Neurodegenerative diseases categorization by applying the automatic model selection and hyperparameter optimization method

Abstract

Neurodegenerative diseases affect a large part of the population in the world and also in Mexico, deteriorating gradually the quality of patients’ life. Therefore, it is important to diagnose them with a high degree of reliability. In order to solve it, various computational methods have been applied in the analysis of biomarkers of human gait. In this study, we propose employing the automatic model selection and hyperparameter optimization method that has not been addressed before for this problem. Our results showed highly competitive percentages of correctly classified instances when discriminating binary and multiclass sets of neurodegenerative diseases: Parkinson’s disease, Huntington’s disease, and Spinocerebellar ataxias.

Keywords

Random forest categorization gait recognition biomarkers

1 Introduction

Neurodegenerative diseases are pathologies that affect the nervous system producing gradually different aspects of motor skills. Gait is one of the motor skills that gets severely affected. A major problem is that their diagnosis and categorization are difficult in the initial stages and they prevail in older adults [9, 30]. Their diagnosis in the early stages is an open problem in medicine. These diseases can be detected as patterns of behavior in the patient’s gait, since as shown in many studies: variations in gait are an early indicator [23 , 32].

Parkinson’s disease (PD) is a progressive disorder of the nervous system that affects movements particularly [13]. PD develops gradually and sometimes starts with a barely noticeable tremor in one upper extremity. In Mexico, this disease represents the main neurodegenerative disease that affects adult patients, and it is more common in men than in women [27, 29]. Huntington’s disease (HD) and spinocerebellar ataxias (SCA) are hereditary neurodegenerative disorders that affect motor coordination, and consequently, the gait and also lead to progressive functional deterioration. HD and SCA patients have a definite molecular diagnosis as the responsible autosomal dominant mutation with trinucleotide repeats was identified in their genome. The most frequent SCAs in Mexico are SCA2 and SCA3 (or Machado-Joseph’s disease). We chose this group of progressive neurological disorders because they represent the most frequent diseases, which cause movement disorders. These disorders affect gait similarly. They alter the balance, slow the gait velocity (bradykinesia) which produces shortening steps, and among other parameters of gait are affected [8, 25]. These diseases can be so subtle that can go undetected by an observer.

From a computational perspective, Gait recognition methods are well-established to provide gait biomarkers that can serve in the categorization of diseases such as PD, Alzheimer’s disease (AD), HD, SCA, among others [1 , 39]. In this work, we explore the use of these types of biomarkers for the classification of neurodegenerative diseases using a method of Machine Learning. The classification consists of predicting a label within a finite set of labels [28]. There are a wide variety of Machine learning classification algorithms that are based on: bayesian networks, artificial neural networks, fuzzy logic, decision trees, association rules, assembled methods, etc. [36]. In particular, in this study, we followed the method proposed in [20, 33]: Automatic model selection and hyperparameter optimization, to identify the appropriate classifier based on characteristics of the gait biomarkers. After exhaustive experimentation, we identified the Random Forest classification algorithm, as the one which better performs.

The rest of this paper is organized as follows: Section 2 depicts related works to this research, the materials and methods are detailed in Section 3, in Section 4 the experiments and results are shown and, finally, the results and possible future works are discussed in Section 5.

2 Related works

The categorization of neurodegenerative diseases based on gait recognition has had important advances. this section is divided into two subsections, the first focuses on studies of neurodegenerative diseases based on gait. The second subsection reviews works that are concerned with the gait classification.

2.1 Gait as an identifier of neurodegenerative diseases: a neurological approach

The following three studies suggest a relationship between gait disorders and neurodegenerative diseases:

In a group of 45 patients with SA and 15 control subjects, Nakamura et al. studied the relationship between gait and SA. Results revealed that as gait skills declined, the same happened with the mental abilities of patients. They established that the gait speed is significantly reduced as the severity of dementia symptoms increased [24].

Theill et al. studied the gait of 1,072 older adults. The tests revealed that as the mental decline known as mild cognitive impairment increased, the striding speed also decreased [32].

Mielke et al. performed evaluations on 1,478 patients. Tests of mental abilities and gait with each patient were carried out. The results showed that a higher gait velocity was associated with a lower cognitive decline [23].

