Age-related gait patterns classification using deep learning based on time-series data from one accelerometer

Gait pattern classification is important for healthcare. Conventional machine learning (CML) approaches based on handcrafted gait features are widely used in gait classification. However, extracting features may lead to suboptimal performance by omitting useful features. End-to-end deep learning (DL) approaches eliminate the need for feature extraction. However, some state-of-the-art DL approaches have not been explored in gait analysis. Furthermore, no consensus exists regarding the window sizes of input acceleration, which affects classification accuracy. In this study, data were collected from one accelerometer during a 3-minute indoor walking task. A total of 267 participants were divided into adults (18–65 years) and older adults (>65) groups. To explore age-related gait patterns classification performance, 5 DL approaches based on raw data and 4 CML approaches based on handcrafted features were compared. The results show that DL outperformed CML, with all AUC (Area under the receiver operator curve) greater than 0.94 compared to the best CML approach of 0.83. This suggests that DL may have learned important gait features related to aging that have not yet been identified by previous research. Furthermore, windows of different sizes ranging from 128 to 5120 samples were tested. The best performance of DL was achieved at a window size of 1024 (including about 20 steps). These findings indicate that the differences and relationship between gait cycles are important factors for classifying age-related gait patterns. This study could contribute to the development of more accurate gait pattern classification and assist in detecting age-related gait patterns in clinical environments.