aBCMI: Affective Brain-Computer Musical Interface

Real-time affective Brain-Computer Musical Interface (BCMI) systems that generate adaptive music based on EEG frontal asymmetry (AF7/AF8 electrodes). This repository contains four complete experimental frameworks for investigating emotion recognition from EEG signals and generating music in real-time.

This repository is a direct contribution supporting the doctoral thesis:
"EEG-Driven Brain–Computer Musical Interfaces for Emotion Self-Induction: A Comparative Study of Deterministic and Machine Learning Approaches"

Repository URL: https://github.com/pamonroy/aBCMI/

Overview

This project implements four different approaches to EEG-based emotion recognition and affective music generation:

1. Asymmetric Frontal Activity Hypothesis (AFAH) - Real-time valence/arousal estimation using alpha/beta band power ratios
2. Multi-Layer Perceptron (MLP) - Neural network classification using power spectral features
3. Linear Discriminant Analysis (LDA) - 2×2 factorial design with statistical classification
4. EEGNet Deep Learning - Convolutional neural network with LSTM for end-to-end learning

Each implementation provides a complete experimental pipeline including real-time EEG acquisition, signal processing, machine learning/deep learning, and adaptive MIDI music generation.

Institutions

• Universitat Pompeu Fabra (Barcelona, Spain)
• Universidad Icesi (Cali, Colombia)

Technical Specifications

Software Environment

• Python Version: 3.10.9
• LSL Recording Software: Lab Recorder version 1.14.0
• Operating System: macOS / Linux (compatible)

Dependencies

Package	Version	Purpose
numpy	1.23.2	Numerical computations and array handling
pandas	2.2.2	Data management and CSV export
mido	1.2.10	MIDI communication for music generation
scikit-learn	1.5.1	MLP and LDA classifiers, StandardScaler
pynput	1.7.7	Keyboard input for self-report collection
pyxdf	1.16.3	XDF file loading for offline simulation
bitalino	1.2.6	BITalino hardware interface
pylsl	1.15.0	Lab Streaming Layer synchronization
tensorflow-macos	2.9.2	EEGNet deep learning framework (macOS)

Note for Linux/Windows users: Replace tensorflow-macos with tensorflow version 2.9.2 or higher.

Citation

If you use this software in your research, please cite:

@phdthesis{Monroy_2026_BCMI,
  author = {Pablo Andrés Monroy D'Croz},
  title = {EEG-Driven Brain–Computer Musical Interfaces for Emotion Self-Induction: A Comparative Study of Deterministic and Machine Learning Approaches},
  year = {2026},
  school = {Universitat Pompeu Fabra},
  address = {Barcelona, Spain},
  note = {Repository: https://github.com/pamonroy/aBCMI/}
}

@software{Monroy_aBCMI_2026,
  author = {Pablo Andrés Monroy D'Croz},
  title = {aBCMI: Affective Brain-Computer Musical Interface},
  year = {2026},
  url = {https://github.com/pamonroy/aBCMI/},
  version = {1.0},
  doi = {[Add DOI if available]}
}

Public Datasets

This thesis contributes four publicly available datasets collected from 82 participants across sequential BCMI experiments, totaling more than 50 hours of synchronized recordings. All datasets are available on Zenodo:

Dataset Overview

Dataset	Paradigm	Participants	Duration (approx.)	EEG Sampling Rate	Key Features
AFAH	Asymmetric Frontal Activity Hypothesis	23	14 min	1000 Hz	Dual-channel EEG, valence-arousal predictions
MLP	Multi-Layer Perceptron	23	20 min	1000 Hz	Labeled training data, trained models
EEGNet	Deep Learning	26	30 min	100 Hz	6 synchronized streams, real-time predictions, self-reports
LDA	Linear Discriminant Analysis	33	30 min	1000 Hz	Structured similar to EEGNet with higher sampling rate
TOTAL	All Paradigms	82	50+ hours	100-1000 Hz	Synchronized recordings + questionnaires