These studies established a strong effect of neurodegenerative diseases on gait. In the next subsection, we review relevant works that focus on gait recognition with several purposes.

2.2 Classification of neurodegenerative diseases based on gait recognition

There are several works focused on PD. El Maachi et al. exposed that intelligent algorithms can reduce the subjectivity of gait analysis, so they applied a convolutional neural network to detect PD and predict the severity. Their results achieved 98.7% accuracy in detection and 85.3% in severity prediction [11]. Khan et al. proposed a method based on the analysis of movement in videos to recognize the gait in patients with PD. The model was based on the idea that a normal human body reaches equilibrium during gait by aligning the body’s posture with the gravity axis using feet as a base of support; In contrast, the posture of patients with PD seems to be leaning forward, as they are less able to align their body. To capture the gait data, 3 patients with PD and 4 control subjects were recorded while walking. Results showed a 100% recognition rate [18]. On the other hand, Barth et al. combined the analysis of the movement of hands and gait with a sensor-based system, for PD recognition. The experiments were carried out with 18 patients with PD and 17 control subjects. On the data, they did an exploration with four classifiers and obtained the highest percentage with AdaBoost, 97% [4]. Finally, Li et al. developed a gait recognition system based on a network of 16 sensors to identify gait patterns in patients with PD. They also developed an algorithm based on local linear embedding to extract and recognize gait characteristics. Results show that the proposed system has a recognition rate of around 95.57% [21]. In addition to the sensor networks, the images use has been proposed. Chien-Wen et al. proposed an image-based diagnostic system that uses an algorithm that combines principal component analysis with linear discriminant analysis. The study was done with 7 patients with PD and 7 control subjects. Results showed a recognition rate of 95.49% [7]. Finally, Barnes and Jafari designed a sensors network that was placed in such a way that apart from gait also recorded data from other movements of the body. The study showed that a sensor system quantitatively measures some of the factors involved in locomotion in real-time and also they proposes that it can be extended to detect SA [3].

Gait has also been related to HD. Grimbergen et al. using sensors, followed gait in 47 patients with HD, and 27 control subjects. The results revealed that concerning controls, patients with HD had a decrease in gait speed (1.15 m/s versus a 1.45 m/s, p < 0.001) and a decrease in stride length [14]. The clinical features included were lateral swaying, spontaneous knee flexion, variable cadence, and parkinsonian characteristics. The biomechanical analysis showed that gait oscillation varied in each foot with an average descent of speed, stride length, and cadence [19].

In ALS case also been proposed diagnostic methods have through gait. Sugavaneswaran et al. explored the concepts of Machine Learning-kernel functions by incorporating the ambiguity of time-frequency space. The proposed technique was evaluated with gait information (acquired by sensors) of 13 patients with ALS and 16 control subjects. A classification percentage of 93.1% was reached [31]. Similarly, Wu and Ng conducted a study on gait cadence (step intervals) in patients with ALS. Gait information was acquired by sensors. The probability density functions of step intervals were estimated with the non-parametric Parzen-window method. Results showed that stride patterns can be effectively distinguished between patients with ALS and healthy subjects with an accuracy rate of 82.8% [37].

So far, works have been directed to a single disease, but the discriminative power for different diseases has also been studied. Elden et al. presented a method for classifying PD, HD, ALS, and healthy control subjects. They implemented Fisher score in feature selection and Support Vector Machine (SVM) in classification. They obtained 95.31% of accuracy for discriminating neurodegenerative diseases versus healthy control [12]. Dutta et al. proposed an algorithm for the gait pathological classification of 15 patients with PD, 13 with ALS, 20 with HD, and 16 control subjects. Their technique consists of time extraction and frequency domain characteristics of correlograms obtained by cross-correlation of gait signals, subsequently, a previously trained Elman’s recurrent neuronal network is used. Results show a precision of 87.5% for PD, 88.9% for HD, and 83.3% for ALS in the problem of a set of neurodegenerative diseases (multi-class) [10]. Merory et al. did a statistical analysis of spatio-temporal motion characteristics in patients gait with dementia with Lewy body (LBD) and compared them with patients with AD and control subjects. In performed tests on a treadmill, it was observed that the stride length and velocity values were significantly reduced in both patients groups compared to the control group at different speeds. Significant correlations were found between gait speeds of patients with LBD and AD, but significantly different from the control group [22]. Verghese et al. present a broader study with gait statistical methods of 422 patients with different neurodegenerative pathologies. For the study, videos of walking patients were analyzed. As a result, it was found that the presence of gait alterations is a significant predictor of the risk of developing dementia, especially the called non-Alzheimer dementia [34]. Finally, Hausdorff et al. compared the gait pace of 11 patients with ALS versus control subjects and PD versus HD. The subjects walked for 5 minutes at their usual pace using sensors. In the study, it was found that the gait of patients with ALS is less stable and more disorganized temporarily compared with healthy people. Therefore, the step interval and gait fluctuations are apparently compromised with ALS [15].

