As the global population ages, the need for scalable, proactive solutions to support health and wellbeing among older adults has become increasingly urgent. By 2050, adults aged 60 and over are projected to comprise 22% of the global population (1). This demographic shift is accompanied by a rise in chronic diseases and cognitive decline (2, 3), placing immense pressure on healthcare systems. In this context, non-pharmacological interventions such as music therapy have shown promise in promoting mental and emotional well-being, particularly by enhancing mood, memory, and social engagement in older adults (1, 3, 4). Additionally, music listening has been linked to modulation of physiological indicators like heart rate variability (HRV) and stress biomarkers (4, 5).

Advancements in wearable technology enable the continuous collection of physiological signals such as heart rate and HRV (6, 7). These data offer new opportunities to develop personalized digital therapeutics that can adapt in real-time to individual needs. Machine learning and deep learning models—including Random Forest (8), XGBoost (9), LSTM (10), and TCN (11)—have been employed to extract insights from physiological data. However, most existing models fail to effectively integrate behavioral and physiological modalities (12). Furthermore, composite health indices that merge affective states (e.g., PANAS) and physiological metrics such as HRV are increasingly used to provide a holistic measure of wellbeing (6, 13).

Despite these advances, the application of music-driven health prediction remains underexplored in real-world aging care contexts. Digital platforms for older adult care often lack dynamic personalization and explainability. Bridging this gap requires novel methods that can simultaneously model complex multimodal data and offer interpretable outputs to support clinician and user trust (7, 14).

Prior research has examined the role of music in health interventions across diverse domains. Faulkner et al.(4) developed Rhythm2Recovery, a rhythmic music and reflection-based program that improved emotional regulation and social reconnection in trauma recovery settings. Davidoff (3) emphasized the parallels between musicianship and medical practice, suggesting music-based training can bolster stress resilience in healthcare professionals.

In the domain of wearable health, Groh et al.(15) introduced lightweight, explainable models for on-device symptom detection using mechano-acoustic signals. Their interpretability strategies mirror our attention-based approach to understanding physiological responses during music listening. Wang et al.(5) demonstrated that musical features such as valence and tempo can be extracted using deep learning and correlate with mental energy—a concept we extend to older adults users.

Meta-analyses by Raglio (1) confirm the effectiveness of music interventions in improving mood and reducing stress across older adult populations. Meanwhile, studies in educational and digital contexts (12, 16, 17) have highlighted how music paired with real-time feedback can enhance self-regulation and emotional wellbeing. Bulaj et al.(14) demonstrated the potential of combining pharmacological treatments with personalized music playlists to empower patients. Liu et al.(18) applied biofeedback to adapt music to passengers' real-time heart-rate states, illustrating the potential for responsive, music-driven systems.

Complementary works in healthcare education and reflective practice have shown that integrating expressive arts—including music—can improve empathy, engagement, and cognitive performance in trainees (19, 20). The PANAS scale, widely used in music therapy studies, provides a validated tool for capturing emotional states and forms a key input to our modeling approach (13).

Despite these promising developments, several challenges remain. Many prior studies rely on small, localized samples, which limits generalizability to broader older adult populations (1, 2). Moreover, few models integrate minute-level physiological signals with high-level behavioral summaries into a unified framework (12). Wearable devices also face resource constraints, requiring models that are both accurate and lightweight (15). Interpretability is another major concern, as black-box models may hinder clinical adoption without transparent mechanisms for decision-making (7, 15). Lastly, user variability in music preferences and sensor engagement demands adaptive systems that remain robust across diverse use cases (16, 18), while minimizing fatigue from self-reporting (13).

We aim to address these gaps by focusing specifically on older adult individuals living in community settings who can benefit from proactive, music-driven digital health interventions. While music therapy has proven beneficial, most existing digital health models ignore music behavior as a variable of interest (1). Likewise, although heart-rate monitors collect minute-level data, they are often underutilized in fusion with subjective and behavioral features. Our proposed FAT-Net model bridges this divide by integrating physiological dynamics with music listening patterns, enabling interpretable predictions of next-day health outcomes. This integration not only improves forecasting accuracy but also supports clinical decision-making and user engagement through attention-based explanations.

We hypothesize that cross-modal attention in FAT-Net will significantly improve next-day health score prediction compared to unimodal baselines. Specifically, we expect reductions in RMSE and MAE greater than 15%, and improvements in R², with attention mechanisms highlighting meaningful interactions (e.g., HRV dips during high-tempo music). These interpretable insights are aligned with clinical understanding of stress responses and user engagement, and contribute to increased transparency and trust.

The key contributions of this paper are as follows:

To improve clarity for interdisciplinary readers, we provide a glossary of key technical terms and acronyms used throughout the manuscript. This glossary includes definitions for commonly used concepts such as FAT-Net, HRV, and PANAS. Readers unfamiliar with these terms may refer to Table 1 for concise explanations.