Included Materials

Each dataset includes:

• Raw EEG recordings (synchronized LSL streams)
• Experimental markers for trial segmentation
• Pre- and post-experiment questionnaires
• Trained models (where applicable)
• Documentation (data format specifications, experimental protocols)
• Code snippets for data loading and analysis

Zenodo Access

The complete datasets are publicly available at: https://zenodo.org/records/19059357

Table of Contents

• Features
• Hardware Requirements
• Software Dependencies
• Installation
• Experimental Methods
  • Method 1: AFAH (Asymmetric Frontal Activity Hypothesis)
  • Method 2: MLP (Multi-Layer Perceptron)
  • Method 3: LDA (Linear Discriminant Analysis)
  • Method 4: EEGNet Deep Learning
• Common Protocol
• MIDI Configuration
• Lab Streaming Layer (LSL) Configuration
• License
• Acknowledgments
• References
• Contact

Features

• Real-time EEG acquisition from BITalino device (AF7, AF8 electrodes)
• Multiple classification approaches:
  • AFAH (Asymmetric Frontal Activity Hypothesis)
  • MLP (Multi-Layer Perceptron)
  • LDA (Linear Discriminant Analysis)
  • EEGNet Deep Learning
• Offline simulation using pre-recorded XDF files
• Lab Streaming Layer (LSL) integration for data synchronization
• Adaptive MIDI music generation with emotion-controlled parameters:
  • Musical mode (7 Greek modes from Lydian to Locrian)
  • Tempo and rhythm density (controlled by arousal)
  • Loudness and brightness (controlled by emotional intensity)
  • Harmonic progression (I-IV-V-I pattern)
• 2×2 factorial design (sad/happy × music/no-music) for LDA and EEGNet
• Self-report mechanism (1–9 scale) for emotional state validation
• Comprehensive data logging for offline analysis

Hardware Requirements

• BITalino (r)evolution with EEG sensor kit
• Electrodes: AF7 (left prefrontal), AF8 (right prefrontal)
• Digital Audio Workstation (e.g., Ableton Live) for audio output
• Computer with Bluetooth capability for BITalino connection
• Keyboard for self-report responses (keys 1–9)

Software Dependencies

Core Dependencies with Version Specifications

Package	Version	Purpose
numpy	1.23.2	Numerical computations and array handling
pandas	2.2.2	Data management and CSV export
mido	1.2.10	MIDI communication for music generation
scikit-learn	1.5.1	MLP and LDA classifiers, StandardScaler
pynput	1.7.7	Keyboard input for self-report collection
pyxdf	1.16.3	XDF file loading for offline simulation
bitalino	1.2.6	BITalino hardware interface
pylsl	1.15.0	Lab Streaming Layer synchronization
tensorflow-macos	2.9.2	EEGNet deep learning framework (macOS)
matplotlib	3.4.3	Training curve visualization
scipy	1.7.3	Signal processing (Welch's method)

Important Notes

• LSL Recording: Lab Recorder version 1.14.0 was used for recording all LSL streams.
• Python Version: All code was developed and tested with Python 3.10.9.
• TensorFlow Compatibility:
  • On macOS, use tensorflow-macos
  • On Linux/Windows, use tensorflow==2.9.2

Installation

Clone the Repository

git clone https://github.com/pamonroy/aBCMI.git
cd aBCMI

Create a Virtual Environment (Recommended)

python3.10 -m venv eeg_env
source eeg_env/bin/activate

On Windows:

eeg_env\Scripts\activate

Install Dependencies

pip install -r requirements.txt

Verify Installation

python -c "import numpy, pandas, mido, sklearn, pynput, pyxdf, bitalino, pylsl, tensorflow; print('All dependencies installed successfully!')"

Configure Working Directory

• Update the os.chdir() path in each script to point to your project directory.
• Ensure the BITalino MAC address is correct in the connection code.

