Performance Evaluation of Gait Classification Using Quaternion Waveforms

Visual gait assessment in physical therapy is subjective and may lack accuracy. Three-dimensional motion analysis is precise, but it is expensive and difficult to use in daily clinical settings. This study examined whether abnormal gait patterns can be analyzed using quaternion time-series data obtained from a single smartphone-based inertial measurement unit (IMU). Nineteen healthy adults performed normal walking and four types of simulated abnormal gait. A smartphone was fixed at the waist, and gait data were recorded during a 10 m walk. Gait cycles were extracted and normalized to 100 points. The four quaternion components were analyzed as multivariate time-series data over one gait cycle. Principal component analysis and clustering showed differences in waveform structure between gait patterns. Clear separation was observed between normal gait and Trendelenburg gait. A support vector machine classifier achieved an accuracy of 0.83 and a Macro-F1 score of 0.84 using cross-validation. When leave-one-subject-out validation was applied, classification performance decreased. This result suggests that individual differences and waveform variability influence gait classification. These findings indicate that quaternion time-series data obtained from a single smartphone IMU can capture structural characteristics of gait patterns. This approach may support objective and practical gait analysis in physical therapy.