Reducing the Complexity of Deep Learning Models for EEG Analysis on Wearable Devices

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Internet of Things (IoT) & Connected Devices · Depth: Advanced, extended

Summary

Deploying deep neural networks (DNNs) for electroencephalogram (EEG) analysis on resource-constrained wearable devices, particularly for epileptic seizure detection, is investigated. The study explores parameter quantization and electrode reduction to balance accuracy and computational complexity. The ResNet-LSTM model, achieving 90.38% AUROC on the TUSZ dataset, was identified as optimal. Applying 8-bit quantization to LSTM and 16-bit to CNN reduced model size by 59% (from 15.75 MB to 6.4 MB) and execution time by 3.4x with minimal accuracy loss. Further, LQ-Net binarization (4-level) reduced the model to 2.3 MB and accelerated execution by 4.1x, but with a 5% accuracy drop (to 86.12% AUROC). Electrode reduction from 20 to 8 channels (10 electrodes) maintained accuracy while reducing input data by 60%. Combining all methods resulted in a 61% network size reduction, 80% input size reduction, 70% acceleration, and less than 1% AUROC decrease.

Key takeaway

For AI Hardware Engineers designing wearable EEG devices, you should prioritize a ResNet-LSTM architecture and implement mixed-precision quantization (e.g., 8-bit LSTM, 16-bit CNN) to achieve significant reductions in model size and execution time. Additionally, consider reducing EEG electrode count to 8 channels to simplify device design and improve user comfort, all while maintaining high accuracy for critical applications like epileptic seizure detection. This approach balances performance with tight resource constraints.

Key insights

DNNs for EEG analysis on wearables can be optimized for resource constraints through quantization and electrode reduction with minimal accuracy loss.

Principles

• Quantization significantly reduces model size and speeds execution.
• CNN layers are more sensitive to bit-width reduction than LSTMs.
• Electrode redundancy allows for input channel reduction.

Method

The study used the TUSZ dataset, comparing DNNs, selecting ResNet-LSTM, then applying 8-bit LSTM/16-bit CNN quantization and reducing 20 EEG channels to 8 (10 electrodes) for epileptic seizure detection.

In practice

• Use ResNet-LSTM for balanced EEG analysis accuracy and speed.
• Apply 8-bit LSTM and 16-bit CNN quantization for efficiency.
• Reduce EEG electrodes to 8 channels for device simplification.

Topics

• Deep Learning Optimization
• EEG Analysis
• Wearable Devices
• Model Quantization
• Epileptic Seizure Detection
• ResNet-LSTM

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Hardware Engineer

Related on AIssential

• Reducing the Complexity of Deep Learning Models for EEG Analysis on Wearable Devices — Parameter quantization and electrode reduction can significantly reduce DNN complexity for wearable EEG analysis with minimal accuracy loss.
• A spectral audit framework reveals task-dependent aperiodic reliance across EEG and ECG deep learning — Deep learning models on physiological data often rely on aperiodic signals, impacting task performance.
• From 32 Bits to 1.58: The Illustrated Guide to LLM Quantization — LLM quantization significantly reduces memory and accelerates inference while maintaining intelligence across various bit precisions.
• Quantization Effects of Artificial Neural Networks for Embedded Edge-Computing Applications — LUT-based 2-bit BNNs trained with GAs enable sub-50ns real-time waveform classification on FPGAs.
• ECG Signal Classification Using Image-Based CNN Deep Learning — ECG signals can be effectively classified for cardiac abnormalities by transforming them into "2-D scalogram images" for CNN deep learning.
• KAN-MLP-Mixer: A comprehensive investigation of the usage of Kolmogorov-Arnold Networks (KANs) for improving IMU-based Human Activity Recognition — Strategically combining KANs for data embedding and classification with MLPs for feature mixing significantly enhances noisy IMU-based HAR.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.