EMG Gesture Classification: A Deep Learning Approach to Biological Signal Processing

Overview

This project implements a comprehensive Data Science pipeline for the classification of Electromyography (EMG) signals. The objective is to decode motor intent from neuromuscular activity by transforming high-dimensional time-series data into spectral feature maps for Convolutional Neural Network (CNN) modeling.

Data Science Pipeline

1. Feature Engineering: Spectral Transformation

Raw EMG signals are non-stationary and stochastic. To extract meaningful features, we employ a frequency-domain transformation:

• Preprocessing: Signal normalization (z-score) and rectification to stabilize variance.
• Windowing Strategy: 200ms segmentation with 50% overlap to preserve temporal dependencies while increasing the dataset size for model robustness.
• Spectrogram Generation: Transformation of 1D signals into 2D log-mel spectrograms (64 mel-bins). This maps biological activity into a representation that highlights frequency-modulated patterns unique to specific gesture classes.

Neurocomputation Context

Beyond the Data Science framework, this pipeline is grounded in neurocomputational principles:

• Motor Unit Recruitment: The spectrograms capture the firing rates and recruitment patterns of motor units within the forearm musculature. By analyzing the spectral envelope, the model decodes the physiological intensity and coordination of muscle groups.
• Signal Rectification & Intent: Full-wave rectification is utilized as a non-linear transformation to extract the neural drive (the signal envelope) from the high-frequency firing patterns, effectively mapping "raw electrical noise" back to "biological motor intent."
• Temporal Neural Dynamics: The 200ms windowing strategy is biologically informed, mirroring the typical integration time of the human motor control system for gesture execution.

2. Dataset Management & Balancing

A critical challenge in biological signal classification is class imbalance and sample scarcity.

• Dynamic Dataset Wrapper: Custom PyTorch SpectrogramDataset implementation for efficient memory management and real-time data loading.
• Class Imbalance Mitigation: Implementation of Inverse Frequency Weighting within the nn.CrossEntropyLoss function. This ensures that the model does not overfit to the majority class (e.g., 'Rest') and maintains high sensitivity to rarer gesture events.
• Validation Strategy: Stratified 80/20 train-test split with fixed seeding for reproducibility.

Neural Architecture Design

The core model is a custom ConvNet designed for low-latency inference on time-series spectral features.

Architecture Specifications:

• Feature Extractor: Three convolutional layers with increasing filter depth (32, 64, 128).
• Regularization: Batch Normalization and 0.3 Dropout to prevent overfitting on small-scale biological datasets.
• Adaptive Pooling: Global Average Pooling (GAP) is used to ensure the model is agnostic to varying window lengths and to reduce the total parameter count to ~155k, optimizing for edge-device deployment.

Statistical Evaluation

Model performance is evaluated using metrics that go beyond simple accuracy to account for class distribution and classification rigor.

Key Performance Indicators:

Metric	Result	Context
Validation Accuracy	83%	Overall classification success across all states.
Macro F1-Score	0.74	Robustness indicator across imbalanced classes.
Model Size	0.60 MB	Efficient memory footprint for embedded systems.

Execution Guide

Dependency Installation:

pip install -r requirements.txt

Pipeline Deployment:

python train.py

Repository Structure

• datasets/: Feature mapping and dataset normalization logic.
• models/: CNN architecture and training optimization loops.
• transforms/: Signal-to-spectral engineering utilities.
• train.py: Main execution script for the end-to-end Data Science pipeline.