Authors:

Chaitra Hegde、Yashar Kiarashi、Allan I Levey、Amy D Rodriguez、Hyeokhyen Kwon、Gari D Clifford

Paper:

https://arxiv.org/abs/2408.10442

Feasibility of Assessing Cognitive Impairment via Distributed Camera Network and Privacy-Preserving Edge Computing

Introduction

Mild cognitive impairment (MCI) is a condition characterized by a decline in cognitive functions that surpasses typical age and education-related expectations. It is often a precursor to dementia, with over half of those diagnosed progressing within five years. Early diagnosis of MCI is crucial as it allows patients and caregivers to develop coping strategies while the individual still retains significant cognitive function. However, diagnosis is often delayed due to limited access to expert assessment or the dismissal of early symptoms as normal aging.

MCI has been linked to reduced social interactions and increased aimless movements. This study aims to automate the capture of these behaviors using a privacy-preserving distributed camera network to enhance longitudinal monitoring. The goal is to distinguish between higher and lower cognitive functioning MCI groups based on movement and social interaction data.

Related Work

Previous studies have explored the use of machine learning to analyze gait and spatial navigation for prescreening MCI. These studies typically involved controlled environments or single-resident households equipped with motion sensors. However, these approaches have limitations in generalizing to real-world scenarios and understanding spatial usage and navigation behaviors. Social interactions in the MCI population remain relatively understudied, often relying on self-reports or caregiver accounts, which may introduce bias.

This study introduces a passive sensing pipeline designed to capture detailed movement and social interaction cues during group activities in an MCI population. The data were collected over fourteen months in a 1700m² space, aiming to distinguish between groups of individuals with varying cognitive function levels.

Research Methodology

Data Collection

Study Site

The data were collected at the Charlie and Harriet Shaffer Cognitive Empowerment Program (CEP), a therapeutic program for individuals with MCI. Participants engaged in weekly activities such as cooking, exercise, and cognitive training within a designated indoor space. The analysis focused on data from three daily breaks when participants moved and socially interacted freely.

Cognitive Assessment

Participants underwent the Montreal Cognitive Assessment (MoCA), a 30-point cognitive screening test. Based on their MoCA scores, participants were divided into cohorts representing high or low levels of cognitive functioning. The study examined 66 subjects divided into six cohorts, categorized as high functioning if the mean MoCA score exceeded 21 and low functioning if it was 21 or below.

Behavior Sensing Framework

Distributed Camera Network with On-device Pose Estimation

A distributed camera system consisting of 39 edge computing and camera devices was deployed throughout the therapeutic space. These devices used a multiperson 2-D pose estimation model to detect keypoints across the body in real-time. The derived poses and bounding box images were stored for downstream analysis, ensuring that personally identifiable data were not stored or transmitted.

Multi-view Multi-person Localization, Body Orientation Estimation, Tracking, and Group Localization

The keypoints and bounding boxes were used to perform multi-view multi-person localization, body orientation estimation, and tracking analysis. Group formations within the indoor environment were detected using a clustering method.

Behavior Features for Group Activities in MCI

Movement and social interaction features were extracted to highlight behavioral differences between high and low cognitive functioning cohorts. These features included:

• Linear Path Lengths: Lengths of uninterrupted walking paths.
• Speed of Walking: Walking speed for each detected individual.
• Direction Change: Frequency of changes in walking direction.
• Entropy of Walking Velocity: Measure of complexity in walking speed.
• Entropy of Orientation Change: Entropy from orientation changes.
• Levy Distribution Parameters for Linear Path Lengths: Parameters of the Levy distribution fitted to linear path lengths.
• Overall Group Formations: Number of detected groups as a proxy for social interactions.
• Region-specific Group Formation: Number of groups formed in distinct regions of the facility.

Cohort-level Classification of Cognitive Impairment

The derived features were used for binary classification of high and low levels of cognitive function using models such as Support Vector Machines (SVM), XGBoost (XGB), logistic regression (LR), and Lasso binary classification algorithms.

Experimental Design

Data Collection and Labeling

315 break sessions were recorded over fourteen months across six cohorts. Each break session served as a single sample for deriving behavior features. Based on the mean MoCA score of each cohort, they were labeled as either high or low cognitive functioning.

Model Training and Evaluation

The behavior features were rescaled and fed to classification models. Model performance was compared using all features, only movement features, and only social interaction features, evaluated by precision, recall, F1 score, and accuracy. The Wilcoxon Rank Sum Test assessed statistical differences in raw feature distributions between high and low-functioning cohorts.

Results and Analysis

Classification Results

XGBoost achieved the highest F1 score (0.68) using all features, slightly outperforming SVM (0.67). The SVM demonstrated the highest performance and stability across all feature combinations. Logistic regression and lasso binary classifier both reached 0.65 F1 score with all features.

Feature Analysis

The Wilcoxon rank sum test indicated statistically significant differences in the distributions of most features between high and low-functioning MCI cohorts. Walking speed, linear path lengths, velocity and orientation change entropies, and group formations were among the most discriminative features.

Feature Importance

Feature importance analysis for the SVM models highlighted the most important features for classification. When using all features, the most important features included direction change, velocity entropy, and group formations in specific regions.

Overall Conclusion

This study demonstrated that a privacy-preserving distributed camera system can effectively quantify longitudinal group activity and movements during break times for cohorts with MCI. Movement and social interaction-based features showed discriminative ability in distinguishing between high and low cognitive functioning cohorts. This approach moves beyond controlled tests and questionnaires to assess cognitive abilities in real-world scenarios, potentially identifying early signs of dementia through longitudinal passive monitoring.

Future work will focus on distinguishing subjects with MCI from individuals with normal cognition and applying temporal deep learning methods to model longitudinal changes in behavior related to cognitive impairment. This study lays the foundation for further exploration to validate the robustness of passive monitoring systems for MCI in real-world settings.