Reliability and validity of OpenPose for measuring HKA angle in dynamic walking videos in patients with knee osteoarthritis

Abstract

The hip-knee-ankle (HKA) angle is essential to assess surgical evaluation and disease progression in patients with knee osteoarthritis (KOA). Rapid, radiation-free assessment methods are a key area of research. This study investigates the reliability and validity of OpenPose, video-based human pose estimation method, for determining the HKA angle in KOA patients. In this study, we analyzed 50 knees affected by osteoarthritis. The HKA angle was measured using the pose estimation method and X-ray imaging before total knee arthroplasty. The pose estimation method demonstrated excellent test-retest reliability (ICC 1,1 = 1.000) and good consistency with radiography (ICC 2,1 = 0.897), with linear regression analysis showing a good correlation (R² = 0.814). Compared with radiography, the pose estimation method exhibited a fixed error of 0.131°. This is the first study to examine the feasibility of measuring the HKA angle from frontal-view videos of patients walking normally by using the pose estimation method. Using the pose estimation method to measure the HKA angle in knee osteoarthritis patients is reliable and valid. The pose estimation method provides a safe, cost-effective, and user-friendly solution for monitoring lower limb alignment, with promising applications in remote healthcare and rehabilitation management. It eliminates radiation exposure, avoiding the health risks associated with X-ray imaging, and it does not require specialized medical equipment, enabling fully automated analysis.

Keywords: Dynamic walking videos; Hip-knee-ankle angle; Knee osteoarthritis; OpenPose.

Fig. 1

Workflow correspondent to the data processing using OpenPose. The illustration shows the workflow for video recording and OpenPose processing. First, a single smartphone captures the walking video of the subject. Then, OpenPose analyzes the video recording, and a custom Python script processes the data. Finally, compare the HKA angle with radiography. (A) Sample image from a participant’s recorded walk; (B) The OpenPose skeletal model and the name of each part. From: Sugiyama Y, Uno K, Matsui Y. Types of anomalies in two-dimensional video-based gait analysis in uncontrolled environments. PLoS Comput Biol. 2023;19(1):e1009989. (C) A sample output of OpenPose with the detected keypoints of interest plotted on the image; (D): Preoperative HKA angle measurement method in Radiographic image. The image was created using BioRender (https://biorender.com/) with publication and licensing rights (Agreement number: US27WFZL9I).

Fig. 2

Bland–Altman plot for assessing measurement error. Mean hip-knee-ankle angle measurements for OpenPose and radiography are plotted on the x-axis and the difference between measurements on the y-axis. HKA, hip-knee-ankle.

Fig. 3

Examples of significant errors in OpenPose pose estimation. (A) Clothing and other factors, which are not accounted for in training data, significantly impact the accuracy of joint detection. (B) Large distances reduce OpenPose performance, leading to inaccurate joint localization and increased errors. (C) As the hip-knee-ankle (HKA) angle deviates further from 180 degrees, the accuracy of joint detection and HKA angle calculation decreases significantly.