Home // HEALTHINFO 2024, The Ninth International Conference on Informatics and Assistive Technologies for Health-Care, Medical Support and Wellbeing // View article

Social Engagement Embeddings of Parkinson's Disease through Autoencoders

Authors:
Ting Xiao
Poulomi Guha
Ali Shah Tajuddin

Keywords: Autoencoder, Parkinson's Disease, Hidden Representation, Mobile Application.

Abstract:
Various survey tools are available to measure social engagement, but they often suffer from infrequent measurement and recall bias. To address this, we developed a mobile application that estimates turn-taking in conversations and generates engagement features. These features were used to create an autoencoder-based hidden representation of individuals, which distinguishes between Parkinson's Disease and control subjects. The study aims to create reduced representations to robustly compare speaker-test outcomes with limited samples. An autoencoder was employed to reduce the number of features related to social engagement. This tailored assessment tool was applied to extract 42 speaker assessment scores, which were distilled into two-dimensional embeddings using a 9-layer autoencoder. We compared the proposed hidden representation with Principal Component Analysis, assessing metrics such as conversation percentage, turn-taking, and total pauses. These embeddings enabled a cross-validated reconstruction of all 42 features, accounting for 58% of the variance and were validated using multiple classification methods, including K-Nearest Neighbors (KNN), Support Vector Machine, Random Forest, and XGBoost. The KNN model, using the embeddings features, achieved a 90% macro precision score. Our results suggest that autoencoder representations provide a concise and effective tool for the holistic assessment of speaker behavior in limited data scenarios.

Pages: 43 to 47

Copyright: Copyright (c) IARIA, 2024

Publication date: November 3, 2024

Published in: conference

ISSN: 2519-8491

ISBN: 978-1-68558-204-3

Location: Nice, France

Dates: from November 3, 2024 to November 7, 2024