Multiscale wavelet-based framework for machine-learning gait classification using inertial signals

This study presents a generalizable multiscale signal processing framework for time-series classification, demonstrated through gait analysis using inertial measurement unit (IMU) signals. The proposed approach leverages the Complex Morlet Wavelet Transform to extract time–frequency representations across multiple spectral scales, enabling enhanced discrimination of temporal dynamics in sequential data. A dataset comprising five locomotion tasks performed by younger and older adults was analyzed to identify scale-dependent descriptors sensitive to inter-group variability and activity type. Statistical feature selection was applied to retain discriminative wavelet-domain attributes. These features were used to train and evaluate four supervised classifiers–Support Vector Machine (SVM), Random Forest, XGBoost, and K-Nearest Neighbors (KNN)–across segmentation windows ranging from 1 to 10 seconds. Results indicate that classification performance improves with increasing window duration, with XGBoost achieving accuracies of 97% in younger adults and 96% in older adults using 10-second segments. Notably, short 1-second windows maintained competitive performance (up to 80%), supporting the feasibility of low-latency real-time classification. The results highlight the effectiveness of multiscale wavelet representations for modeling non-stationary temporal signals and support the proposed framework as a robust approach for time–frequency signal classification.