Elsevier

Gait & Posture

Volume 80, July 2020, Pages 130-136
Gait & Posture

A data-driven approach for detecting gait events during turning in people with Parkinson's disease and freezing of gait

https://doi.org/10.1016/j.gaitpost.2020.05.026Get rights and content

Abstract

Background

Manual annotation of initial contact (IC) and end contact (EC) is a time consuming process. There are currently no robust techniques available to automate this process for Parkinson's disease (PD) patients with freezing of gait (FOG).

Objective

To determine the validity of a data-driven approach for automated gait event detection.

Methods

15 freezers were asked to complete several straight-line and 360 degree turning trials in a 3D gait laboratory during the off-period of their medication cycle. Trials that contained a freezing episode were indicated as freezing trials (FOG) and trials without a freezing episode were termed as functional gait (FG). Furthermore, the highly varied gait data between onset and termination of a FOG episode was excluded. A Temporal Convolutional Neural network (TCN) was trained end-to-end with lower extremity kinematics. A Bland-Altman analysis was performed to evaluate the agreement between the results of the proposed model and the manual annotations.

Results

For FOG-trials, F1 scores of 0.995 and 0.992 were obtained for IC and EC, respectively. For FG-trials, F1 scores of 0.997 and 0.999 were obtained for IC and EC, respectively. The Bland-Altman plots indicated excellent timing agreement, with on average 39% and 47% of the model predictions occurring within 10 ms from the manual annotations for FOG-trials and FG-trials, respectively.

Significance

These results indicate that our data-driven approach for detecting gait events in PD patients with FOG is sufficiently accurate and reliable for clinical applications.

Introduction

Freezing of gait (FOG) is a devastating gait disorder manifesting itself frequently in Parkinson's disease (PD). FOG is defined by Nieuwboer and Giladi as “an episodic inability to generate effective stepping in the absence of any known cause other than Parkinsonism or high-level gait disorder” [1]. Patients describe a FOG episode as “the feeling that their feet are glued to the ground” [2]. FOG occurs most reliably during complex gait tasks, such as turning with fast speed or walking while performing a dual task [3]. To study FOG and the highly abnormal steps leading up to it, gait analysis has been adopted, using instrumented gait analysis systems based on 3D motion capturing techniques [4], [5]. The gait data generated from these systems are typically normalized to a gait cycle. This normalization requires accurate timing of initial contact (IC) and end contact (EC) of the foot. The detection of these gait events is based on visual inspection by a clinical expert [4], [5]. Due to the small and shuffling steps, reduced heel strike and inadequate swing phase prior to FOG [6], and altered steps between FOG episodes [7], this process is imprecise. In addition, visual detection of gait events are more time consuming, during more complex gait tasks such as 360 degree turning [8].

To find a solution for this problem, this paper aimed to investigate the validity of an automated approach for gait cycle detection. Heuristic based methods are most commonly used to automatically detect the defined gait events. These methods utilize domain knowledge to extract kinematic features that correlate with the timing of gait events. However, owing to the variable gait patterns apparent in PD patients with FOG, these features do not necessarily generalize to this pathology. Furthermore, heuristic methods typically lack validation in challenging movement sequences, such as turning and dual tasking, commonly used to trigger FOG [3].

Powered by large datasets, data-driven approaches, such as recurrent neural networks (RNN), have shown great success in many problems that contain temporal information. These approaches can infer relevant features directly from the raw input data, a technique called end-to-end learning [9]. The success of these approaches for gait event detection was recently demonstrated [10], utilizing a long short-term memory (LSTM) network to classify gait events in children. The focus of this paper was to provide a robust tool to automatically annotate gait events for PD patients with FOG during straight-line gait and turning, which can be trained end-to-end with minimal data pre-processing.

Section snippets

Sequence to sequence learning

In this study, gait event detection is cast as a sequence to sequence classification problem [11]. Each input sample x is associated with a ground-truth label yobs. A model is trained to learn a function f : x → y that transforms a given input sequence X = x0, …, xt into an output sequence Y = y0, …, yt that closely resembles the manual annotations Yobs. Separate datasets are generated for each gait event by encoding each sample as a binary vector yobs ∈ {0, 1}. The input sequence Xins×t is

Results

For the freezing trials (FOG), a total of 506 IC and 491 EC events were acquired. The TCN model shows F1-scores of 0.995 and 0.992 for IC and EC, respectively. The LSTM model shows F1-scores of 0.989 and 0.976 for IC and EC, respectively. The heuristic method shows F1-scores of 0.976 and 0.956 for IC and EC, respectively. For the functional gait trials (FG), a total of 741 IC and 669 EC events were acquired. The TCN model shows F1-scores of 0.997 and 0.999 for IC and EC, respectively. The LSTM

Discussion

We evaluated two data-driven approaches for the detection of gait events that were trained end-to-end on a small dataset of straight-line gait and 360 degree turning of PD patients with FOG. A total of 2407 events have been manually annotated and these events were used to quantitatively validate the algorithms in terms of accuracy and timing agreement. A commonly used heuristic method proposed in [21] was reproduced to allow a quantitative comparison with the deep learning models on the same

Data availability

The input set was imported and labelled using Python version 2.7.12 with Biomechanical Toolkit (btk) version 0.3 [29]. Peak detection was done with Scipy [26]. Deep learning models were trained on an NVIDIA Tesla K80 GPU using Python version 3.6.8 and Tensorflow version 1.14 [30]. Hyperparameters were optimized using the Hyperopt python library [31], with cross validation splits created with scikit-learn version 0.21.3 [32]. Utility functions for processing c3d files were adopted from [10]. All

Financial disclosure

None reported.

Conflict of interest

None reported.

References (33)

  • A. Nieuwboer et al.

    Abnormalities of the spatiotemporal characteristics of gait at the onset of freezing in Parkinson's disease

    Mov. Disord.

    (2001)
  • J. Nonnekes et al.

    Short rapid steps to provoke freezing of gait in Parkinson's disease

    J. Neurol.

    (2014)
  • M. Plotnik et al.

    The role of gait rhythmicity and bilateral coordination of stepping in the pathophysiology of freezing of gait in Parkinson's disease

    Mov. Disord.

    (2008)
  • M. Mancini et al.

    Clinical and methodological challenges for assessing freezing of gait: future perspectives

    Mov. Disord.

    (2019)
  • Z. Wang et al.

    Time Series Classification From Scratch with Deep Neural Networks: A Strong Baseline

    (2016)
  • Ł. Kidziński et al.

    Automatic real-time gait event detection in children using deep neural networks

    PLOS ONE

    (2019)
  • View full text