GFASNet: Gait feature attention-driven deep sequential network for dementia-related gait pattern analysis.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Deep learning models show potential for dementia prediction using human activity data like gait.
• Limited interpretability and clinical relevance of current models hinder their application in cognitive health research.

Purpose of the Study:

• Introduce GFASNet (Gait Feature Attention-driven Deep Sequential Network) for identifying dementia-related gait alterations.
• Enhance model transparency and quantify gait parameter contributions using attention mechanisms.
• Explore potential digital biomarkers for early dementia detection.

Main Methods:

• Collected spatiotemporal gait data from 232 participants using a pressure-sensor walkway.
• Trained and evaluated four GFASNet variants (LSTM, BiLSTM, GRU, BiGRU) on gait sequences (eight strides).
• Utilized feature-level attention mechanisms within deep sequential architectures.

Main Results:

• All GFASNet models outperformed non-attention baselines in dementia classification tasks.
• Attention weight analysis revealed consistent focus on specific gait features for dementia case identification.
• Demonstrated GFASNet's ability to provide interpretable gait analysis.

Conclusions:

• GFASNet enhances dementia identification accuracy through interpretable gait analysis.
• Identified gait features via attention mechanisms show promise as digital biomarkers for cognitive health.
• GFASNet facilitates clinically relevant gait analysis for dementia research.