Fig. 1

a) Sensor network, b) sensor network topology.

Other studies have focused on identifying the best Machine Learning classifier. Iram et al. divided neurological diseases into three stages: retrogenesis, cognitive impairment, and gait disorder. For tests, they used information acquired from sensors of 15 patients with PD, 13 with ALS, 20 with HD, and 16 control subjects. They evaluated 11 classifiers employing confusion matrices to determine the exact degree of disease. Results show a better performance with the Bayes quadratic classifier, 90% for PD, 50% for HD, and 50% for ALS in multi-class classification [16]. Similarly, Banaie et al. proposed new attributes starting from statistical processes on a dataset with gait data. For tests, they used data from 15 patients with PD, 13 with ALS, 20 with HD, and 16 control subjects. They evaluated 17 classifiers and the best performance was obtained with the Bayes quadratic classifier with percentages of 100% in control subjects, 71.429% in HD, 80% in PD, and 100% in ALS, in the problem of four classes [2]. Yang et al. implemented an SVM classifier to examine four types of attribute selection. Tests were performed on data from 15 patients with PD, 13 with ALS, 20 with HD, and 16 control subjects. Results showed that with a set of four attributes it is easier to distinguish PD from the control subjects, the HD, and ALS with percentages of 86.43%, 79.04%, and 85.47% respectively; ALS is better distinguished than the control subjects and HD with 93.96% and 86.52%; likewise, HD is better distinguished from the control subjects with 84.17% [39]. Xia et al. applied the Leave-One-Out Cross-Validation method obtaining 96.83% of accuracy [38]. Ren et al. studied the Empirical Mode Decomposition method for decomposing the time series of gait rhythms into intrinsic mode functions. Their general values of AUC got a good performance in binary classification [26]. Bilgin through wavelet function “bior2.6” decomposed the compound force signal for determination of features and applying Näive Bayesian classifier achieved 90.93% of accuracy in distinguishing ALS from the control subjects [5].

Based on these studies, it is evident that there is no method that allows obtaining very acceptable results in both classification cases: binary and multiclass. On the other hand, the use of a sensor network allows obtaining biomarkers by means of sensors placed on the lower extremities. In addition, to our knowledge, there is no database with gait information of a group of patients with neurodegenerative diseases in Mexico, so this study is considered valuable, in the sense that other researchers can carry out experiments from the database. Finally, the automatic model selection and hyperparameter optimization method has not yet been used to address a medical problem such as the one proposed in this research.

3 Material and methods

3.1 Topology of sensor network

To design the sensor network five ADXL-335 3-axis accelerometers were used, wired on an Arduino MEGA-2560 card, these accelerometers cover the extremities of the knees (left and right), ankles (left and right), and the chest (Fig. 1). From the sensor network, the information of the cartesian axes x, y, and z were obtained. The sensor network is low-cost.

3.2 Dataset

An ethics committee of the National Institute of Neurology and Neurosurgery (NINN) approved the implementation of a gait laboratory inside NINN to obtain gait biomarkers of patients suffering neurodegenerative diseases. On the other hand, each patient with his family member signed consent reports authorizing the public availability of data for studies with scientific purposes, as long as the anonymity of patients was protected. The population was 82 patients suffering neurodegenerative diseases: 47 with PD, 13 affected with HD, and 22 with SA; of the total population were 48 men and 34 women. In addition, gait information from 19 control subjects was collected, 7 men and 12 women. The age and gender distributions of patients are shown in Table 1. The estimated time per person during the capture process was two minutes. The exclusion criteria were patients who used a cane, who was helped to walk by their caregivers, or who used wheelchairs.

