PPGCN: Phase-Aligned Periodic Graph Convolutional Network for Dual-Task-Based Cognitive Impairment Detection

Abstract

Early detection methods for cognitive impairment are crucial for its effective treatment. Dualtask-based pipelines that rely on skeleton sequences can detect cognitive impairment reliably. Although such pipelines achieve state-of-the-art results by analyzing skeleton sequences of periodic stepping motion, we propose that their performance can be improved by decomposing the skeleton sequence into representative phase-aligned periods and focusing on them instead of the entire sequence. We present the phase-aligned periodic graph convolutional network, which is capable of processing phase-aligned periodic skeleton sequences. We trained it with a cross-modality feature fusion loss using a representative dataset of 392 samples annotated by medical professionals. As part of a dual-task cognitive impairment detection pipeline that relies on two-dimensional skeleton sequences extracted from RGB images to improve its general usability, our proposed method outperformed existing approaches and achieved a mean sensitivity of 0.9231 and specificity of 0.9398 in a four-fold cross-validation setup.