Automated Physical Performance Battery as a Digital Marker for Alzheimer’s Disease and Mild Cognitive Impairment

Ehsan AdeliAlzheimer 2024 · DOI

TL;DR

This paper introduces a novel computer vision (CV) approach to automatically analyze physical performance from video recordings of the Short Performance Physical Battery (SPPB) test. By extracting detailed body movement data, the method offers a more objective and accessible way to identify early signs of Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI), potentially overcoming limitations of current diagnostic tools.

What Problem Does This Paper Address?

Current methods for screening Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI) often rely on cognitive assessments, daily living activity evaluations, and brain imaging. However, these approaches can be time-consuming, costly, dependent on individual education and language proficiency, and require long-term monitoring. Emerging research highlights a strong link between non-tremor motor dysfunction, particularly in gait and hand movements, and AD biomarkers, even in the early stages of cognitive decline. This suggests that physical performance, especially gait, could serve as an early and trackable indicator of cognitive impairment. The paper aims to address the need for more objective, efficient, and accessible tools for early AD/MCI detection by leveraging computer vision to analyze physical performance data.

Methods

The core of this paper's methodology lies in the application of advanced computer vision (CV) techniques to analyze video recordings of the 5-minute Short Performance Physical Battery (SPPB) tests. The process involves:

  • Automated Body Marker Extraction: Using CV, the system automatically processes video data to generate body markers, essentially creating a tracked human skeleton over time. This is a key novel aspect, moving beyond manual annotation or specialized sensors.
  • Feature Extraction: From the tracked skeleton, several quantitative features related to mobility are extracted. These include:
  • Gait speed
    • Mean torso inclination angle
    • Double support time (the period when both feet are on the ground during walking)
    • A gait-summary score
  • Clinical Score Derivation: The extracted physical features are then used to derive clinical SPPB test scores. This transforms raw movement data into clinically relevant metrics.

The paper highlights that this CV approach automates the quantification of mobility in a data-driven fashion, offering a more objective and potentially scalable alternative to traditional manual scoring of physical performance tests.

Datasets

The study utilizes data collected at the Stanford AD Research Center. The paper does not specify the exact number of participants or the breakdown of control, MCI, and AD groups within this dataset. It also does not mention the creation of a new benchmark dataset.

Key Insights and Ideas

  • Motor Dysfunction as an Early AD/MCI Indicator: The research reinforces the growing understanding that subtle changes in physical movement, particularly gait, can precede or co-occur with cognitive decline in AD and MCI. This opens up new avenues for early detection.
  • CV for Objective Mobility Assessment: Applying CV to analyze physical performance tests like the SPPB offers a powerful way to move beyond subjective human observation. It allows for precise, quantitative measurement of movement patterns that might be imperceptible to the naked eye.
  • Potential for Scalable Screening: Automating the analysis of physical performance through CV could significantly reduce the cost and time associated with traditional diagnostic methods, making screening more accessible.
  • Bridging Animal and Human Research: The authors suggest that this CV mobility analysis can be extended to animal models of AD, providing a consistent framework for studying disease progression and therapeutic interventions across species.

Key Results

  • Significant Difference in Chair Stand Test: The study observed a statistically significant difference in the "repeated chair stand test score" between healthy controls (HC) and both the MCI and AD groups. This indicates that even this specific physical task can reveal distinctions between cognitive health statuses.
  • Inverse Correlation Between Gait Speed and Cognitive Score: An inverse correlation was found between the Montreal Cognitive Assessment (MoCa) score and gait speed. This means that as gait speed decreases, MoCa scores (indicating cognitive function) also tend to decrease, further linking physical and cognitive health.
  • Visualizations of Movement Data: The paper includes visualizations showing knee angle over time during movement and foot contact events, demonstrating the detailed kinematic data captured by the CV system.

Limitations and Future Work

  • Dataset Size and Diversity: While the study used data from the Stanford AD Research Center, the specific size and demographic diversity of the dataset are not detailed, which could limit the generalizability of the findings.
  • Validation Against Gold Standard: The paper focuses on deriving clinical SPPB scores from CV analysis. Further validation against established diagnostic criteria or more comprehensive AD assessments would strengthen the findings.
  • Generalizability of CV Model: The robustness of the CV model across different video qualities, lighting conditions, and participant variations would need to be explored.
  • Future Directions: The authors explicitly mention future work involving the application of this CV method for detecting behavioral changes in animal AD models and exploring its broader implications for future human AD research. This suggests a path towards developing a comprehensive digital marker for AD.

Summary updated

Abstract

Historically, screening for incidence of AD-related MCI or conversion from MCI to AD dementia has relied on cognitive, activities of daily living, and brain imaging measures. Limitations of this diagnostic approach include dependency on education and language, time-consuming and costly measures, and long-term monitoring. Emerging studies suggest that non-tremor motor dysfunction in dementias is known to be highly associated with AD biomarkers, with signs of cognitive decline visible in gait and hand movement at various stages of the illness. With the evidence that gait and physical disturbances are early predictors of cognitive impairment and that their trajectories could readily be tracked, we utilize recent advances in computer vision (CV) to quantify mobility in a data-driven fashion from the video-recorded 5-minute Short Performance Physical Battery (SPPB) tests. We use the data collected at Stanford AD Research Center and show that our CV methods can automatically reduce videos to body markers (human skeleton tracked through time) and extract several features (such as gait speed, mean torso inclination angle, double support time, gait-summary score, etc.) and finally turn those into clinical SPPB test scores. Our initial data observed a significant difference between healthy controls (HC) and the two MCI and AD groups for the repeated chair stand test score. Similarly, an inverse correlation between the MoCa cognitive test score and the gait speed is observed. At the end of the talk, I will also discuss how this CV method for mobility can be used for detecting behavioral changes in animal AD models and implications for future human AD research.
BibTeX
@inproceedings{adeli2024automated,
  author = {Ehsan Adeli},
  title = {{Automated Physical Performance Battery as a Digital Marker for Alzheimer’s Disease and Mild Cognitive Impairment}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.086276},
}
Figure