AI-Enhanced Behavioral Biomarkers for Early Neurodegenerative Disorder Prediction

Abstract

Neurodegenerative disorders such as Alzheimer's disease(AD) have a silent period of years before clinical manifestation and there is a great need for early, non-invasive and scalable predictive tools. However, current screening methods involve a lot of clinical visitations, cognitive testing or biomarker assays that are expensive, episodic and often are insensitive to detect early subtle changes. This study tackles the issue of identifying prodromal risk for neurodegenerative disease based on naturalistic behavioral clues that could be passively and continuously obtained by everyday digital devices. The aim is to develop and test a new AI-enhanced framework based on identification of behavioral biomarkers of potential for future conversion using multimodal sensor data. We propose TGAT-X, which is a Temporal Graph-Adaptive Transformer with multimodal graph fusion, individual baseline deviation modelling and self-supervised longitudinal pretraining for extracting robust behavioural signatures. Experiments done on a longitudinal dataset of 1800 participants indicate that TGAT-X performs significantly better than five state-of-the-art baselines with an AUROC score of 0.91, AUPR score of 0.56 and C-index score of 0.86 for predicting conversion risk in the next six months. These results show that there are rich early warning signals of neurodegenerative decline in the dynamics of behavior and that advanced multimodal artificial intelligence models can unlock their predictive value. Overall, this study proposes the use of a scalable and privacy-preserving method with promising implications for the early detection of the population and timely clinical intervention.

Ähnliche Arbeiten