Music Clinical Usage Dataset

Music therapy, as a non-pharmaceutical intervention in evidence-based medicine, requires systematic knowledge of clinically effective musical content. However, existing datasets are rarely purpose-built for clinical music therapy and often lack fine-grained clinical annotations, limiting personalized music therapy guidance. We introduce MCUD, a clinically grounded dataset curated from randomized controlled trials indexed in PubMed (2002–2024), comprising 42 playlists and 378 songs spanning diverse clinical contexts, including neurodegenerative disease, affective and anxiety disorders, cognitive impairment, and cardiovascular conditions, with multi‑level annotations at both playlist (treatment goal, disease information, medical procedure, intervention modality) and song levels (genre, emotion, lyrics), plus three audio representations (traditional acoustic features, deep audio embeddings, and mel‑spectrogram images). We validate MCUD through baseline machine learning experiments on three tasks: clinical usage prediction, emotion recognition, and genre classification, achieving best macro F1 scores of 0.4238, 0.2482, and 0.5077, respectively. Correlation analyses further link rhythmic, spectral, and tonal features to clinical usage, emotion, and genre. MCUD provides a high-quality, evidence-based clinical dataset that will be iteratively updated, enabling the development of intelligent systems for personalized music therapy.