The feasibility of using deep learning technologies to preliminarily identify patients with advanced knee osteoarthritis via smartphone videos

Knee osteoarthritis (KOA) is a prevalent condition that often leads to a decline in patients’ physical function. Many patients with KOA struggle to discern whether the decline in their physical function results from natural physical deconditioning or the disease itself. Assessment of physical function holds promise as a potential method for identifying patients with advanced KOA. We developed and validated a simple at-home artificial intelligence method for identifying patients with advanced KOA. 357 participants independently recorded videos of sit-to-stand test and gait. By employing AlphaPose and VideoPose algorithms, we extracted three-dimensional spatiotemporal information from the videos. Subsequently, we employed the discrete wavelet transform (DWT) to analyze the data qualitatively and constructed time series models to identify patients with KOA. We extracted time series data directly collected by the participants. The analysis of spatiotemporal information revealed that the primary differences between patients with advanced KOA and individuals with declining physical function were in the overall movement patterns. Using the STS spatiotemporal information and demographic characteristics to construct the model, we achieved optimal performance with an AUC of 0.981 (95% CI 0.977–0.985). Our low-cost, user-friendly method effectively captures spatiotemporal information differences between patients with advanced KOA and those with declining physical function by smartphones and demonstrates high performance in distinguishing between these two populations. These findings provide compelling evidence for the feasibility of our low-cost, user-friendly method for large-scale initial screening of advanced KOA in targeted populations.