📄 medicine

Stable and Interpretable Cognitive Status Screening Using Longitudinal Wearable Data

This study presents an interpretable hierarchical framework that utilizes longitudinal wearable data to distinguish between cognitively normal, mild cognitive impairment, and dementia states with promising accuracy, while identifying specific physiological patterns associated with each condition to support risk screening and hypothesis generation.

Original authors: Haeun Kim, Taewan Cho, Andrew Jaeyong Choi, Seon-Heui Lee

Published 2026-10-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haeun Kim, Taewan Cho, Andrew Jaeyong Choi, Seon-Heui Lee

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The human mind changes slowly over time, often in ways that are too subtle to catch during a brief visit to a doctor's office. For decades, the standard way to check for cognitive decline has been to ask a person a series of questions about memory and orientation, a method that captures a single moment in a person's life. However, the brain does not stop working when the test is over; it continues to shape how a person moves, sleeps, and reacts to their environment every day. Recent science has begun to explore whether the quiet, continuous data collected by wearable devices—like smartwatches that track steps, heart rate, and sleep patterns—could reveal these gradual shifts long before a formal diagnosis is possible. The challenge lies in distinguishing genuine changes in a person's health from the noise of daily life, such as a bad night's sleep due to stress or a temporary illness, and doing so without relying on the very cognitive tests the system is meant to screen for.

A team of researchers at Gachon University in South Korea has developed a new way to use this everyday data to screen for three distinct stages of cognitive health: normal function, a mild intermediate stage known as mild cognitive impairment, and dementia. Instead of trying to guess a person's status from a single snapshot, the researchers built a system that looks at seven-day windows of data, capturing the rhythm, variety, and stability of a person's life over time. They analyzed data from 174 adults aged 55 and older, using only the raw signals from wearable sensors like activity counts, sleep stages, and heart rate, while deliberately excluding any scores from standard memory tests. The goal was to see if the patterns of daily living alone could reliably sort people into these three groups.

The researchers found that a single, simple model struggled to tell the three groups apart, often missing the middle group entirely. To solve this, they designed a two-step process that mimics a careful triage system. First, the system scans the data to see if a person shows signs of dementia. If the evidence is strong, the person is flagged for further attention. If not, the system moves to a second step to decide whether the person is cognitively normal or in that tricky middle stage of mild impairment. This approach proved much more effective. When tested on a separate group of 38 participants who had not been seen during the training phase, the system correctly identified 83 percent of the people with dementia and 70 percent of those with mild impairment. It correctly classified the overall group with an accuracy of 71 percent, a significant improvement over trying to do everything in one step.

What makes this work particularly notable is not just the numbers, but what the system actually "saw" to make its decisions. The researchers used a method to peel back the layers of the computer model and understand which specific behaviors mattered most. For the dementia group, the system pointed to four clear patterns: a loss of complexity in daily activities, a more rigid organization of movement, increased time spent resting, and a specific change in the structure of light sleep. For the group with mild impairment, the signals were more about instability. The system noticed that these individuals had trouble sustaining activity for long periods, experienced erratic shifts in their sleep-wake timing, and showed fluctuating patterns in their nighttime heart rate and body temperature.

The study is careful to frame these findings as a tool for screening and hypothesis generation rather than a definitive diagnostic test. The researchers emphasize that the system is designed to flag people who might need a deeper look from a medical professional, not to replace the doctor. The patterns identified, such as the specific changes in sleep structure or activity fragmentation, are described as preliminary clues that align with what is already known about how the brain and body interact, but they are not yet confirmed as standalone medical markers. The team acknowledges that their dataset was small, particularly for the dementia group, and that wearable devices vary in how accurately they measure things like sleep stages.

Ultimately, this work demonstrates that the story of cognitive health is written in the long, continuous narrative of daily life, not just in the short chapters of a clinical exam. By listening to the subtle rhythms of movement and rest over weeks rather than minutes, it is possible to spot the early signs of decline that a single visit might miss. The researchers suggest that this kind of longitudinal monitoring could become a valuable part of community health, helping to identify those who need support before a crisis occurs. However, they conclude that before such a system can be used in hospitals or clinics, it must be tested on much larger groups of people across different locations and with different types of devices to ensure the patterns hold true for everyone.

