Dementia affects approximately 57 million people worldwide, yet nearly three-quarters of cases remain undiagnosed. Current screening methods—such as PET scans, MRIs, and fluid biomarkers—are expensive, invasive, and difficult to scale for general population screening. While sleep disturbances are well-known symptoms of neurodegeneration, clinical-grade monitoring (Polysomnography) is cumbersome and often alters a patient's natural sleep patterns due to the unfamiliar environment. There is a critical need for "invisible" technology that can monitor sleep longitudinally in a person's own home to catch the subtle physiological shifts that precede a formal dementia diagnosis.
Researchers have developed a machine learning pipeline that uses non-invasive, under-the-mattress sensors to identify dementia risk by calculating a "Sleep Age Index" (SAI). This study reveals that sleep patterns in dementia do not follow typical aging trajectories; instead, they exhibit distinct irregularities in heart rate, breathing, and sleep-wake cycles that can serve as early digital biomarkers for cognitive decline.
Methods
The proposed framework utilizes a two-stage machine learning approach:
Age Estimation: Using nocturnal activity and physiological data (heart rate, respiratory rate, sleep stages), an XGBoost regression model was trained to predict a participant's "sleep age."
Dementia Risk Classification: A Sleep Age Index (SAI) was calculated as the difference between chronological age and predicted sleep age. A Logistic Regression model then used these SAI scores to classify individuals into dementia vs. control groups.
To enhance clinical utility, the researchers implemented a "traffic light" stratification system (Red, Amber, Green) based on model confidence. Interpretability was provided via SHAP (SHapley Additive exPlanations), identifying specific sleep metrics—such as heart rate variability or time spent out of bed—that drove individual predictions.
Data Cohorts
The study leveraged three distinct cohorts to train and validate the models:
Population Control: Data from 1,567 individuals (15,229 samples) using Withings sleep mats in the general population to establish "typical" aging patterns.
Dementia Cohort: 105 participants (3,140 samples) from the Minder study, consisting of people living with dementia in their own homes.
Pilot High-Risk Cohort: 50 individuals from the Resilient study with multi-morbidity, used to validate the model's real-world clinical applicability against expert clinician judgment.
Sample distribution by age and sex in the Population Control group.
The study yielded several critical findings regarding the relationship between sleep and cognitive decline:
Dementia is not "Accelerated Aging": Surprisingly, the model often predicted individuals with dementia to be younger than their chronological age (negative SAI). Dementia-related sleep fragmentation—irregular bedtimes and frequent awakenings—mimics the erratic sleep schedules of younger adults rather than the stable patterns of healthy older adults.
Physiological Signatures: Specific markers of decline included elevated mean breathing and heart rates, and lower variation in deep sleep duration.
Performance Metrics: The model predicted chronological age with a Mean Absolute Error (MAE) of 5.52 years. For classification, it achieved a sensitivity of 75.7% and a specificity of 74.7% in distinguishing dementia cases from controls.
Clinical Validation: In the pilot study, the model’s risk scores showed strong agreement with expert clinical judgment (mean difference of 0.98 units).
Discussion and Limitations
While promising, the authors acknowledge several limitations. The current model does not account for medications or comorbidities like sleep apnea. Environmental factors also play a role; researchers noted differences in SAI between winter and summer months, suggesting seasonality should be integrated into future models. Furthermore, a significant data imbalance exists between age groups (e.g., 50-59 vs. 80+), which may affect generalizability across all demographics.
Ultimately, this work demonstrates that under-the-mattress sensors can act as passive health monitors, flagging the earliest signs of cognitive decline years before clinical symptoms become apparent.
Summary updated
Abstract
Abstract BACKGROUND Sleep disturbances and altered circadian rhythms are well-documented in both physiological and biological studies of dementia. The exact causal relationship remains unclear. Several other long-term health conditions may also influence sleep patterns. Examining sleep patterns in relation to chronological ageing and the variations and progression of dementia, considering factors such as age, sex, and disease stage, can offer new insights for developing early screening and risk identification measures. Remote sleep monitoring enables routine assessment of cognitive decline symptoms in high-risk groups, aiding early risk identification. Integrating predictive models with routine sleep data holds promise for shaping more proactive approaches to assessing and caring for older adults and people living with dementia (PLWD). METHODS We developed a machine learning pipeline to estimate Sleep Age Index (SAI) from longitudinal remote sleep monitoring data collected using under-the-mattress sleep sensors and use it to identify dementia risk. The study utilised a dataset of nocturnal activity and physiology data (n=1,672; 18,369 person-samples) collected from individuals in the general population and a cohort of PLWD. Dementia risk scores were stratified into high, medium and low-risk categories to support clinical monitoring and decision-making. RESULTS Our study indicates that sleep patterns in dementia do not follow typical ageing processes. For individuals with dementia, our pre-trained machine learning model predicted age in a negative direction. Further investigation revealed distinct patterns associated with these predictions, including irregular times to bed and rise, lower variation in deep sleep duration, and elevated breathing and heart rates. While age predictions were largely similar for female and male participants, the patterns observed in the female group were more consistent. Chronological age was predicted from sleep data with a mean absolute error of 5.52 (95% CI: 5.37 - 5.67) on held-out test data. A sensitivity of 75.7% (95% CI: 71.4% - 79.9%) and specificity of 74.7% (95% CI: 69.2% - 80.0%) was achieved post-stratification on unseen data in dementia versus control. To evaluate the real-world clinical applicability of the model, we conducted a pilot study in a population with a higher risk of cognitive decline (n = 50). Pilot data analysis indicated a slight positive bias between model predictions and the clinical experts’ judgement (mean difference 0.98 units, limits of agreement from −0.83 to 2.78 units, reported to 3 s.f.). This analysis also revealed that the traffic light system, designed to indicate the risk of cognitive decline, can serve as a complementary source of information to enhance decision support. This is especially evaluated for its applicability in improving clinical decision-making for high-risk groups, where there is often insufficient data available regarding an individual’s cognitive status. CONCLUSIONS The study offers new insights into sleep and dementia, highlighting how age and sex differences manifest differently in typical ageing compared to dementia. We demonstrate the utility of leveraging sleep monitoring data and predictive analysis in identifying individuals who may benefit from further clinical evaluation and early disease-modifying interventions.
BibTeX
@misc{fletcherlloyd2025real,
author = {Nan Fletcher‐Lloyd and Nathalia Céspedes and Alexander Capstick and Antigone Fogel and Marirena Bafaloukou and Mohammad Heydari and Alyssa Cairns and Chloe Walsh and Jessica True and Behnam Shariati and Ramin Nilforooshan and Payam Barnaghi and Payam Barnaghi},
title = {{Real-world deployment of remote sleep monitoring technologies reveals distinct patterns associated with cognitive decline}},
year = {2025},
doi = {10.1101/2025.08.29.25334735},
howpublished = {medRxiv},
}