---

Experimental Methods

Method 1: AFAH (Asymmetric Frontal Activity Hypothesis)

File: afah.py

Real-time affective music generation based on EEG frontal asymmetry without machine learning. The system estimates emotional dimensions directly from EEG power spectra.

Key Features

• Real-time valence and arousal estimation
• No training required (works immediately)
• Normalization with baseline recording
• Smoothed valence (10-second moving average)

Emotion Metrics

Valence = (AF8_alpha / AF8_beta) - (AF7_alpha / AF7_beta)

Arousal = (AF7_beta + AF8_beta) / (AF7_alpha + AF8_alpha)

Protocol

• 15 seconds baseline (eyes closed)
• 20 trials: 10 happy + 10 sad (semi-randomized order)
• 30 seconds music per trial

Usage

python afah.py

---

Method 2: MLP (Multi-Layer Perceptron)

File: mlp.py

EEG-based emotion classification using a Multi-Layer Perceptron neural network trained on power spectral features.

Key Features

• 10 power spectral features (theta, alpha, low beta, high beta, gamma) extracted from AF7 and AF8
• MLP classifier with 30-30 hidden layers
• Calibration phase with 3 conditions (sad, happy, neutral)
• Online real-time classification

Calibration Protocol

• 6 trials: sad, happy, neutral (counterbalanced order)
• Each trial: 10s silence + 20s music
• Feature extraction and model training

Online Protocol

• 12 trials: 6 happy + 6 sad (randomized order)
• Real-time classification and adaptive music

Usage

python mlp.py

---

Method 3: LDA (Linear Discriminant Analysis)

File: lda.py

2×2 factorial design experiment using Linear Discriminant Analysis for emotion classification with music/no-music conditions.

Key Features

• 2×2 factorial design (sad/happy × music/no-music)
• LDA classifier with 10 spectral features
• 3-minute meditation phase
• Self-report after each trial (1–9 scale)
• 4-second and 10-second emotion smoothing buffers

Experimental Design

• Independent variables: Emotion (sad/happy) × Auditory Condition (music/no-music)
• Dependent variables: EEG features, self-report ratings, classifier accuracy

Protocol

• 3 minutes meditation (baseline)
• 6 calibration trials (sad, happy, neutral)
• 20 online trials (5 per condition, randomized)
• 40-second trials with self-report after each

Usage

python lda.py

---

Method 4: EEGNet Deep Learning

Files

• eegnet.py — Main experiment script
• eegnet_train.ipynb — Model training notebook

End-to-end deep learning using EEGNet architecture with LSTM for raw EEG classification.

Key Features

• Raw EEG signal processing (no hand-crafted features)
• EEGNet architecture with convolutional and LSTM layers
• 10-second windows, 100 Hz sampling rate
• Binary classification: sad vs. happy

Model Architecture

Input (2 channels × 1000 timesteps)
→ Conv2D (8 filters, 1×50)
→ DepthwiseConv2D (depth multiplier 4)
→ SeparableConv2D (16 filters)
→ LSTM (16 units)
→ Dense (2 units)
→ Output (1 unit, sigmoid)

Protocol

• 3 minutes meditation (baseline)
• 12 calibration trials (sad, happy, neutral)
• 24 online trials (6 per condition in 2×2 design)
• 30-second trials with self-report after each

Usage

python eegnet.py

---

Common Protocol

All four methods share a common experimental structure.

