The goal of this pilot study is to learn whether group music therapy improves the emotional health of residents living in long-term care facilities. It will also examine the feasibility of integrating an artificial intelligence (AI)-based emotion recognition model into routine psychosocial interventions.

The main questions it aims to answer are:

Does group music therapy improve positive affect and reduce negative affect, depression, and loneliness among long-term care residents? Are AI-based facial emotion recognition results consistent with residents' self-reported emotional assessments?

Researchers will use a one-group pretest-posttest quasi-experimental design to evaluate changes before and after a 6-week group music therapy program. The study will also compare subjective questionnaire results with objective facial emotion recognition outputs generated by the PaLI Gemma 2 multimodal model.

Participants will:

Attend one 60-minute group music therapy session per week for 6 weeks Complete emotional health questionnaires before the first session and after the sixth session Be recorded during sessions using a non-invasive camera system for facial emotion analysis Have their questionnaire results compared with AI-based emotion recognition outputs to evaluate consistency and feasibility

This pilot study will provide preliminary evidence regarding both the psychological benefits of group music therapy and the feasibility of applying AI-supported multimodal emotion assessment in long-term care settings.

Taiwan is entering a super-aged society, and an increasing number of older adults reside in long-term care (LTC) facilities. Although institutional care provides medical and daily living support, many residents experience loneliness, depressive symptoms, and reduced emotional well-being. Non-pharmacological interventions such as group music therapy have demonstrated beneficial effects on mood regulation, social interaction, and psychological health. However, most previous studies have relied primarily on self-report questionnaires or observer-rated scales, which may be influenced by cognitive status, expressive ability, or assessment bias.

This pilot study aims to evaluate both the psychological effects of group music therapy and the feasibility of integrating an artificial intelligence (AI)-based multimodal emotion recognition model into emotional assessment in LTC settings.

Study Design

This study uses a quasi-experimental one-group pretest-posttest design. Approximately 20 residents from a long-term care facility in central Taiwan will be recruited using convenience sampling.

Inclusion criteria include:

Age 65 years or older

Residency in the facility for at least 3 months

Possesses basic cognitive functions (0-2 errors on the SPMSQ scale)

Ability to participate in group activities

Provision of informed consent for participation and facial image collection

Residents with severe dementia, major psychiatric disorders, neurodegenerative diseases, severe sensory impairments, or recent participation in other psychological or music therapy programs will be excluded.

Intervention

Participants will receive a 6-week group music therapy program. Each participant will attend one 60-minute session per week, for a total of 6 sessions. Sessions will be conducted in small groups (approximately 10 participants per group) and led by a trained music therapist with support from nursing staff and research personnel.

Each session will include:

Music listening using culturally familiar songs

Instrument interaction (e.g., tambourines, hand bells, xylophones)

Singing and vocal expression activities

Short group sharing discussions to promote emotional expression and social connection

The intervention is designed to progressively enhance emotional engagement, social interaction, and psychological comfort in a supportive group environment.

Outcome Measures

Emotional health outcomes will be assessed at two time points:

T0 (baseline): Before the first session

T1 (post-intervention): Within 24 hours after the sixth session

Subjective measures include:

International Positive and Negative Affect Schedule Short Form (I-PANAS-SF)

Geriatric Depression Scale - 15 item version (GDS-15)

UCLA Loneliness Scale Version 3

Barthel Index (functional status, baseline only)

AI-Based Multimodal Emotion Analysis

During each music therapy session, participants' facial images will be captured using a non-invasive camera system under standardized environmental conditions.

The PaLI Gemma 2 multimodal model will analyze static facial images to generate:

Categorical emotion outputs (e.g., happy, sad, angry, fear, surprise, disgust, neutral)

Continuous emotion dimensions (valence and arousal values)

Emotion recognition outputs from Session 1 (baseline representation) and Session 6 (post-intervention representation) will be compared with corresponding subjective questionnaire scores. Correlation analyses (Pearson or Spearman) will examine the consistency and complementary value between AI-based emotion detection and self-reported emotional states.

Statistical Analysis

Descriptive statistics will summarize demographic and baseline characteristics.

Pre-post differences in emotional outcomes will be examined using:

Paired t-tests for normally distributed data

Wilcoxon signed-rank tests for non-normally distributed data

Effect sizes (Cohen's dz or r) and 95% confidence intervals will be reported to support future sample size estimation for larger trials.

Correlation analyses will evaluate agreement between subjective measures and AI-derived emotional indicators.

Ethical Considerations

The study will be approved by an Institutional Review Board (IRB) prior to implementation. Written informed consent will be obtained from all participants. Facial image data will be anonymized and used solely for emotion recognition analysis. No identity recognition will be performed. All data will be securely stored in encrypted databases, and participants may withdraw at any time without affecting their care.

Significance

This pilot study will provide preliminary evidence regarding:

The effectiveness of group music therapy in improving emotional well-being among LTC residents.

The feasibility of integrating AI-based multimodal emotion assessment into long-term care practice.

Findings will inform the development of future large-scale trials and contribute to the advancement of technology-assisted psychosocial care models in aging societ

Inclusion Criteria:

Aged 65 years or older

Resident of the long-term care facility for at least 3 months

SPMSQ scale error 0-2 questions

Medically stable and able to participate in group activities

Able to communicate and complete questionnaires (with assistance if needed)

Willing to provide written informed consent, including consent for facial image collection

Exclusion Criteria:

Diagnosed with severe dementia

Diagnosed with major psychiatric disorders

Diagnosed with neurodegenerative diseases that significantly impair participation

Severe hearing or visual impairment that prevents engagement in music therapy activities

Participation in other psychological therapy or music therapy programs within the past 3 months

Acute medical condition requiring hospitalization or intensive treatment

Individual Participant Data (IPD) will not be publicly shared due to the sensitive nature of the data collected in this study. The dataset includes psychological assessment results and facial image data used for AI-based emotion recognition analysis. Although all data will be de-identified and coded, the combination of emotional health information and facial image records may pose a potential risk of re-identification.

To protect participant privacy and comply with Institutional Review Board (IRB) regulations and local data protection policies, individual-level data will not be made publicly available. Aggregated, de-identified results may be shared in scientific publications or upon reasonable request for academic collaboration, subject to ethical approval and data use agreements.