Study on the path of combining music and digital health technology to promote the health of older adult groups.

OBJECTIVE

As the global population ages, non-pharmacological interventions such as personalized music therapy show promise for wellbeing in older adults. We propose the Fusion-Attentive Temporal Network (FAT-Net). This dual-stream model processes minute level heart-rate and music on/off data alongside daily summary features to predict a composite health score.

METHODS

Data from 92 participants over 45 ± 10 days were augmented fourfold using jittering, time-warping, magnitude scaling, and SMOTE. The temporal stream uses Conv1D, BiLSTM, and self-attention pooling. The summary stream uses a three-layer MLP. Cross-modal attention fuses both embeddings.

RESULTS

Over ten runs, FAT-Net achieved RMSE = 0.35 ± 0.005 (22.7% reduction vs. Random Forest), MAE = 0.28 ± 0.005 (19.5% reduction), and R 2 = 0.87 ± 0.008 (17.3% improvement). Pearson's r between predictions and true values was 0.93.

CONCLUSION

FAT-Net's attention-based fusion provides a robust, interpretable approach for forecasting daily wellbeing in older adults.