Identifying changes in dynamic plantar pressure associated with radiological knee osteoarthritis based on machine learning and wearable devices

• Published in: Journal of neuroengineering and rehabilitation Vol. 21; no. 1; pp. 45 - 10
• Main Authors: Li, Gege, Li, Shilin, Xie, Junan, Zhang, Zhuodong, Zou, Jihua, Yang, Chengduan, He, Longlong, Zeng, Qing, Shu, Lin, Huang, Guozhi
• Format: Journal Article
• Language: English
• Published: London BioMed Central 03.04.2024; BioMed Central Ltd; BMC
• Subjects: Algorithms; Arthritis; Biomarkers; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Body mass index; Body size; Care and treatment; Center of pressure; Data mining; Diagnosis; Feature selection; Gait; Gait analysis; Health aspects; Heels; Humans; Knee; Knee osteoarthritis; Learning algorithms; Machine Learning; Methods; Model accuracy; Neurology; Neurosciences; Osteoarthritis; Osteoarthritis, Knee - diagnostic imaging; Peak pressure; Plantar fasciitis; Plantar pressure; Pressure; Radiography; Rehabilitation Medicine; Sensors; Shoes & boots; Support vector machines; Surgery; Walking; Wearable computers; Wearable device; Wearable Electronic Devices; Wearable technology; China; Wearable device; Plantar pressure; Knee osteoarthritis; Gait analysis; Machine learning
• ISSN: 1743-0003; 1743-0003
• DOI: 10.1186/s12984-024-01337-6

Background Knee osteoarthritis (KOA) is an irreversible degenerative disease that characterized by pain and abnormal gait. Radiography is typically used to detect KOA but has limitations. This study aimed to identify changes in plantar pressure that are associated with radiological knee osteoarthritis (ROA) and to validate them using machine learning algorithms. Methods This study included 92 participants with variable degrees of KOA. A modified Kellgren–Lawrence scale was used to classify participants into non-ROA and ROA groups. The total feature set included 210 dynamic plantar pressure features captured by a wearable in-shoe system as well as age, gender, height, weight, and body mass index. Filter and wrapper methods identified the optimal features, which were used to train five types of machine learning classification models for further validation: k-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), AdaBoost, and eXtreme gradient boosting (XGBoost). Results Age, the standard deviation (SD) of the peak plantar pressure under the left lateral heel (f_L8PPP_std), the SD of the right second peak pressure (f_Rpeak2_std), and the SD of the variation in the anteroposterior displacement of center of pressure (COP) in the right foot (f_RYcopstd_std) were most associated with ROA. The RF model with an accuracy of 82.61% and F1 score of 0.8000 had the best generalization ability. Conclusion Changes in dynamic plantar pressure are promising mechanical biomarkers that distinguish between non-ROA and ROA. Combining a wearable in-shoe system with machine learning enables dynamic monitoring of KOA, which could help guide treatment plans.