Reliable gait phase classification is important for wearable-based locomotion analysis, rehabilitation monitoring, and health assessment. Traditional gait analysis systems can be expensive and confined to laboratory settings, while inertial measurement units (IMUs) offer a portable way to capture walking dynamics in daily or clinical environments.
In a paper published in Artificial Intelligence and Autonomous Systems , Huanghe Zhang of Shandong University reports a transfer learning framework that connects two related gait-analysis tasks: predicting continuous gait cycle percentage and classifying discrete gait phases such as stance and swing.
The method first trains neural network backbones on gait cycle percentage prediction, a regression task that encourages the model to learn smooth temporal patterns in walking. These learned representations are then transferred to gait phase classification through model transfer or feature transfer.
The study compares a compact deep neural network (DNN) with a Transformer model across several sliding-window sizes. The best model-transfer configuration reached an F1-score of 0.9788, outperforming feature transfer baselines and models trained from scratch.
Efficiency was a central part of the evaluation. The compact DNN contains about 0.3 million parameters and achieved CPU inference latency below 0.07 ms, supporting real-time processing for resource-constrained wearable systems.
The framework was also validated on an independent dataset from healthy young adults, where it achieved 92.3% classification accuracy. This cross-population result suggests that representations learned from continuous gait progression can capture gait dynamics that generalize beyond the original older-adult dataset.
The findings point to a practical workflow for wearable gait analysis: a single regression-pretrained model can be adapted to downstream gait phase recognition with less redundant training while maintaining high classification performance.
This paper "Transfer learning from gait cycle percentage prediction to gait phase classification using wearable sensors" was published in Artificial Intelligence and Autonomous Systems .
Zhang H. Transfer learning from gait cycle percentage prediction to gait phase classification using wearable sensors. Artif. Intell. Auton. Syst. 2026(1):0004, https://doi.org/10.55092/aias20260004.
Artificial Intelligence and Autonomous Systems
Experimental study
Not applicable
Transfer learning from gait cycle percentage prediction to gait phase classification using wearable sensors
28-May-2026