Phase 1: Setup

• Participant code entry
• LSL outlet initialization
• EEG hardware connection check

Phase 2: Volume Setting (Optional)

• Alternating sad/happy music
• Comfortable listening level adjustment

Phase 3: Meditation / Baseline (Methods 3 & 4 Only)

• 3-minute eyes-closed rest period
• Baseline physiological state acquisition

Phase 4: Calibration

• 10 seconds silence (idle baseline)
• 20 seconds fixed music (sad/happy/neutral)
• Counterbalanced order
• Self-report collection (Methods 3 & 4)

Phase 5: Model Training (Methods 2–4)

• Feature extraction and preprocessing
• Model training (MLP, LDA, or EEGNet)
• Model serialization for online use

Phase 6: Online Phase

• Real-time emotion prediction
• Adaptive music generation
• Randomized trial order (no consecutive duplicates)
• Self-report after each trial

Phase 7: Debriefing

• Experiment completion signals
• Data saving and cleanup

---

MIDI Configuration

Channel Mapping

Channel	Instrument	Purpose
0	Piano	Chord voicing (upper register)
1	Cello	Chord voicing (middle register)
2	Bass	Bass line
3	Piano	Melodic patterns
5	Control	MIDI CC messages (Auditory cues)

Control Change (CC) Messages

CC	Purpose	Value
13	Volume setting request	127
14	Calibration start	127
15	Online phase start	127
16	Meditation start	127
17	Experiment end	127
18	Stop all	127
19	Model training start	127
80–82	Target emotion (80=happy, 81=neutral, 82=sad)	127
83	Close eyes instruction	127
84	Self-report request	127
85	No-music condition	127
123	All notes off	0

---

Lab Streaming Layer (LSL) Configuration

LSL Streams

All methods create the following LSL streams (recorded with Lab Recorder 1.14.0):

• eeg-bitalino - Raw EEG data (2 channels, 1000Hz or 100Hz)
• markers - Experiment markers (integers)
• predicted-emotion - Predicted emotion values (2-3 channels, 2Hz)
• emotion - Predicted emotion fed to the music generator at the start of each bar
• self-report - Self-report ratings (EEGNet and LDA methods only)
• harmony - Chord progression markers (LDA method only)

Marker Codes

Marker	Value	Description
setVolume	6	Volume setting phase
silence	0	Silence period
neutral	2	Neutral emotion
sad	1	Sad emotion
happy	3	Happy emotion
modeling	8	Model training phase
online	7	Online phase start
music	9	Start music condition
min_power	-1	Minimum power alarm
pause	10	Pause between phases
askForReport	11	Self-report request
stopMusic	12	Stop music condition
startNoMusic	13	Start no-music condition
stopNoMusic	14	Stop no-music condition

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

Copyright (c) 2024–2026 Pablo Andrés Monroy D'Croz

---

Acknowledgments

• Lawhern et al. (2018) for the EEGNet architecture
  • EEGNet: A Compact Convolutional Network for EEG-based Brain-Computer Interfaces
• EEGNet implementation inspired by:
  • Pre-trained-EEG-for-Deep-Learning
• BITalino for providing open-source biomedical hardware
• The Lab Streaming Layer (LSL) community for synchronization tools and Lab Recorder 1.14.0
• Doctoral Committee:
  • Universitat Pompeu Fabra
  • Universidad Icesi
• Contributors:
  • This research was supported by the doctoral programs at Universitat Pompeu Fabra and Universidad Icesi

---

References

1. Lawhern, V. J., Solon, A. J., Waytowich, N. R., Gordon, S. M., Hung, C. P., & Lance, B. J. (2018).
   EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces.
   Journal of Neural Engineering, 15(5), 056013.

2. Davidson, R. J. (2004).
   What does the prefrontal cortex "do" in affect: Perspectives on frontal EEG asymmetry research.
   Biological Psychology, 67(1–2), 219–234.

3. Schirrmeister, R. T., Springenberg, J. T., Fiederer, L. D. J., Glasstetter, M., Eggensperger, K., Tangermann, M., ... & Ball, T. (2017).
   Deep learning with convolutional neural networks for EEG decoding and visualization.
   Human Brain Mapping, 38(11), 5391–5420.

---

Contact

Author: Pablo Andrés Monroy D'Croz

Institutions:

• Universitat Pompeu Fabra (Barcelona, Spain)
• Universidad Icesi (Cali, Colombia)

GitHub: pamonroy

Repository:
https://github.com/pamonroy/aBCMI/