Table 1
Age and sex distribution in the NINN-Database

Case Age groups Sex

18-29 30-39 40-49 50-59 60-69 70-84 F M

Control 3 2 4 5 4 1 12 7

PD 2 2 12 12 19 17 30

HD 3 2 3 4 1 5 8

SA 5 4 6 7 11 11

Case	Age groups	Sex
Control	3	2	4	5	4	1	12	7
PD		2	2	12	12	19	17	30
HD	3	2	3	4	1		5	8
SA	5	4	6	7			11	11

The first step to build datasets was grouping the information by disease, i.e, having a file for each patient, we proceeded to unite all the files by disease (including control subjects files), which resulted in four datasets which to be processed in the classification phase to records level and with the raw data 1 . Having the data grouped, we proceeded to construct the datasets in two categories:

Binary sets: {control, SA}, {control, HD}, {control, PD}, {SA, HD}, {SA, PD}, {PD, HD},

Multiclass sets: {SA, HD, PD}, {control, SA, HD, PD}.

3.3 Automatic model selection and hyperparameter optimization method

The automatic model selection and hyperparameter optimization method is used to select the classification algorithm [20, 33]: Given a set of algorithms $A$ and a limited amount of training data $D = {(x_{1}, y_{1}), . . ., (x_{n}, y_{n})}$ , the objective is to determine algorithm $A^{*} \in A$ with an optimal generalization that is estimated by dividing $D$ into training sets and disjoint validation D $D_{train}^{(i)}$ and $D_{valid}^{(i)}$ for i = 1, . . . , k and then the f_i learning function, applying A^* to $D_{train}^{(i)}$ , and evaluating the predictive performance of these functions in $D_{valid}^{(i)}$ . This allows the selection problem to be written as: $A^{*} \in \underset{A \in A}{argmin} \frac{1}{k} . \sum_{i = 1}^{k} L (A, D_{train}^{(i)}, D_{valid}^{(i)}),$ (1)

Where $L (A, D_{train}^{(i)}, D_{valid}^{(i)})$ is the loss (here: misclassification rate) achieved by A when training in $D_{train}^{(i)}$ and it is evaluated in $D_{valid}^{(i)}$ . Cross-validation is used to divide training data into k partitions of equal size $D_{valid}^{(1)}, . . ., D_{valid}^{(k)}$ , and it is established $D_{train}^{(i)} = D \ D_{valid}^{(i)}$ for i = 1, . . . , k.

Based on the previous criterion, it was determined that the Random Forest algorithm was appropriate for classification. The algorithm, which belongs to the assembled methods, consists of a collection of classifiers structured by tree {h (x, Θ_k) , k = 1, . . .} where Θ_k are independent random vectors distributed identically and each tree throws a unitary vote for the most popular class at the input x [6]. The algorithm is shown below:

Algorithm 1. Random Forest

Data: dataset T = (x, y), number of trees m,

number of random levels k

Result: RF, a set of grown trees

Initialization RF

fori = 1 to mdo

T’ ← boostrap(T)

Tree ← trainDT (T’k)

add Tree to RF

end

In this technique, the inputs are the data (T) which are previous recordings (x) with their corresponding diagnosis (y), the number of decision trees to generate (m), and the maximum number of levels that each tree will have (k). For each tree, a new set (T′) of training data is created through bootstrapping (i.e., sampling with replacement). Once each data is selected a tree is selected by randomly selecting two features and choosing the best partition among both features, this is repeated until all the features have been analyzed or it reaches the maximum number of levels [6].

