High Accuracy Machine Learning Model for Sarcopenia Severity Diagnosis based on Sit-to-stand Motion Measured by Two Micro Motion Sensors

Keer Wang, Hongyu Zhang, Clio Yuen Man Cheng, Meng Chen, King Wai Chiu Lai, Calvin Kalun Or, Yonghua Chen, Yong Hu, Vellaisamy A. L. Roy, Cindy Lo Kuen Lam, Ning Xi, Vw Lou, Wen J. LimedRxiv 2023 · DOI

Abstract

Sarcopenia, an age-related syndrome characterized by the loss of skeletal muscle mass and function, is a significant contributor to falls, fractures, and decreased quality of life in the elderly. While clinical standards like MRI and DEXA scans are accurate, their high cost and lack of portability limit routine screening. This study proposes a portable, high-accuracy diagnostic system using two micro-Inertial Measurement Units (μIMUs) and machine learning to analyze sit-to-stand (STS) motions. By segmenting the STS movement into four distinct phases and extracting 510 features, the system achieved a peak accuracy of 98.32% in binary classification (healthy vs. sarcopenia) and 90.44% in identifying four specific severity levels.

Introduction

As global populations age, sarcopenia has become a critical geriatric concern. Early diagnosis is essential for effective intervention, yet current methods often provide only binary results or require expensive hospital-grade equipment. The researchers in this study addressed the need for an objective, portable tool that can not only detect sarcopenia but also quantify its severity (Risk Levels 0 to 3), enabling more personalized treatment plans.

Methodology

The study recruited 53 elderly subjects (average age 70) categorized into healthy (12 subjects) and sarcopenic (41 subjects) groups based on AWGS 2019 criteria. Participants performed a single sit-to-stand motion while wearing two μIMUs on their waist and thigh. These sensors recorded acceleration and angular velocity, which were transmitted via Bluetooth for analysis.

Overview of the diagnostic system pipeline, illustrating data acquisition, phase segmentation, feature extraction, and classification into risk levels.

Phase Segmentation and Feature Extraction

A core innovation of this research is the granular analysis of the STS motion. Rather than treating the movement as a single event, the researchers divided it into four distinct phases based on five critical biomechanical points: initiation, peak flexion, seat-off, peak extension, and termination.

Detailed breakdown of the four phases of the sit-to-stand motion based on angular velocity and angle data.

From these segments, 510 features were extracted across time and frequency domains. To optimize performance and avoid the "curse of dimensionality," a Sequential Feature Selection (SFS) algorithm was employed to identify the most relevant features for five different machine learning classifiers: Support Vector Machines (SVM), K-Nearest Neighbor (KNN), Decision Trees (DT), Linear Discriminant Analysis (LDA), and Multilayer Perceptron (MLP).

Key Results

  • Binary Classification: The system achieved a peak accuracy of 98.32% in distinguishing healthy subjects from those with sarcopenia using an SVM model.
  • Severity Classification: The model achieved 90.44% accuracy in classifying subjects into four risk levels (0, 1, 2, and 3) using the KNN algorithm.
  • Efficiency: The Decision Tree model proved highly efficient, requiring only 8 features to maintain high accuracy, suggesting the potential for lightweight, real-time diagnostic applications.

Discussion and Future Work

The study demonstrates that phase-specific biomechanical signatures, such as those captured during "seat-off" or "peak extension," are far more predictive of muscle health than simple duration timers. While the current results are promising, the authors note that the small sample size (particularly in Risk Level 2) is a limitation. Future research will focus on larger datasets and the integration of these sensors into everyday items like smart belts or clothing for continuous, non-invasive monitoring of muscle health.

Summary updated

Abstract

Abstract In this ageing society, sarcopenia as a geriatric condition that can have significant negative impacts on an individual’s quality of life. Sarcopenia is a kind of aged syndrome associated with loss of muscle mass and function, which may lead to falls, fractures, gait disorders or even mortality. There are multiple ways to diagnose sarcopenia, such as using Magnetic resonance imaging (MRI), Dual-energy X-ray absorptiometry (DEXA) and Bioelectrical impedance analysis (BIA) etc. to calculate muscle mass; using handgrip or sit-to-stand to measure muscle strength; using short physical performance battery (SPPB), gait, and 5-time sit-to-stand to evaluate physical performance. In this work, we use two μIMUs worn on subjects to record their sit-to-stand motion, and then used several machine learning models to diagnose the severity of sarcopenia of the subjects. We recruited 53 elderly subjects in total for this work. The youngest subject is 65 years old and the oldest is 84 years old. Their average age is 70 years old. Among these 53 subjects, there are 12 healthy ones and 41 sarcopenia patients with different severity. The subject is instructed to do the single sit-to-stand (STS) three times, and two μIMUs attached to the subject’s waist and thigh transfer the data to a computer by Bluetooth. We separated the STS motion process into 4 phases based on the angle and angular velocity, extracted a total of 510 features for motion analytics. These features were futher analyzed by sequential feature selection with 5 different machine learning models (SVM, KNN, decision tree, LDA, and multilayer perceptron). With our proposed methodology, all 53 subjects could be classified as healthy or having sarcopenia with risk level 1, 2, or 3. The best accuracy to distinguish the healthy or sarcopenia subjects is 98.32%, and the best results to distinguish sarcopenia risk levels from 0 (healthy) to 3 (most severe) is 90.44%.
BibTeX
@misc{wang2023high,
  author = {Keer Wang and Hongyu Zhang and Clio Yuen Man Cheng and Meng Chen and King Wai Chiu Lai and Calvin Kalun Or and Yonghua Chen and Yong Hu and Vellaisamy A. L. Roy and Cindy Lo Kuen Lam and Ning Xi and VW Lou and Wen J. Li},
  title = {{High Accuracy Machine Learning Model for Sarcopenia Severity Diagnosis based on Sit-to-stand Motion Measured by Two Micro Motion Sensors}},
  year = {2023},
  doi = {10.1101/2023.05.18.23289933},
  howpublished = {medRxiv},
}
Figure