Technical Summary: Stable and Interpretable Cognitive Status Screening Using Longitudinal Wearable Data

Problem Statement
Wearable devices continuously capture activity, sleep, and nocturnal physiology, offering a longitudinal view of daily life that could complement traditional cognitive screening. However, their utility for distinguishing three distinct cognitive states—Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and Dementia (DEM)—remains uncertain. Existing studies often focus on binary CN–MCI discrimination or cognitive-score prediction, failing to address the overlapping boundaries and class imbalance inherent in a three-class task. Furthermore, standard machine learning approaches face significant risks of overfitting due to high-dimensional, correlated features and subject-level data leakage when repeated observations from the same participant cross training and validation partitions. There is a need for a framework that distinguishes genuine longitudinal behavioral changes from measurement artifacts while providing interpretable insights without relying on the Mini-Mental State Examination (MMSE) or proprietary device composite scores.

Methodology
The authors propose a hierarchical, participant-level inference framework designed to prioritize risk screening. The system processes 7-day windows of wearable data (activity, sleep, heart rate, temperature) derived from 174 adults (55+ years).

Feature Engineering: Daily records are organized into overlapping (3-day stride) and non-overlapping (7-day stride) 7-day windows. Features encode level, variability, trend, temporal transitions, bout structure, and signal complexity (e.g., Shannon entropy). Longitudinal slopes and changes are calculated across the observation period.
Two-Stage Architecture:
- Stage 1 (DEM vs. Non-DEM): A Random Forest classifier distinguishes Dementia from non-Dementia (CN + MCI). To prevent dilution of strong signals by lower-probability windows, the participant-level score is the mean of the top-3 window probabilities. Participants with a score $\ge$ 0.21 are routed to DEM.
- Stage 2 (CN vs. MCI): For participants routed as non-DEM, a second Random Forest classifier distinguishes CN from MCI. This stage uses participant-median aggregation of longitudinal slope-and-change features to reduce sensitivity to transient windows. Participants with a probability $\ge$ 0.35 are classified as MCI; otherwise, CN.
Validation and Robustness: The study employs strict participant-level cross-validation (no participant data leakage between folds). Thresholds were fixed on development data after sensitivity analyses across repeated 3- to 6-fold schemes. Evaluation was performed on an internal holdout set of 38 participants.
Interpretability: Post-hoc explanation uses TreeSHAP. To move from model features to clinical insights, the authors applied correlation grouping ( $|\rho| \ge 0.85$ ) and filtered features based on consistency, directionality, and physiological interpretability, explicitly separating SHAP rank from clinical biomarker claims.

Key Contributions

Longitudinal Representations: Construction of a feature set describing activity complexity, transitions, bout structure, and nocturnal physiology without using MMSE or device composite scores.
Hierarchical Inference: Formulation of cognitive screening as a two-stage process (DEM detection followed by CN/MCI separation) with stage-specific aggregation strategies (top-3 mean vs. median).
Rigorous Evaluation Protocol: Implementation of participant-level validation, sensitivity-oriented threshold selection, and explicit error routing analysis to track how misclassifications propagate through the hierarchy.
Interpretable Domains: Derivation of preliminary behavioral patterns and exploratory domains by filtering SHAP-ranked features through clinical plausibility and group difference consistency.

Results
On the internal holdout set (38 participants), the framework achieved:

Overall Performance: Accuracy of 0.711, Balanced Accuracy of 0.738, and Macro F1 of 0.692.
Class Recall: 0.833 for DEM, 0.700 for MCI, and 0.682 for CN.
Comparison: A direct one-stage three-class baseline failed significantly (MCI recall of 0.000), demonstrating the value of the hierarchical approach in amplifying minority-class sensitivity.
Robustness: Stage 1 sensitivity remained stable at 0.833 across repeated cross-validation schemes.

Interpretability Findings

DEM Patterns: Four preliminary patterns emerged: (1) Lower activity complexity/diversity, (2) Specific device-coded state organization (more class 1, fewer class 3 bouts), (3) Higher rest and inactivity, and (4) Higher device-recorded light-sleep measures.
MCI Domains: Three exploratory domains were identified: (1) Activity persistence and fragmentation (shorter sustained activity), (2) Sleep-wake timing instability (extreme shifts in sleep midpoint), and (3) Longitudinal changes in nocturnal autonomic and temperature dynamics.

Significance and Claims
The authors position this work as a step toward risk screening and hypothesis generation, explicitly stating that the findings do not constitute validated biomarkers or a diagnostic tool. The framework supports the identification of individuals whose longitudinal changes in activity, sleep, or physiology warrant formal clinical assessment. The study emphasizes that the engineering value lies not in the Random Forest classifier alone, but in the combination of longitudinal feature construction, participant-level validation, and stage-specific aggregation. The authors conclude that while the results are promising for nursing-assisted monitoring and triage, external multicenter validation is required before clinical deployment, particularly given the small sample size of the DEM group and the device-dependent nature of consumer wearable signals.

Technical Summary: Stable and Interpretable Cognitive Status Screening Using Longitudinal Wearable Data

More like this