Democratizing Music Therapy: LLM-Based Automated EEG Analysis and Progress Tracking for Low-Cost Home Devices

Abstract: Home-based music therapy devices require accessible and cost-effective solutions for users to understand and track their therapeutic progress. Traditional physiological signal analysis, particularly EEG interpretation, relies heavily on domain experts, creating barriers to scalability and home adoption. Meanwhile, few experts are capable of interpreting physiological signal data while also making targeted music recommendations. While LLMs have shown promise in various domains, their application to automated physiological report generation for music therapy represents an unexplored task. We present a prototype system that leverages LLMs to bridge this gap -- transforming raw EEG and cardiovascular data into human-readable therapeutic reports and personalized music recommendations. Unlike prior work focusing on real-time physiological adaptation during listening, our approach emphasizes post-session analysis and interpretable reporting, enabling non-expert users to comprehend their psychophysiological states and track therapeutic outcomes over time. By integrating signal processing modules with LLM-based reasoning agents, the system provides a practical and low-cost solution for short-term progress monitoring in home music therapy contexts. This work demonstrates the feasibility of applying LLMs to a novel task -- democratizing access to physiology-driven music therapy through automated, interpretable reporting.

• The paper demonstrates a novel LLM-based approach for automated EEG analysis and progress tracking in home-based music therapy.
• It integrates consumer-grade EEG and cardiovascular sensors with advanced signal processing and frequency analysis to generate personalized music recommendations.
• The system achieved superior expert evaluations and objective emotional correlations, underscoring its potential to democratize accessible music therapy.

LLM-Driven Democratization of Music Therapy: EEG-Based Progress Tracking for Home Devices

Introduction

The paper "Democratizing Music Therapy: LLM-Based Automated EEG Analysis and Progress Tracking for Low-Cost Home Devices" (2601.12280) proposes a framework leveraging LLMs for the automated interpretation of physiological signals—specifically EEG and cardiovascular data—to facilitate music therapy progress tracking and therapeutic recommendation in non-clinical, home settings. The work addresses the ongoing challenge of translating raw biosignal data into actionable, interpretable feedback for non-expert users, bridging the gap between consumer-grade hardware and expert-level analytics.

Figure 1: Pipeline of the proposed LLM-based framework, integrating signal acquisition, feature extraction, LLM-based reasoning, and personalized music recommendation.

Motivation and Background

Music therapy has shown efficacy for a spectrum of neuropsychiatric conditions and general wellness. However, its home deployment faces two principal obstacles: (1) the need for expert interpretation of complex physiological signals, and (2) the lack of scalable, adaptive approaches to music selection based on individualized therapy outcomes. Existing research has mainly focused on real-time physiological feedback systems [dongre2024integratingphysiologicaldatalarge, jiao_advancing_2025] or clinical-grade post-hoc assessment, with limited emphasis on interpretable, accessible reporting for lay users.

The presented system combines affordable EEG and heart rate acquisition hardware with a multimodal AI pipeline, centering around post-session analytics and personalized feedback, enabling self-monitoring and adjustment of therapeutic interventions.

Hardware and Signal Acquisition

The therapeutic setup described includes a single-channel EEG headband positioned at the prefrontal cortex (AF3, dry electrode, 256–512 Hz sampling), and a consumer-grade pulse oximeter for HR and $\mathrm{SpO}_2$ acquisition.

Figure 2: A user wearing the single-channel EEG headband targeting prefrontal activity relevant to emotional and attentional modulation.

The integration of these sensors allows acquisition of neural oscillations correlated to affective states, alongside cardiovascular markers that can be directly mapped to music tempo for personalized recommendations.

Figure 3: The heart rate and blood oxygen saturation probe used for cardiovascular parameter acquisition.

Figure 4: The hardware connections setup of the Therapy Pod, illustrating the flow from sensors to processing and the immersive output systems.

