PABformer: Multi-Channel Transformer-Based Physical Activity Representation Learning from Wearable Accelerometry for Prediction of Parkinson Disease
Year: 2025
Type: preprint
Abstract: Parkinson disease (PD) is a prevalent neurodegenerative disorder whose early diagnosis and prediction remain challenging due to the delayed onset of overt motor symptoms and the complexity of long-term behavioral changes. Wearable accelerometers provide a scalable and non-invasive approach for monitoring physical activity (PA) in free-living environments and have emerged as a promising tool for PD prediction. However, existing accelerometer-based methods often rely on coarse summary statistics or models with limited ability to capture long-range temporal dependencies in multi-day accelerometer recordings. In this study, we propose PABformer, a multi-channel Transformer framework designed for long-term PA behavior representation learning from accelerometer data. PABformer introduces a channel-separation strategy to disentangle heterogeneous activity streams and leverages self-supervised pretraining to learn generalized behavioral representations. Using week-long accelerometer recordings from 96,463 participants in the UK Biobank, we pretrained PABformer and subsequently fine-tuned it for demographic attribute inference, PD diagnosis, and incident PD prediction. We compared PABformer against traditional statistical, machine learning approaches, recurrent deep learning models, and a standard Transformer without channel separation. Experimental results demonstrated that PABformer achieved superior performance in both PD classification and survival prediction tasks, as measured by precision-recall area under the curve (PRAUC) and concordance index (C-index), respectively. In addition, Grad-CAM analysis improved interpretability by identifying salient channel contributions associated with PD-related behavioral patterns. These findings demonstrate the potential of foundation models for wearable accelerometer data to advance early PD detection, disease risk prediction, and future precision health applications.
Source: medRxiv
Authors: Kai He, Liuyi Fan, Hanyu Qian, Chunyu Li, Meng Wang
Institution: University of Michigan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7124429059