Deep learning-based extension of gait segmentation to abnormal patterns using inertial measurement units

Inertial measurement unit (IMU)-based gait segmentation is widely employed in medical applications and plays a crucial role in recognizing gait phases. However, existing methods primarily focus on normal gait patterns, limiting their applicability to pathological cases such as small-stepped, dragging, and circumduction gaits. In this study, we propose a novel gait segmentation framework that can effectively handle both normal and abnormal gait patterns, thereby enhancing generalization of medical applications. Our main contribution is to expand gait segmentation to abnormal gait patterns in two ways: (1) We propose a new definition of gait segmentation to ensure equal treatment of normal and abnormal gaits, facilitating a more inclusive approach. (2) We also propose a novel network called gait segmentation neural network (GaitSeg Net), a deep learning model that integrates a convolutional neural network, bidirectional long short-term memory and transformer for robust feature extraction. This architecture employs wide-kernel CNNs to mitigate noise-related issues and a convolutional feedforward layer to filter out irrelevant information, significantly improving segmentation accuracy. We recorded a new dataset encompassing various normal and abnormal gaits for training and validation. Experimental results demonstrate that GaitSeg Net outperforms existing methods, achieving an F1-score of 98.16 %. Compared to a previous study, our method improves accuracy from 96.88 % to 97.50 % in walking and running tasks. Furthermore, our model maintains high accuracy for abnormal gaits (small-stepped gait: 96.1 %, dragging gait: 96.6 %, circumduction gait: 97.6 %), confirming its robustness. These results highlight the potential of our approach in extending gait segmentation to pathological movement patterns, marking a significant advancement in both artificial intelligence applications and biomedical engineering.