Neurocomputing 315 (2018) 1–8
Contents lists available at ScienceDirect
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
A hybrid spatio-temporal model for detection and severity rating of
Parkinson’s disease from gait data
Aite Zhao
a
, Lin Qi
a , 1
, Jie Li
a
, Junyu Dong
a , ∗
, Hui Yu
b
a
Department of Computer Science and Technology, Ocean University of China, Qingdao, China
b
University of Portsmouth, Portsmouth, UK
a r t i c l e i n f o
Article history:
Received 9 August 2017
Revised 23 January 2018
Accepted 12 March 2018
Available online 20 March 2018
Communicated by Dr Chenguang Yang
Keywords:
Parkinson’s disease
diagnosis
Gait
temporal data
LSTM
CNN
a b s t r a c t
When diagnosing Parkinson’s disease (PD), medical specialists normally assess several clinical manifesta-
tions of the PD patient and rate a severity level according to established criteria. This rating process is
highly depended by doctors’ expertise, which is subjective and inefficient. In this paper, we propose a
machine learning based method to automatically rate the PD severity from gait information, in particular,
the sequential data of Vertical Ground Reaction Force (VGRF) recorded by foot sensors. We developed a
two-channel model that combines Long Short-Term Memory (LSTM) and Convolutional Neural Network
(CNN) to learn the spatio-temporal patterns behind the gait data. The model was trained and tested on
three public VGRF datasets. Our proposed method outperforms existing ones in terms of prediction ac-
curacy of PD severity levels. We believe the quantitative evaluation provided by our method will benefit
clinical diagnosis of Parkinson’s disease.
©2018 Published by Elsevier B.V.
1.
Introduction
Parkinson’s disease (PD) is a degenerative brain disorder charac-
terized by a loss of midbrain dopamine (DA) neurons [1] . It affects
mainly elderly people, causing movement problems such as static
tremors, rigidity, bradykinesia, gait disturbance, and postural insta-
bility [2] . As apart of these motor symptoms, gait disturbance oc-
curs in early stages and shows obvious manifestations. Some gait
disturbances such as festinating gait, short gait, and freezing gait
have been used in literature to identify a prognosis of Parkinson’s
disease by characterizing gait analysis [3–5] .
When assessing the severity level of Parkinson disease, nu-
meric scales are preferred. The Hoehn & Yahr scale (H & Y scale)
was widely adopted, which consisted of 5 stages originally and
was further extended with additional stage 1.5 and 2.5 [6] . The
Unified Parkinsons Disease Rating Scale (UPDRS) is more com-
plex and consists of more levels [7] . When medical specialists em-
ploy these scales to rate the PD severity, subjectivity and low effi-
∗
Corresponding author.
E-mail addresses: zhaoaite@stu.ouc.edu.cn , junyu.dong@outlook.com (A. Zhao),
qilin@ouc.edu.cn (L. Qi), lijie@stu.ouc.edu.cn (J. Li), dongjunyu@ouc.edu.cn (J. Dong),
hui.yu@port.ac.uk (H. Yu).
1
The author contributed equally to this work and should be considered co-first
author.
ciency are inevitable as most of the diagnostic criteria use descrip-
tive symptoms, which cannot provide a quantified diagnostic ba-
sis. Therefore, the development of computer-assisted diagnosis and
computer-expert system is very important.
Modern computer-assisted diagnosis, such as automatic medical
image processing and large scale medical data analysis, has been
widely used in corresponding medical fields, where machine learn-
ing plays a core role in these systems.
For PD detection from gait data, machine learning methods,
such as kernel Fisher discriminant, naïve Bayesian, and support
vector machine, have been employed and achieved promising
results [8–15] . However, these approaches only deal with it as a
two-category classification problem, i.e. detecting PD from gait in-
formation, whereas the severity grading, which requires a finer in-
vestigation, has not been well studied.
In addition, the machine learning methods used in literature
are not specifically designed to deal with temporal sequential data,
whereas the gait data captured by sensory devices (cameras, force
sensors) contains important temporal information that is critical
for PD diagnosis. As one of the popular deep learning model,
Long Short-Term Memory (LSTM) has recently been used in var-
ious fields, including action recognition and gait recognition [16–
22]
as it is good at handling time series with long intervals and
multi-classification problems. Besides, the Convolution Neural Net-
works (CNN) can automatically learn commendable features from
the given gait data, which considers spatial proximity using a
https://doi.org/10.1016/j.neucom.2018.03.032
0925-2312/© 2018 Published by Elsevier B.V.