluding zero, and the authors argue that apparent genre effects are plausibly proxies for age composition, since preference for metal and rap concentrates in younger respondents who also report higher distress. One genuinely non-linear pattern emerged: the association between daily listening hours and predicted depression was near zero between roughly two and five hours and rose sharply above about six hours, concentrating the signal in the upper tail of listening quantity rather than spreading it across the range.

The authors also quantified how much of the low ceiling reflects the outcome measures themselves rather than the music variables. Under classical test theory, an unreliable outcome attenuates any achievable correlation. The composite index, with a Cronbach alpha of 0.659, is the only outcome whose reliability is estimable from these data; correcting for it lifts the observed R-squared of 0.058 to 0.088, so roughly a third of the shortfall is attributable to measurement error. For the single items, under assumed reliabilities of 0.60 to 0.80, the depression value of 0.068 corresponds to disattenuated values between 0.085 and 0.114. Even under the most favorable assumption, however, the ceiling remains below 0.12, supporting population-level description rather than individual-level inference. Learning curves reinforced the point: validation performance flattened by roughly 400 respondents, indicating the models are not data-starved but have extracted nearly everything these variables contain.

The implications reach well beyond music psychology. The authors situate their benchmark within a broader literature on digital mental health biomarkers, where smartphone mobility, phone use, sleep and vocal characteristics are cataloged as candidate markers of psychiatric state. A scoping review of voice-based screening for depression and anxiety cited in the paper found a median total sample of 118 participants, external validation absent in 68 percent of evaluations, and only 11 percent rated as having high clinical readiness, with accuracy frequently reported without the class distribution needed to interpret it. Other work has shown that reported classification accuracy in small neuroimaging studies of depression falls systematically as sample size grows, the opposite of what a genuine signal should do. The authors argue that an intuitive behavioral proxy is a hypothesis about information content, not a biomarker, and should be required to clear a trivial baseline and a permutation null under nested validation before being treated as one.

The study carries substantial caveats of its own. The design is cross-sectional and observational, so nothing in it speaks to whether music actually affects mental health. Self-selection into a survey about music and mental health plausibly over-represents people who believe music helps, consistent with the 74.5 percent endorsement rate. The archive includes 153 respondents under 18 without documentation of parental consent, so the authors report adult-only sensitivity analyses throughout and draw no conclusions specific to minors. They also acknowledge collective overfitting: a dataset analyzed publicly many times accumulates analytic choices shaped by results already seen on it, a risk no internal validation can detect, though they note it is difficult to overfit a reused dataset toward a negative result. Their practical recommendations are concrete: report a trivial baseline, report a permutation null, nest hyperparameter selection inside cross-validation, and gate post-hoc interpretation on the model having beaten its own null. On the evidence assembled here, musical preference is not a usable individual-level marker of psychological state, though the finding carries one consolation for listeners: taste reveals so little about distress that the privacy risk of inferring mental health from streaming histories is, with variables of this kind, smaller than sometimes feared.

Subject of Research: Interpretable machine learning benchmark testing how much music genre preference and listening behavior predict self-reported anxiety, depression, insomnia and obsessive–compulsive symptoms

Article Title: How much can music listening habits tell us about mental health? An interpretable machine learning benchmark using genre preference and self-reported anxiety, depression, insomnia, and obsessive–compulsive symptoms

Article References: Li, Y., & Ren, B. (2026). How much can music listening habits tell us about mental health? An interpretable machine learning benchmark using genre preference and self-reported anxiety, depression, insomnia, and obsessive–compulsive symptoms. Frontiers in Psychology, 17, Article 1970053. https://doi.org/10.3389/fpsyg.2026.1970053

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1970053

Keywords: music preference, mental health, machine learning, cross-validation, class imbalance, self-reported distress, SHAP, predictive modeling, digital mental health, convenience sample, permutation test, streaming data