Real-Time Forecasting of Pathological Gait via IMU Navigation: A Few-Shot and Generative Learning Framework for Wearable Devices

Current gait analysis faces challenges in various aspects, including limited and poorly labeled data within existing wearable electronics databases, difficulties in collecting patient data due to privacy concerns, and the inadequacy of the Zero-Velocity Update Technique (ZUPT) in accurately analyzing pathological gait patterns. To address these limitations, we introduce GaitMotion, a novel machine-learning framework that employs few-shot learning on a multitask dataset collected via wearable IMU sensors for real-time pathological gait analysis. GaitMotion enhances data quality through detailed, ground-truth-labeled sequences and achieves accurate step and stride segmentation and stride length estimation, which are essential for diagnosing neurological disorders. We incorporate a generative augmentation component, which synthesizes rare or underrepresented pathological gait patterns. GaitMotion achieves a 65% increase in stride length estimation accuracy compared to ZUPT. In addition, its application to real patient datasets via transfer learning confirms its robust predictive capability. By integrating generative AI into wearable gait analysis, GaitMotion not only refines the precision of pathological gait forecasting but also demonstrates a scalable framework for leveraging synthetic data in biomechanical pattern recognition, paving the way for more personalized and data-efficient digital health services.

💡 Research Summary

The paper addresses three major shortcomings of current wearable‑based gait analysis: scarce and poorly labeled datasets, privacy‑driven difficulty in collecting patient data, and the inadequacy of the Zero‑Velocity Update (ZUPT) technique for pathological gait patterns. To overcome these issues, the authors propose GaitMotion, a machine‑learning framework that combines few‑shot meta‑learning with generative data augmentation, built on a multitask dataset recorded from foot‑mounted inertial measurement units (IMUs) synchronized with a gold‑standard GaitRite pressure mat.

Data collection involved ten healthy adults (average age 24.8 ± 3.5 years) performing three gait conditions—normal, Parkinson‑like “shuffle”, and stroke‑like asymmetric walking—six repetitions each, yielding 180 sequences. For every trial, raw 3‑axis accelerometer and gyroscope streams were recorded alongside precise step‑segmentation labels (heel‑strike, toe‑off) and a suite of gait parameters (swing/stance times, stride length, stride time, etc.). This results in a rich multimodal dataset comprising raw IMU signals, segmentation masks, and seven derived gait metrics.

Methodologically, GaitMotion consists of four key components: (1) a few‑shot learning strategy that enables rapid adaptation to new gait types with only a handful of labeled examples; (2) a generative augmentation pipeline that employs a pretrained time‑series GAN to synthesize 50 realistic pathological IMU sequences (asymmetric swing, foot‑drop, delayed toe‑off), thereby balancing the class distribution; (3) a convolutional neural network (three 1‑D Conv layers with Leaky‑ReLU, batch normalization, and a Softplus‑activated fully‑connected regression head) trained to predict continuous stride length; and (4) a transfer‑learning protocol that fine‑tunes the pretrained model on the eGait dataset (101 elderly subjects) using minimal patient labels. Training used a learning rate of 5 × 10⁻⁵, batch size 64, and was performed on an NVIDIA RTX 4060 GPU.

Experimental results demonstrate that GaitMotion outperforms ZUPT by a substantial margin: stride‑length root‑mean‑square error drops from 0.159 m (ZUPT) to 0.056 m, a 65 % improvement. Adding synthetic data improves F1‑score on pathological gait detection by roughly 12 % compared with a model trained on real data alone. In the eGait transfer scenario, the fine‑tuned model achieves performance within 3 % of a fully supervised baseline while requiring only about 5 % of the labeled data.

The authors discuss several limitations. First, the “mimicked” pathological gait performed by healthy participants may not capture the full biomechanical complexity of genuine patient data. Second, synthetic samples inherit any bias present in the GAN’s training distribution, potentially skewing model behavior. Third, the original dataset’s small size (10 subjects) constrains scalability. Future work is suggested to incorporate authentic patient recordings, explore more advanced generative models such as VAE‑GANs or diffusion models, and optimize the architecture for on‑device inference on low‑power wearables.

Ethical considerations are addressed through UBC Institutional Review Board approval (H21‑02052), informed consent, anonymization, secure storage, and careful sensor design to minimize skin irritation.

In conclusion, GaitMotion provides a viable, data‑efficient solution for real‑time pathological gait forecasting on wearable devices. By leveraging few‑shot learning and synthetic data generation, it achieves higher accuracy than traditional inertial navigation methods, reduces dependence on large labeled patient cohorts, and opens pathways toward personalized digital health services and adaptive rehabilitation programs.