Two-stream Graph Convolutional Networks with Task-specific Loss for Dual-task Gait Analysis.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Dual-task gait analysis is a promising method for assessing cognitive decline in the elderly.
• Current methods face challenges in fully utilizing dual-task cost representations and optimizing their extraction for diagnostic purposes.

Purpose of the Study:

• To propose a novel deep learning framework for cognitive impairment detection in elderly individuals using gait data.
• To address limitations in dual-task cost representation extraction and utilization for improved diagnostic accuracy.

Main Methods:

• Implementation of a spatio-temporal graph convolutional network (ST-GCN) with distinct single-task and dual-task pathways.
• Introduction of a task-specific loss function to ensure distinguishable representations.
• Calculation of dual-task cost representations as the difference between dual-task and single-task representations to enhance robustness.

Main Results:

• The proposed framework demonstrated high performance in cognitive impairment detection.
• Achieved a sensitivity of 0.969 and a specificity of 0.940, outperforming existing methods.
• The dual-task cost representations proved resilient to individual differences, contributing to the framework's robustness.

Conclusions:

• The developed deep learning framework offers a robust and accurate method for detecting cognitive impairment through dual-task gait analysis.
• The novel approach effectively leverages dual-task cost representations, paving the way for improved diagnostic tools in geriatric care.