Hidden Markov Model based stride segmentation on unsupervised free-living gait data in Parkinson’s disease patients

Abstract

Background To objectively assess a patient’s gait, a robust identification of stride borders is one of the first steps in inertial sensor-based mobile gait analysis pipelines. While many different methods for stride segmentation have been presented in the literature, an out-of-lab evaluation of respective algorithms on free-living gait is still missing. Method To address this issue, we present a comprehensive free-living evaluation dataset, including 146.574 semi-automatic labeled strides of 28 Parkinson’s Disease patients. This dataset was used to evaluate the segmentation performance of a new Hidden Markov Model (HMM) based stride segmentation approach compared to an available dynamic time warping (DTW) based method. Results The proposed HMM achieved a mean F1-score of 92.1% and outperformed the DTW approach significantly. Further analysis revealed a dependency of segmentation performance to the number of strides within respective walking bouts. Shorter bouts ($$< 30$$ < 30 strides) resulted in worse performance, which could be related to more heterogeneous gait and an increased diversity of different stride types in short free-living walking bouts. In contrast, the HMM reached F1-scores of more than 96.2% for longer bouts ($$> 50$$ > 50 strides). Furthermore, we showed that an HMM, which was trained on at-lab data only, could be transferred to a free-living context with a negligible decrease in performance. Conclusion The generalizability of the proposed HMM is a promising feature, as fully labeled free-living training data might not be available for many applications. To the best of our knowledge, this is the first evaluation of stride segmentation performance on a large scale free-living dataset. Our proposed HMM-based approach was able to address the increased complexity of free-living gait data, and thus will help to enable a robust assessment of stride parameters in future free-living gait analysis applications.

Review—Emerging Portable Technologies for Gait Analysis in Neurological Disorders

“…Particularly in times of global pandemics, where the number of regular face-to-face visits is reduced (Roy et al, 2020), automatically extracted and shared parameters from wearables have the potential to support clinical decisions. Automatic evaluation methods of data from wearables in clinical gait and balance assessments (e.g., Karatsidis et al, 2017;Nguyen et al, 2019) but also in unrestricted activities of daily living (e.g., Roth et al, 2021) are constantly investigated. Machine learning techniques are the driving force behind this rapid growth of applications.…”

The placement of foot-mounted IMU sensors does affect the accuracy of spatial parameters during regular walking

“…Therefore, all strides in the dataset were labeled manually by gait experts independently based on the medial-lateral gyroscope signal of the instep sensor. The start and the end of each stride were marked based on the clearly visible minimum before the terminal contact (for more information about the labeling process see [ 31 ]). To ensure consistency, the manual labels were moved to the exact sample of the minima in a 50 ms window around the labeled point.…”

The Diverse Gait Dataset: Gait Segmentation Using Inertial Sensors for Pedestrian Localization with Different Genders, Heights and Walking Speeds

“…From the perspective of frequency signal decomposition, wavelet analysis provided insight into determining stride borders, and it is suggested that better performance could be achieved in the frequency domain than in the time domain [42][43][44][45]. Another kind of method employs Hidden Markov Models (HMM) [46][47][48], residual neural networks [29], etc. These methods could achieve better detection accuracy in stride segmentation or gait recognition assignments with the assurance that large-scale training data should be offered and massive computing resources supplied.…”