Hannah Willms

Advisors
Arne Küderle (M.Sc.), , Prof. Dr. Björn Eskofier

Duration
07/2020 – 11/2020

Abstract
How well we can walk is an important factor in predicting the quality of life [1]. Further, limitations of gait can be a symptom of several neurological and musculoskeletal diseases such as Parkinson’s disease (PD). Consequently, gait analysis can be used as a tool to monitor disease progression [2]. However, this requires objective measurements of gait characteristics such as spatio-temporal parameters. Video-based motion-capture systems or computerized walkways have shown excellent reliability, but are costly and are confined to a lab environment [3]. Wearable inertial measurement units (IMUs) have been shown to overcome these limitations in the recent years. Because they are small and lightweight, they can be worn unobtrusively during everyday life and may provide reliable gait parameters from unconstrained environment [4]. In particular, sensors attached to the foot result in accurate single-stride parameters [5].

One disadvantage of IMUs is that in order to quantify gait and analyze clinically meaningful gait characteristics, complicated processing of the recorded signals is required. In the first step, this usually requires a segmentation of the continuous signal into individual strides [6]. The accuracy of this segmentation is important for the overall performance of the analysis. In result, multiple algorithms have been developed to solve this task as robust and as accurate as possible. Three widely used methods are peak detection, Hierarchical Hidden Markov Models (hHMMs), and multi-dimensional subsequence Dynamic Time Warping (msDTW) [6, 7].

In direct comparison of the three methods, HMMs performed the best, followed by msDTW [7]. However, in contrast to HMMs, msDTW does not require extensive training and only has a small number of parameters that need to be optimized. Further, the method can easily be personalized to individual patients by creating a new template based on a small number of hand-labeled strides. In addition, the algorithm inherently mimics the way a human labeler would segment strides based on the raw signal. This makes the output easy to understand, which could lead to a higher clinical acceptance. In result, msDTW remains a relevant method. However, to ensure that using msDTW does not require compromises when it comes to performance, some of the problems of msDTW must be addressed. For example, the msDTW algorithm is sensitive to changes of gait speed and
highly pathological or abnormal gait patterns. It is also prone to sometimes detect a segment of a stride or a short foot lift as a full stride [6]. This is especially a problem in home-monitoring datasets, because a variety of movements are present in the data. Further, because of the large number of strides in such datasets, even rare error will occur multiple times.

The goal of this thesis is to explore potential ways to improve the msDTW method on foot worn IMU sensors and to decrease needed manual fine-tuning of parameters. Preprocessing of the raw data, adding postprocessing steps to remove wrongly detected strides, and modification of the DTW algorithm itself will be explored as possibilities to improve the method. The modifications will be tested on lab datasets containing different gait tests from healthy subjects and an unsupervised home-monitoring dataset containing data from healthy subjects and subjects with Parkinson’s disease. The modified algorithms are compared against expert labels of stride borders, generated by manual visual inspection of the raw gyroscope data, and to the existing msDTW algorithm [6].

References:
[1] Ellis, Terry et al.: Which measures of physical function and motor impairment best predict quality of life in Parkinson’s disease?. Parkinsonism & related disorders, 2011; 17(9)
[2] Mirelman, Anat et al.: Gait impairments in Parkinson’s disease. Lancet Neurol, 2019.
[3] Verghese, Joe et al.: Quantitative gait markers and incident fall risk in older adults. The journals of gerontology Series A, Biological sciences and medical sciences. 2009; 64(8)
[4] Schlachetzki, Johannes C. M. et al.: Wearable sensors objectively measure gait parameters in parkinsons disease. PLoS ONE 12(10): e0183989., 2017.
[5] Rampp, Alexander et al.: Inertial Sensor-Based Stride Parameter Calculation From Gait Sequences in Geriatric Patients. IEEE Transactions on Biomedical Engineering, 62.4, 2015
[6] Barth, Jens et al.: Stride Segmentation during Free Walk Movements Using Multi-Dimensional Subsequence Dynamic Time Warping on Inertial Sensor Data shoes. Sensors 15.3, 6419-6440, 2005.
[7] Ghassemi, Nooshin Haji et al.: Segmentation of Gait Sequences in Sensor-Based Movement Analysis: A Comparison of Methods in Parkinson’s Disease. Sensors (Switzerland), 18.1, .018