We recruited community-dwelling adults aged ≥60 years through a multi-pronged outreach strategy, which included flyers at senior centers, announcements at local health clinics, and targeted invitations via online forums and email lists. Prospective participants accessed a secure Google Form link where they provided informed consent (IRB#2025-065) before enrollment. Of the 714 individuals approached, 132 (18.5%) initiated the survey; 92 (65% of initiators) completed daily reporting for the entire study duration (45 ± 10 days, mean ± SD). Dropout reasons (n = 40) were categorized as technical difficulties (30%), loss to follow-up (45%), and withdrawal of consent (25%). Participant demographics included a mean age of 67.8 ± 5.1 years, 58% female, and a baseline BMI of 26.4 ± 3.8 kg/m². This cohort size and adherence rate provided sufficient statistical power (>0.8) to detect moderate effect sizes (Cohen's d = 0.5) in health-score changes over time.

Data were captured via two complementary modalities:

We engineered a comprehensive set of daily features spanning demographics, music intervention characteristics, physiological signals, and psychological outcomes. Table 2 details each feature, its source, and collection frequency. Table 2 shows that our dataset balances self-reported and sensor-derived measures to capture multidimensional aspects of participant health and behavior.

To ensure data integrity and analytic validity, we applied the following refinement steps:

This multi-step pipeline yielded a clean, analysis-ready dataset, with balanced representation across key variables.

To augment the limited time-series data from 92 participants, we applied four augmentation techniques. Table 3 summarizes these approaches. It shows how each method modifies feature distributions to enhance model generalizability.

The primary target variable of our predictive models is a day-level composite Health Score, integrating three key dimensions of wellbeing for each participant on day i:

To improve the reproducibility of our methods, we provide further details on the instrumentation and procedures used in the study. All physiological data were collected using the Empatica E4 wristband, a medically validated wearable device equipped with a 64 Hz photoplethysmography (PPG) sensor. This sensor captures continuous heart-rate data and enables the extraction of heart rate variability (HRV) metrics such as RMSSD and SDNN. The E4 also records movement via a 3-axis accelerometer, which was used to compute sleep-related parameters including total sleep time and sleep efficiency. Participants were instructed to wear the device on their non-dominant wrist during waking hours, removing it only for charging or bathing. Before data collection began, each participant received a brief orientation on proper usage of the device and how to complete the daily online survey. The survey included self-reported measures of affect (PANAS), music listening logs (track title, listening time, enjoyment rating), and perceived sleep quality. Physiological data were preprocessed using Empatica's SDK to ensure consistency and accuracy. Raw signals were cleaned, and outliers (e.g., implausible heart-rate values) were removed prior to feature extraction. This multimodal framework allowed us to integrate objective minute-level signals with self-reported behavioral data, providing a comprehensive view of each participant's daily health status.

Let each participant-day i be represented by:

where XiTS contains minute-resolution signals (e.g., heart rate, music on/off) of length T, and xiDS comprises aggregated daily summaries (e.g., BPM, sleep efficiency).

Multi-head self-attention highlights salient temporal segments:

This mechanism allows the model to focus on critical heart-rate fluctuations during or after music sessions.

To bring in high-level behavioral context, we encode daily summaries via a lightweight MLP:

Fusing modalities via attention enables bidirectional contextualization:

The final fused embedding F_i feeds into a regression head:

Model parameters θ are optimized by minimizing:

where y_i+1 is the true Health Score, and λ (set to 1 × 10⁻⁵) controls weight decay. We train using AdamW (lr = 3e-4, batch size = 16) with early stopping on validation MAE.

FAT-Net is built to address the complex, multimodal nature of predicting health status in older adults. It processes two complementary data streams: minute-level physiological signals such as heart rate and music on/off states, and daily-level summary features like sleep efficiency and average tempo of music. These are passed through separate encoders and later fused via cross-modal attention, enabling the model to learn both intra- and inter-modal interactions relevant to next-day health prediction. The time-series stream captures short-term temporal patterns using stacked convolutional layers and BiLSTM units, allowing the model to identify changes in physiological signals during or after music listening. Meanwhile, the summary stream captures behavioral context from features like PANAS scores, sleep metrics, and music characteristics using a lightweight MLP. The fusion layer integrates both streams through bi-directional attention, followed by a regression head that outputs the predicted health score. We evaluate FAT-Net's performance using standard regression metrics: Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R² (coefficient of determination). These metrics assess the accuracy and consistency of the model's predictions. We compare FAT-Net to a set of well-established baseline models: Random Forest and XGBoost represent strong classical methods suited for tabular features, while LSTM and TCN provide competitive deep learning alternatives for sequence data. Our experiments demonstrate that FAT-Net not only achieves lower error rates but also provides interpretable insights via attention mechanisms, linking music behavior to health outcomes.