Signal Processing and EEG Feature Extraction

Signal preprocessing comprises DC offset correction, band-pass filtering (Butterworth, canonical $\delta$, $\theta$, $\alpha$, $\beta$), Hilbert transform-based envelope extraction, and time-domain analysis of power variations across session segments. The system computes median amplitude shifts per band, providing trend analysis of neural states—identifying relaxation, attentional engagement, and potential hyperactivation markers.

Dominant frequency extraction via FFT further contextualizes the oscillatory dynamics, supporting intervention planning across frequency bands (e.g., increasing theta for introspection, reducing beta for anxiety). This low-computation extraction makes the framework suitable for consumer hardware lacking multi-channel clinical instrumentation.

LLM-Based Report Generation and Therapeutic Recommendation

The reasoning core employs the Qwen3 32B LLM [yang2025qwen3technicalreport], interfacing with domain-specific computational routines via function calls in accordance with the OpenAI protocol. Prompt engineering ensures the LLM can contextualize extracted features, integrating neuroscientific theory for interpreting user states, generating comprehensive yet intelligible reports, and mapping these states to specific music therapy interventions.

Sample generated reports reflect nuanced multi-band EEG analysis, followed by targeted recommendations:

• Theta resonance (e.g., 5.33 Hz pulsed sounds for internal focus)
• Alpha-based ambient reinforcement (e.g., rainfall soundscapes)
• Beta modulation via acoustic braking (violin harmonics for downregulation)
• Binaural beat configuration for functional rebalancing (dual-frequency protocol)

Music recommendations leverage both EEG-driven insights and HR-based tempo matching—facilitating scenario-specific selections (relaxation, physical activity) via a retrieval-augmented generation (RAG) architecture [gao2024retrievalaugmentedgenerationlargelanguage].

Music Corpus and Recommendation Strategy

All recommended music is custom-composed, spanning classical, electronic, and Chinese traditional genres, with strict parameter control (BPM, timbral qualities, dynamic range) to enable robust mapping and assessment. BPM-aligned recommendation directly exploits the relationship between HR and music-induced entrainment effects [van_dyck_adopting_2017, bason_control_1972].

Experimental Validation

User Experience and Data Collection

A typical session involves users immersed in a 6-minute soundscape, with synchronized visualizations and real-time biosignal logging.

Figure 5: A typical music therapy experience in the immersive Therapy Pod.

Expert Review and Ranking

Blind evaluations were performed by neuroscience and music therapy specialists. The LLM-driven pipeline with integrated signal processing tools achieved a mean expert score of 15.15, surpassing direct LLM prompting baselines (Qwen-plus: 12.73; GPT-5: 2.12)—underscoring the importance of structured feature extraction and domain-augmented interaction.

User Study and Emotional Correlation

Among 33 volunteer participants, emotional keyword analysis demonstrated the system’s alignment with reported subjective states. Higher baseline emotion groups yielded greater positive tendency scores (8.20 vs. 6.89), evidencing the framework’s capacity to reflect and track psychophysiological shifts objectively, irrespective of user self-report.

Implications and Future Directions

The presented architecture operationalizes physiology-driven music therapy in consumer contexts, setting benchmarks for interpretability and accessibility in biofeedback applications. Its modular design—decoupling feature extraction from reasoning and recommendation—supports extensibility to additional biomarkers and therapeutic protocols.

Potential avenues include:

• Extending music database diversity for user preference adaptability
• Real-time integration of musical features with physiological monitoring for closed-loop interventions
• Personalization algorithms reflecting session-to-session dynamics and long-term progress
• Transitioning fixed procedural workflows to adaptive, user-tailored therapeutic routines

The framework demonstrates feasibility for home-based, scalable music therapy, with implications for digital therapeutics, self-regulation, and decentralized wellness interventions.

Conclusion

This work introduces an authoritative end-to-end physiological signal analytics and music recommendation system, utilizing LLMs for post-hoc interpretation and report generation, validated by expert and user assessments. By uniting consumer-grade hardware, robust signal processing, and multimodal AI reasoning, the framework sets the stage for future research and deployment of democratized, interpretable music therapy—enabling longitudinal self-monitoring, optimized recommendation, and broader accessibility to effective non-pharmacological interventions in mental health and wellness domains.