msna-detect 0.1.2

A deep learning framework for automated detection of bursts in Muscle Sympathetic Nerve Activity.

MSNA Burst Detection

A deep learning framework for automated detection of bursts in Muscle Sympathetic Nerve Activity (MSNA) signals.

Overview

Muscle Sympathetic Nerve Activity (MSNA) provides direct measurement of sympathetic outflow to skeletal muscle vasculature, offering insight into neural mechanisms governing cardiovascular control. This repository presents a robust and efficient deep learning framework for automated detection of bursts in MSNA signals, addressing a challenging task that traditionally requires time-consuming manual annotation by experts. Our approach utilizes a 1D convolutional neural network based on U-Net architecture that reformulates burst detection by modeling probabilistic distributions around burst points. The framework exploits CNN translation invariance with random window sampling for efficient training, and can process recordings of arbitrary length without boundary artifacts during inference. In fact, models can be trained on most laptop CPU in under 4 minutes.

This library provides a simple API for both training and inference, along with pre-trained models for immediate application.

Installation

You can install the package from PyPI:

pip install msna-detect

Or you can install from source:

# Clone the repository
git clone https://github.com/ryanirl/msna-detect.git
cd msna-detect

# Install dependencies
pip install -r requirements.txt

# Install the package
pip install -e .

See requirements.txt for the full list of dependencies.

Command Line Interface

This directory contains command-line scripts for training, evaluating, and using the MSNA burst detection model. These scripts provide a convenient way to work with the MSNA detection library without writing Python code.

Before using these scripts, you need to install the MSNA detection library.

For the data format. The scripts expect CSV files with the following columns:

• Integrated MSNA: The MSNA signal values
• Burst: Binary annotations of true bursts (1 for burst, 0 for no burst)

For prediction output, the following columns are added:

• Predicted Burst: Binary predictions of bursts
• Predicted Probability: Probability scores for each prediction

Training Script

Trains a new MSNA burst detection model on your data.

msna-detect-train -i /path/to/training/data -o /path/to/save/model.pt [options]

Prediction

Uses a trained model to detect bursts in MSNA signals.

msna-detect-predict -i /path/to/input.csv -o /path/to/output.csv -m /path/to/model.pt [options]

Evaluation

Evaluates model performance on a test dataset.

msna-detect-eval -i /path/to/test/data -m /path/to/model.pt [options]

All scripts can be run with the -h of --help tag to get the additional options.

Codebase Quick Start

Inference with Pre-trained Model

from msna_detect import MsnaModel

# Load your MSNA signal (should be shape [channels, time] or [time])
signal = ...

# Load the pre-trained model
model = MsnaModel.from_pretrained("pretrained/model.pth")

# Get burst probabilities
burst_probs = model.predict_proba(signal)

# Find burst peaks
burst_locations = model.find_peaks(signal, height = 0.4, distance = 100)

print(f"Found {len(burst_locations)} bursts")

Training a New Model

from msna_detect import MsnaModel

# Prepare your training data
signals = [signal1, signal2, ...]  # List of numpy arrays, each of shape [channels, time]

bursts = [burst1, burst2, ...]     # List of numpy arrays with binary burst annotations

# Create and train a model
model = MsnaModel(sampling_rate = 250, device = "cuda")
model.fit(
    train_signal = signals,
    train_bursts = bursts,
    epochs = 50,
    lr = 0.01,
    batch_size = 32
)

# Save the trained model
model.save("my_trained_model.pth")

For more advanced usage examples, see the examples/ directory.

Dataset Format

The model expects MSNA data as NumPy arrays. For training, both signals and burst annotations should be provided:

# Example format for training data
signals = [signal1, signal2, ...]  # List of numpy arrays, each of shape (time,)
bursts = [burst1, burst2, ...]     # List of numpy arrays with binary burst annotations

Visualization Dashboard

The package includes an interactive visualization dashboard built with Bokeh that allows you to explore MSNA signals and their detected bursts. The dashboard provides:

• Interactive visualization of the integrated MSNA signal
• Overlay of true and predicted burst locations
• Probability distribution of burst predictions
• Navigation tools for exploring long recordings
• Hover tools for detailed signal inspection

To use the dashboard you can run the dashboard on a CSV file containing MSNA data

python dashboard.py --input path/to/your/data.csv

You can also export the dashboard as an HTML file

python dashboard.py --input path/to/your/data.csv --save

The input CSV file should contain the following columns:

• Integrated MSNA: The normalized MSNA signal
• Burst: Binary annotations of true bursts
• Predicted Burst: Binary annotations of predicted bursts
• Predicted Probability: Probability scores for burst predictions

A file like this can be generated from the scripts/predict.py file.

Architecture

Our approach utilizes a 1D convolutional neural network based on the U-Net architecture:

1. Input: MSNA integrated signal. Normalization is done by the model.
2. Processing: Four downsampling stages with ResNet-like encoder blocks followed by corresponding upsampling stages with skip connections.
3. Output: Probability map where peaks correspond to detected bursts. Peak finding is done on this probability map to get out predicted bursts.

The model is trained by:

1. Transforming sparse binary annotations into soft distributions using Gaussian convolution
2. Random window sampling for computational efficiency
3. Optimizing with mean squared error loss
4. Post-processing with calibration to ensure consistent probability scaling

License

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

Citation

If you use this code in your research, please cite our paper:

@article{msna2025,
  title={Muscle Sympathetic Nerve Activity Burst Detection},
  author={Peters, Ryan},
  year={2025}
}