Real-Time Gait Anomaly Detection Using SVM Time Series Classification

In this paper, a real-time implementation of Support Vector Machines (SVM) - Real-Time tsSVM Anomaly Detection (RTtsSVM-AD) algorithm is proposed. Here, real-time abnormality detection is referencing to the ability of the algorithm to detect true gait anomaly occurrence during the swing phase of ongoing step. Anomaly detection is presented with "earliness" measure. For comparative research, eight different human gait deviations were simulated by two healthy volunteers. Corresponding gyroscope angular velocities, from the sensor placed on the forefoot, were recorded. F1 score, true positive rate (TPR), false positive rate (FPR) and "earliness" values were estimated and analyzed. Real-time classification results, where classification is performed during the ongoing step, are different from regular classification results, where classification is performed after the full step. Thus, they can not be compared directly. Achieved results prove the concept, that it is possible to detect anomalies in real-time during the swing phase of a step with RTtsSVM-AD algorithm. Best F1 scores for first person's gait recordings were 57%, 53% and 52% for Steppage, Parkinsonian and Ataxic gait types respectively. For the second person's gait recordings, best F1 scores were 65%, 58% and 50% for Slap, Steppage and Hemiplegic gait types respectively. RTtsSVM-AD algorithm would be developed further and could be used as a base method for comparison with other algorithms.