Session: 12-03: Structural Health Monitoring III

Paper Number: 147561

147561 - Medical Informed Machine Learning Method for Wearable Knee Health Monitoring System 

Abstract: 

Wearable electronics have undergone rapid development due to the emergence of flexible electronics and artificial intelligence. With the emphasis on sports and fitness among individuals, the prevalence of chronic diseases and musculoskeletal injuries in the knee joint continues to escalate, exerting significant pressure on the healthcare system. In response to this scenario, a knee joint health detection method is proposed based on flexible piezoelectric sensors and machine learning, designed for routine home medical care. The piezoelectric sensor utilized in this study is derived from previous research, wherein graphene particles were incorporated into lead magnesium niobate-lead titanate and polyvinylidene fluoride (PVDF) to fabricate a flexible ternary composite characterized by superior flexibility and excellent piezoelectric properties. The sensor layout is tailored to accommodate the biomechanical dynamics of knee joint motion. While machine learning methodologies have been extensively employed to extract health-related insights from collected motion data for diagnostic purposes, their efficacy relies heavily on the volume and quality of available data. In instances where data volume is limited or its quality is suboptimal, the reliability of these models may be compromised. Conversely, medical diagnoses based on clinical expertise retain inherent advantages in terms of credibility and accuracy. This study pioneers a novel machine learning approach that integrates empirical data with clinical insights and medical knowledge to elucidate the correlations between data patterns and underlying medical principles, thereby facilitating more precise diagnostics. Leveraging Bayesian networks, we utilize medical indicators and signal characteristics as inputs for model training, and incorporate expert knowledge to refine the network structure for the discrimination of individuals into healthy, sub-healthy, and unhealthy categories. Our findings demonstrate that our method achieves an accuracy rate of 83%, surpassing that of conventional machine learning techniques.

Presenting Author: Jingjing He School of Reliability and Systems Engineering, Beihang Universit

Presenting Author Biography: N/A

Authors:

Wenxi Zhang School of Reliability and Systems Engineering, Beihang University
Rundong Qu Department of Orthopaedics, Capital Medical University China-Japan Friendship School of Clinical Medicine
Zehao Fang School of Reliability and Systems Engineering, Beihang University
Ziwei Fang School of Reliability and Systems Engineering, Beihang University
Chenjun Gao School of Reliability and Systems Engineering, Beihang University
Jingjing He School of Reliability and Systems Engineering, Beihang University
Wei Sun Department of Orthopaedics, Capital Medical University China-Japan Friendship School of Clinical Medicine
Chang Wen Medical Center of Beihang Univerisity, Beijing
Jing Lin School of Reliability and Systems Engineering, Beihang University