Annika Mücke

Advisors
Martin Ullrich, M.Sc., Prof. Dr. Björn Eskofier

Duration
05/2020 – 10/2020

Abstract
Motor impairments, often accompanied with postural instability and freezing of gait, are a severe symptom in Parkinson’s disease (PD). Consequently, the risk for injury-causing falls is increased for PD patients. On average, 39 % of all Parkinson’s patients fall recurrently [1]. Sensor-based gait analysis using inertial measurement units (IMUs) has been shown to be a helpful tool for determining the severity of the disease objectively. Based on gyroscope and accelerometer data from IMUs, spatio-temporal parameters such as stride length, stance and swing phase time or
gait velocity can be determined from supervised 4×10 m or other standardized gait tests. These parameters can be utilized to assess the stage of disease [2] or predict the risk of falling [3] of a patient.

The recording of data can either be done in a clinical setting or under free-living conditions using long-term mobile gait monitoring. The latter is advantageous since it provides a comprehensive, continuous insight into the patient’s condition rather than a momentary impression [3].
However, the analysis of IMU signals recorded in free-living conditions is challenging as it contains the data from arbitrary activities in contrast to supervised gait tests where only walking is recorded. Therefore, semi-standardized gait tests recorded in the patients’ home environment can provide helpful to obtain walking data from repetitive conditions in terms of location and time during the day. For an efficient analysis of these gait tests, a reliable timestamp annotation of their executions is required for the identification of the respective data from the continuous sensor signal streams.

The MaD Lab currently runs a joint study with the University Hospital (FallRiskPD) where machine learning algorithms are investigated for the prediction of fall risk from long-term IMU gait data [4]. For a period of two weeks, patients were instructed to wear IMU sensors, attached to
both feet and the lower back, during their daily activities. Additionally, they were asked to perform 4x10m tests at different speeds as well as a Timed Up and Go test in their home environment three times a day. To extract those gait test sequences from the long-term data for further processing, the corresponding timestamps are required. For approximate annotations, a smartphone application was provided to the participants. Because this constitutes a possible source of error, a manual verification and refinement or correction of false labels is necessary.
The main goal of this thesis is to develop an algorithm for gait test detection in free-living gait data that does not require manual labeling or smartphone annotations. The key benefits are not only to reduce the amount of time needed for data processing, but also to decrease the effort
for the patients. Previously, a pipeline for automatic detection and labeling of clinical gait tests was proposed, consisting of the following steps [5]: First, gait sequences are extracted by detecting patterns of harmonics in the frequency domain of the gyroscope signal [6]. Afterwards, the gait sequences are classified using subsequent dynamic time warping. It allows scaling and stretching invariant template matching [7] and is thus suitable for gait tests carried out at different selfselected speeds and under varying conditions regarding the exact length of the walking path. In a postprocessing step, the results are verified by checking the amount of steps and turning sequences within the detected gait tests.

For this thesis the existing pipeline (including gait sequence detection and template matching) will be adapted for the applicability to home monitoring data as recorded in the FallRiskPD study. One persisting problem is the detection of individual gait tests executed without any considerable resting periods between them. Furthermore, the sensitivity of the algorithm for varying walking speeds needs to be improved. Within this thesis, the pipeline is going to be adapted accordingly and validated against the expert-labeling of the free-living gait data.

References:
[1] Allen, Natalie et al.: Recurrent Falls in Parkinson’s Disease: A Systematic Review. Parkinson’s disease, 906274, 2013.
[2] Schlachetzki, Johannes et al.: Wearable sensors objectively measure gait parameters in Parkinson’s disease. PLOS ONE 12.10, e0183989, 2017.
[3] Del Din, Silvia et al.:Analysis of Free-Living Gait in Older Adults With and Without Parkinson’s Disease and With and Without a History of Falls: Identifying Generic and Disease-Specific Characteristics. The Journals of Gerontology: Series A 74.4, 500-506, 2019.
[4] https://www.mad.tf.fau.de/research/projects/fallriskpd/, accessed: 2020-04-15.
[5] Fischer, Stefan et al.: Macro Analysis of free-living Gait in Parkinson’s Disease. Bachelors Thesis, Friedrich-Alexander-Universität Erlangen-Nürnberg, 2019.
[6] Ullrich, Martin et al.: Detection of Gait From Continuous Inertial Sensor Data Using Harmonic Frequencies. IEEE Journal of Biomedical and Health Informatics, 1-1, 2020.
[7] Barth, Jens et al.: Stride Segmentation during Free Walk Movements Using Multi-Dimensional Subsequence Dynamic Time Warping on Inertial Sensor Data. Sensors 15.3, 6419-6440, 2005.

 

Full thesis