Toward Accurate Orthopedic Diagnosis: Improving Deep Learning-based Gait Anomaly Detection

We propose an anomaly detection framework for orthopedic disease recognition using gait models extracted via MediaPipe. Using LSTM-autoencoders, we analyzed gait data from 117 sarcopenia (SA) and 49 Parkinson’s disease (PD) patients, achieving detection rates of 97% and 88%, respectively. The knees emerged as the most sensitive joints for detecting gait abnormalities, while hips and nose-shoulder regions showed lower detection rates due to smaller angle variations and dataset limitations. To address environmental errors such as frame imbalance, clothing interference, and background noise, we implemented optimized video capture protocols and preprocessing using YOLO and semantic segmentation. These improvements increased usable gait data by 38% for healthy individuals, 33% for SA patients, and 2% for PD patients. Our results demonstrate the potential of combining markerless gait tracking with deep learning-based anomaly detection to advance orthopedic diagnostics and improve clinical decision-making.