Back to search

University of Minnesota

Deep Learning-Based Incident Dementia Prognosis from Wearable ECG and Accelerometer Signals

Abstract

dc:description.abstract

Dementia represents a growing global health crisis, yet current screening approaches remain reactive, often detecting disease only after irreversible neurological damage has occurred. Recent evidence supporting the heart-brain axis motivates the use of cardiac and physical activity biomarkers as accessible, non-invasive proxies for early dementia risk. To address the need for proactive dementia prediction, we present a deep learning framework for predicting incident dementia within a 5-year period using single-lead ECG and tri-axial accelerometry data from the Atherosclerosis Risk in Communities (ARIC) cohort, recorded via the Zio® XT wearable patch across 1911 participants (234 dementia, 1677 controls). We construct a multimodal spectral tensor by transforming 7 days of continuous ECG, heart rate variability (HRV), and accelerometry signals into power spectral density (PSD) representations across 30-minute epochs, yielding a (3, 336, 128) input tensor per patient. A lightweight 5-layer CNN trained on this tensor achieves a mean validation AUROC of 0.735 under stratified 5-fold cross-validation, outperforming classical machine learning baselines including Logistic Regression (0.705), Random Forest (0.695), and XGBoost (0.686). To interpret the model's predictions, we employ Cohen's d standardized difference maps and Integrated Gradients (IG), revealing that daytime physical activity and nocturnal cardiac dynamics are key discriminative features. Channel-level ablation confirms that accelerometry is the dominant predictive modality, with its removal causing a 12% relative drop in AUROC. Temporal ablation shows that performance plateaus at 3 days for AUROC but continues to improve in AUPRC through 7 days, underscoring the value of longitudinal monitoring for identifying the minority class. Finally, mechanistic interpretability via linear probing confirms that the model's latent space encodes clinically relevant markers, including resting heart rate, rMSSD, M10, and relative amplitude, without explicit supervision. Together, these results establish a strong proof-of-concept for wearable-based deep learning dementia prognosis, and motivate future work on considering confounding medications and out-of-distribution generalization.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Al-Sharawi, Ramzi

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Al-Sharawi, Ramzi. (2026). Deep Learning-Based Incident Dementia Prognosis from Wearable ECG and Accelerometer Signals. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/281672.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Al-Sharawi, Ramzi. Deep Learning-Based Incident Dementia Prognosis from Wearable ECG and Accelerometer Signals. 2026. https://hdl.handle.net/11299/281672