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

This paper presents a prototype system that uses Large Language Models (LLMs) to transform raw EEG and cardiovascular data into interpretable therapeutic reports and personalized music recommendations, aiming to democratize home-based music therapy by making physiological progress tracking accessible to non-experts.

The Gist

Imagine you’ve just finished a relaxing session with a high-tech "meditation pod" at home. You feel great, but you’re left wondering: “Was that actually working? How did my brain react to that music? What should I listen to tomorrow to feel even better?”

Usually, answering those questions requires an expensive brain doctor and a room full of medical wires. This paper describes a way to bring that "brain doctor" expertise into your living room using Artificial Intelligence.

Here is the breakdown of how they did it:

1. The Problem: The "Language Barrier" of the Brain

Think of your brain waves (EEG) like a massive, complex symphony playing inside your head. To a regular person, it just sounds like static or noise. To a scientist, it’s a series of complex mathematical graphs.

Currently, there is a huge "language barrier":

• The Brain speaks in electrical pulses.
• The User speaks in feelings ("I feel stressed" or "I feel calm").
• The Music speaks in beats and melodies.

Right now, there isn't a "translator" that can take those electrical pulses and turn them into a helpful, friendly conversation.

2. The Solution: The AI "Translator"

The researchers built a system that acts like a highly skilled, multilingual interpreter.

Instead of just giving you a confusing chart of squiggly lines, they use a Large Language Model (the same kind of technology behind ChatGPT) to act as a Music Therapy Expert.

Here is how the "Translation Pipeline" works:

• Step 1 (The Ears): A simple, low-cost headband listens to your brain waves and a small sensor checks your heart rate.
• Step 2 (The Math): A specialized computer program "cleans up" the noise and calculates the "rhythm" of your brain (e.g., are your "relaxation waves" going up or down?).
• Step 3 (The Brain/LLM): This is the magic part. The system feeds those math results into the AI. The AI doesn't just see numbers; it reasons. It says, "Aha! The user's 'anxiety waves' are high, but their 'relaxation waves' are increasing. This means the music is helping, but they are still a bit restless."
• Step 4 (The Voice): The AI writes you a beautiful, easy-to-read report in plain English, telling you exactly what happened in your session.

3. The "Personalized DJ"

The system doesn't stop at just talking; it also acts as your Personalized DJ.

If the AI detects that your heart is racing or your brain is stuck in a "high-stress" frequency, it doesn't just suggest "calm music." It uses a technique called RAG (Retrieval-Augmented Generation) to look through a library of music and pick a specific track that matches your exact biological needs—like a tailor picking the perfect fabric to fit your specific body shape.

4. Why does this matter? (The "Democratization")

The authors use the word "Democratizing." In plain English, this means "making it available to everyone."

Instead of needing a $10,000 medical setup and a PhD to understand your mental health, this system aims to work with cheap, consumer-grade gadgets you can wear at home. It turns "data" into "wisdom," helping regular people track their mental progress over time, just like you might track your steps on a Fitbit.

Summary Metaphor

Think of this system as a Smart Fitness Coach for your Soul.

• The EEG headband is your heart-rate monitor.
• The Signal Processing is the math that counts your reps.
• The LLM (AI) is the coach who looks at your stats and says, "Hey, you're doing great, but you're looking a bit tired today. Let's try some slower, gentler music to help you recover."

Technical Summary

Technical Summary: Democratizing Music Therapy via LLM-Based Automated EEG Analysis

1. Problem Statement

Traditional music therapy relies heavily on domain experts to interpret complex physiological signals (such as EEG) to assess therapeutic efficacy. This creates two significant barriers:

• Scalability & Cost: Professional EEG analysis is expensive and difficult to implement in home settings.
• Interpretability Gap: While consumer-grade wearable devices can collect rich physiological data (EEG, heart rate, SpO2), most users lack the expertise to translate raw data into actionable insights. They cannot determine if a session was effective or how to adjust their therapy.
• Limitation of Existing Systems: Most current "physiology-driven" music systems focus on real-time adaptation (modulating music during listening) but neglect post-hoc analysis and interpretable reporting, which are essential for long-term progress tracking.

2. Methodology

The authors propose an end-to-end intelligent agent system integrated into an "AI-powered music therapy pod." The framework moves away from direct LLM processing of raw signals (which leads to hallucinations) and instead uses a tool-augmented reasoning agent approach.

A. System Architecture:
The system consists of three primary components:

1. Signal Processing Module: A lightweight module that performs digital signal processing (DSP) on raw EEG and heart rate data.
   • EEG Processing: Uses band-pass filtering (Butterworth) to decompose signals into Delta, Theta, Alpha, and Beta bands. It employs the Hilbert transform to extract amplitude envelopes and Fast Fourier Transform (FFT) to identify dominant frequencies. It also compares median power between the first and second halves of a session to determine energy trends (Increasing/Decreasing).
   • Cardiovascular Processing: Calculates average Heart Rate (BPM) and SpO2.
2. LLM-Powered Cognitive Reasoning Agent: Utilizing the Qwen3 32B model, the agent acts as a "Brain Signal Analysis and Music Therapy Expert." It uses a function-calling mechanism to invoke the DSP tools. This prevents the LLM from "hallucinating" signal patterns by forcing it to rely on structured numerical outputs (trends and frequencies) provided by the tools.
3. Physiologically-Informed Music Recommendation Engine: A Retrieval-Augmented Generation (RAG) system. It maps the interpreted neurophysiological state and the user's average BPM to a curated library of music (categorized by classical, electronic, and Chinese traditional styles) to provide personalized recommendations.

B. Workflow:

• Input: Raw EEG voltage and pulse oximetry data.
• Process: The LLM receives the file path $\rightarrow$ invokes process_eeg tool $\rightarrow$ receives JSON results $\rightarrow$ synthesizes a natural language report $\rightarrow$ selects music via RAG.
• Output: A human-readable report (analyzing mental states) and specific music recommendations (e.g., "Theta-wave pulsed sounds" or specific BPM-matched tracks).

3. Key Contributions

• Novel Task Definition: The paper explores the unexplored task of using LLMs for automated physiological report generation specifically for music therapy.
• Hybrid Modeling Approach: It bridges the gap between low-level signal processing (DSP) and high-level semantic reasoning (LLM) through a tool-augmented agent framework.
• Democratization Framework: It provides a blueprint for low-cost, home-deployable music therapy that offers clinical-style insights without requiring a professional intermediary.
• Integrated Multimodal Pipeline: Unifies EEG-based neural markers with cardiovascular (BPM) indicators for a holistic therapeutic recommendation.

4. Results

• Expert Evaluation: In a blind test, the proposed LLM-agent system significantly outperformed baseline models. It achieved a mean score of 15.15, compared to 12.73 for Qwen-plus and a much lower 2.12 for GPT-5 (which struggled with long-sequence signal analysis without specialized tools). Experts rated the proposed method higher in clinical appropriateness, coherence, and therapeutic relevance.
• User Study: The system demonstrated high alignment with user emotional states. Users in "high-emotion" groups (those feeling more relaxed/focused) received reports with significantly higher positive keyword scores (11.00) compared to "low-emotion" groups (9.11), validating the system's ability to objectively reflect psychophysiological states.

5. Significance

This work demonstrates that LLMs, when combined with specialized computational tools, can transform "black-box" physiological data into meaningful, personalized therapeutic guidance. By shifting the focus from real-time modulation to post-session interpretability, the research paves the way for scalable, user-centered mental health technologies that allow individuals to monitor and manage their own neurophysiological well-being in a home environment.