← Research
Nov 2025 – Present Neuroengineering

LearnTune: Closed-Loop Music Neurofeedback

Student Researcher · Shing-Tung Yau High School Science Award project

I designed and built a closed-loop musical neurofeedback prototype — a custom scalp-EEG headband combined with camera-based sensing — to study whether music that responds to a learner’s state can steady that state during learning.

Role
Student Researcher (first author)
Timeline
Nov 2025 – Present
Collaborator
Ting Jiang, Faculty of Psychology, Beijing Normal University
Participants
29 behavioral · 24 EEG state analysis
Tools
Custom PCB · nRF52840 · Python · Machine learning
Read the paper (PDF) Download
Diagram of the LearnTune closed-loop signal flow, with photos of the custom EEG headband PCB and the experimental setup
System architecture: multimodal sensing → calibration and fusion → Kalman filtering → negative-feedback music control.

The question

Two learners can have the same average level of arousal but very different moment-to-moment trajectories. Research has linked this variability to attention and memory, but an association is not an intervention. I wanted to test whether music that continuously adjusts to a learner’s own state could reduce that variability — without asking the learner to actively self-regulate.

What I built

  1. 01

    Custom EEG headband

    A PCB with a KS1092 dual-channel EEG analog front end, an MS88SF3 (nRF52840) Bluetooth MCU, a BMI260 6-axis IMU, battery management and a low-noise regulator.

  2. 02

    Multimodal sensing

    A frontal EEG power ratio, camera-derived heart rate and facial arousal, combined into one composite state index.

  3. 03

    Calibration & fusion

    Each learner is calibrated to their own reference centre and scale; inputs are weighted by their real-time signal quality.

  4. 04

    Closed-loop control

    Kalman filtering and temporal smoothing feed a negative-feedback controller that continuously adjusts the acoustic and note parameters of the music.

The experiment

A within-participant design compared closed-loop music driven by each participant’s real-time state with music from the same engine using static parameters. Each participant completed four word-learning blocks, two per condition, followed by recognition tests and attention probes.

Of 34 sessions, 30 produced complete records; after data-quality exclusions, 29 participants entered the behavioral analysis and 24 — with sufficient raw EEG — entered the primary state analysis.

What I found

  • Over the full feedback period, the standard deviation of the composite index was 12.8% lower under closed-loop feedback (Wilcoxon p = 0.0231; paired t-test after BH correction p = 0.0554). Because the two tests disagree after correction, this is an exploratory result, not a confirmed effect.
  • No significant difference was found in the mean level; changes in individual channels appeared mainly in facial arousal.
  • No improvement was detected in recognition discriminability or reaction time — so the paper presents the system and its measurement limits, not a learning benefit.

Gallery

Custom EEG headband circuit board with labelled components
Headband PCB with labelled components
The finished EEG headband
The finished headband
Paired chart of within-block standard deviation for each participant
Paired within-participant comparison (N = 24)

Video

Experiment sessions

Paper

LearnTune: An Exploratory Pilot Study of Multimodal Closed-Loop Music Biofeedback During Word Learning

Fan, Z., & Jiang, T. Manuscript (LNCS format). Co-authored with Ting Jiang, Faculty of Psychology, Beijing Normal University.

Open PDF →

“This was my first time carrying a research idea through the full pipeline. The hardest problem in neuroengineering is often not model accuracy, but the tension between real-world signal noise and experimental control.”