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

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

Problem Statement
Wearable healthcare devices represent a rapidly growing sector of the Internet of Things (IoT), offering continuous monitoring of vital signs like ECG and EEG. While Deep Neural Networks (DNNs) have become the premier approach for analyzing complex biological signals such as EEG for tasks like epileptic seizure detection, their deployment on wearable devices is severely hindered by resource constraints. Wearable systems typically rely on miniaturized processors and small batteries, lacking the computational power, energy capacity, and memory bandwidth required to run state-of-the-art DNNs in real-time. Although offloading data to the cloud is a common workaround, it introduces issues regarding power consumption, data privacy, security, and the need for constant connectivity. Consequently, there is a critical need to adapt DNNs for "edge processing" within the wearable device itself, balancing high accuracy with extreme computational efficiency.

Methodology
This study investigates the feasibility of deploying state-of-the-art DNN models for EEG analysis on resource-constrained wearable devices. The research focuses on epileptic seizure detection using the TUSZ dataset, which contains 20 bipolar EEG channels resampled at 200 Hz and segmented into 4-second windows. The methodology involves a three-pronged approach to complexity reduction:

1. Model Selection: The authors evaluated several architectures, including ResNet18, ChronoNet, 2D-CNN-LSTM, ResNet-GRU, ResNet-BiLSTM, and ResNet-LSTM. The ResNet-LSTM architecture was selected as the baseline due to its superior performance in balancing accuracy and execution speed. This model utilizes a front-end ResNet-based CNN for spatial feature extraction and a back-end LSTM for temporal dependency learning.
2. Parameter Quantization: The study explored reducing the bit-width of model parameters and activations to lower memory footprint and accelerate inference.
   • Post-Training Quantization (PTQ): The authors compared Static Quantization (PTSQ) and Dynamic Quantization (PTDQ). They found that uniform 8-bit quantization across the entire network caused significant accuracy degradation. Through granular experimentation, they determined that the CNN component is highly sensitive to bit-width reduction, while the LSTM is more robust. An optimal configuration was identified using 13-bit precision for the CNN and 8-bit precision for the LSTM.
   • Binarization (LQ-Net): The study further applied the state-of-the-art LQ-Net method, which utilizes non-uniform multi-level binarization to reduce parameters to 4 bits.
3. Electrode Reduction: Recognizing that clinical EEG setups often use 20+ electrodes, which is impractical for wearables, the authors investigated the impact of reducing the number of input channels. They tested various combinations of electrodes, prioritizing those most sensitive to epileptic signals, to determine the minimum set required to maintain diagnostic accuracy.

Key Results
The application of these complexity reduction techniques yielded significant improvements in efficiency with minimal impact on performance:

• Optimal Configuration: The ResNet-LSTM model achieved the highest baseline accuracy (90.40% AUROC) among the tested architectures.
• Quantization Impact:
  • Using a hybrid precision approach (13-bit CNN, 8-bit LSTM), the model size was reduced from 15.75 MB to 6.4 MB (a 59% reduction). Execution time improved by 3.4× compared to the 32-bit baseline.
  • Applying LQ-Net binarization (4-bit) further reduced the model size to 2.3 MB and increased execution speed by 4.1×, though this came with a 5% drop in AUROC (from ~90.33% to 86.12%).
• Electrode Reduction: Reducing the input from 20 channels (19 electrodes) to 8 channels (10 electrodes) resulted in an AUROC of 89.23%, a negligible decrease from the full setup. This configuration maintains high accuracy while significantly lowering data volume and hardware complexity.
• Combined Optimization: By combining the 8-channel input, 16-bit input data, 13-bit CNN weights, and 8-bit LSTM/FC weights, the final optimized model achieved:
  • A 61% reduction in network size.
  • An 80% reduction in network input size.
  • A 70% acceleration in network inference time.
  • A decrease in AUROC of less than 1% and a 2.1% decrease in False Positive Rate (FPR).

Significance and Claims
The paper claims that these findings demonstrate the explicit trade-offs between accuracy and complexity reduction when adapting DNN-based online EEG analysis for wearable devices. The authors assert that their approach proves the feasibility of deploying sophisticated deep learning models on edge devices without relying on cloud offloading. By identifying an optimal network configuration and applying judicious quantization and electrode reduction, the study shows that it is possible to achieve a 3× reduction in model size and a 5× reduction in execution time with less than 1% accuracy loss.

The significance of this work lies in its practical contribution to the Internet of Medical Things (IoMT). It provides a pathway for creating lightweight, portable, and energy-efficient wearable devices capable of continuous, real-time neurological monitoring. This advancement is crucial for improving the autonomy, privacy, and responsiveness of automated healthcare services, particularly for high-risk patients requiring constant seizure surveillance outside of clinical environments. The paper explicitly states it does not propose new DNN architectures but rather evaluates existing methods under specific complexity-reduction constraints to guide future hardware and software integration in wearable health technology.