All models were implemented in Python 3.8 using PyTorch 1.12 for deep networks and scikit-learn 1.1 for tree-based methods, running on an NVIDIA RTX 3090 GPU (24 GB VRAM), Intel Core i9-11900K CPU, and 64 GB RAM. The dataset was split into 80% training, 10% validation, and 10% testing sets. We trained for up to 100 epochs with early stopping (patience = 10) for LSTM (10), TCN (11), and FAT-Net; batch size was 16, optimizer was AdamW with learning rate 3 × 10⁻⁴ and weight decay 1 × 10⁻⁵. XGBoost (9) used 100 trees, max depth 6, and learning rate 0.1; Random Forest (7, 8) used 100 estimators and max depth 10.

Figure 5 shows that FAT-Net achieves the lowest RMSE across 10 independent runs, significantly outperforming Random Forest (7, 8), XGBoost (9), LSTM (10), and TCN (11). Figure 6 illustrates tighter MAE distributions for FAT-Net, indicating more consistent prediction accuracy. Figure 7 demonstrates that FAT-Net explains more variance (higher R²) in next-day Health Score than all baselines. Figure 8 highlights average percentage improvements of FAT-Net over each baseline across RMSE, MAE, and R², with the greatest gains against Random Forest. Finally, Figure 9 presents a strong alignment between predicted and actual Health Scores (Pearson's r = 0.93), confirming FAT-Net's calibration and generalizability.

In this study, we demonstrated that the proposed Fusion-Attentive Temporal Network (FAT-Net) significantly outperforms conventional baselines, like Random Forest, XGBoost, LSTM, and TCN, in predicting next-day composite health scores for older adults based on minute-level physiological signals and music-listening behavior. The notched boxplots (Figures 5–7) and the improvement heatmap (Figure 8) confirm that FAT-Net reduces prediction error by up to 23% and increases explained variance by up to 17%. Our cross-modal attention mechanism enables the model to dynamically weight salient heart-rate fluctuations during music sessions and high-level summary features such as sleep efficiency, resulting in more robust and interpretable forecasts. The strong alignment of predicted vs. actual health scores (Figure 9, Pearson's r = 0.93) further attests to FAT-Net's calibration and practical utility.

Beyond demonstrating model performance, our findings also offer insights into the health effects of music-based interventions for older adults. The attention visualization results reveal that the model consistently attends to moments of elevated heart-rate variability and specific music characteristics, such as increased tempo or valence, during and after listening sessions. These patterns correspond with existing literature showing that upbeat or emotionally engaging music can elevate mood, reduce stress, and support autonomic regulation. For example, attention peaks often aligned with post-listening heart-rate stabilization or during periods of high arousal music, suggesting potential physiological benefits of music exposure. This suggests that FAT-Net does not merely rely on technical time-series correlations but identifies semantically meaningful episodes where music engagement appears to mediate health-related changes. In this sense, the model helps illuminate the dynamic relationship between music behavior and well-being, offering a computational pathway to validate and interpret real-world music therapy effects. By coupling prediction with explainability, FAT-Net thus serves as both a forecasting tool and a mechanism to investigate the role of music in everyday health regulation. Future iterations of this work may further disentangle causal effects through controlled music intervention studies, but our current results already highlight the practical potential of integrating music behavior into digital health frameworks.

The superior performance of FAT-Net has several real-world implications. First, smartphone or wearable applications incorporating our model can deliver personalized music-therapy recommendations to older adults, adapting in real time to their physiological state and listening habits. Second, healthcare providers and caregivers could leverage daily health-score forecasts to monitor well-being remotely, triggering timely interventions, such as adjusting exercise regimens or recommending relaxation playlists, to prevent declines in mood or sleep quality. Finally, the interpretability afforded by the attention weights allows end-users and clinicians to understand which features (e.g., tempo, valence, HRV dips) most strongly influenced the prediction, fostering trust and facilitating shared decision-making in digital health platforms.

While our augmented dataset (92 participants with 4 × synthetic expansion) enabled thorough model training, the relatively small cohort size and self-selected sample may limit generalizability. Moreover, the reliance on self-reported PANAS scores and Google Form logging introduces potential reporting bias. Our minute-level “music on/off” signal did not account for nuances such as multitasking or background noise, which could affect physiological responses.

Building on these results, future work should (a) validate FAT-Net on larger, more diverse cohorts like different age groups and cultural backgrounds, to assess robustness; (b) integrate additional sensor modalities (e.g., skin conductance, accelerometry) to capture broader physiological and contextual cues; (c) explore online learning schemes that adapt model parameters as new user data arrive, supporting lifelong personalization; (d) implement real-time on-device inference for privacy-preserving mHealth deployments; and (e) conduct randomized controlled trials to measure the clinical efficacy of FAT-Net–driven music-therapy interventions in improving long-term health outcomes.