3.4 Metrics

To assess the performance, confusion matrix (see Table 2), true positive rate or recall (Equation 2), false positive rate (Equation 3), precision (Equation 4), f-measure (Equation 5), and ROC area were used as evaluation metrics. In the confusion matrix in each column the number of predictions of each class is represented and each row represents the real values. The parameters used are true positive (TP), true negative (TN), false positive (FP), and false negative (FN). The confusion matrix allows calculating the true positive rate as the classification of correctly classified instances. On the other hand, the FP rate represents the rate of incorrectly classified negative instances. The precision gets the fraction of instances classified in the positive class that is in fact, positive class. F-measure captures the properties of TP rate and precision into a single measure. Finally, the ROC curve represents the true positive rate versus the false positive rate.

Table 2
Confusion matrix

True class

Positive Negative

Predicted Positive TP FP

Class Negative FN TN

		True class
Predicted	Positive	TP	FP
Class	Negative	FN	TN

TPrate = \frac{TP}{TP + FN}

(2)

FPrate = \frac{FP}{FP + TN}

(3)

Precision = \frac{TP}{TP + FP}

(4)

F - measure = \frac{2 \cdot precision \cdot recall}{precision + recall}

(5)

4 Experiments and results

For binary sets, the percentages of correctly classified instances shown in Table 3 were obtained. On the other hand, for multiclass sets, the percentages of correctly classified instances are shown in Tables 4 –6.

Table 3
Classification results of binary sets using Random Forest classifier and their respective confusion matrix

Subsets % correctly classified Confusion matrix

{control,SA} 98.8547 a b

control = a 2529 21

SA = b 46 3254

{control,HD} 98.8444 a b

control = a 36 14

HD = b 38 1912

{control,PD} 99.3036 a b

control = a 39 11

PD = b 36 4163

{SA,HD} 98.7619 a b

SA = a 3296 4

HD = b 61 1889

{SA,PD} 99.5999 a b

SA = a 3297 3

PD = b 27 4172

{HD,PD} 99.4959 a b

HD = a 1939 11

PD = b 20 4179

Subsets	% correctly classified	Confusion matrix
{control,SA}	98.8547		a	b
		control = a	2529	21
		SA = b	46	3254
{control,HD}	98.8444		a	b
		control = a	36	14
		HD = b	38	1912
{control,PD}	99.3036		a	b
		control = a	39	11
		PD = b	36	4163
{SA,HD}	98.7619		a	b
		SA = a	3296	4
		HD = b	61	1889
{SA,PD}	99.5999		a	b
		SA = a	3297	3
		PD = b	27	4172
{HD,PD}	99.4959		a	b
		HD = a	1939	11
		PD = b	20	4179

Table 4

Classification results of multiclass sets using Random Forest classifier

Subsets	% correctly classified
{SA,HD,PD}	99.3227
{control,SA,HD,PD}	99.1166

Table 5

Confusion matrix of multiclass set {SA,HD,PD}

Classified as	a	b	c
SA = a	3283	8	9
HD = b	19	1926	5
PD = c	15	8	4176

Table 6

Confusion matrix of multiclass set {control,SA,HD,PD}

Classified as	a	b	c	d
Control = a	2534	7	0	9
SA = b	7	3278	3	12
HD = c	15	15	1913	7
PD = a	11	13	7	4168

It is observed in a general way (Table 3) that Random Forest generates classifications results highly competitive, around 99% . In particular: sick against healthy, confused with healthy 46 of 3300 with SA, also confused with healthy 38 of 1950 with HD and confused with healthy 36 of 4163 with PD; sick against sick, confused 65 of a total of 5250 with SA and PD, confused 30 out of a total of 7499 with SA and PD and confused 31 of a total of 6149 with HD and PD.

It is observed that Random Forest generates, again, highly competitive results, very close to 100% . When considering four classes we can see in the confusion matrix (Table 5) that numbers confused are very low when classifying the instances. In addition, the statistical metrics of Table 7 confirm in detail by class, highly competitive values.

Table 7

Detailed metrics by class

	TP Rate	FP Rate	Precision	F-Measure	ROC Area	Class
	0.994	0.003	0.987	0.990	1.000	Control
	0.993	0.004	0.989	0.991	1.000	SA
	0.981	0.001	0.995	0.988	1.000	HD
	0.993	0.004	0.993	0.993	1.000	PD
Weighted Avg.	0.991	0.003	0.991	0.991	1.000

The experiments were performed in Waikato Environment Knowledge Analysis v.3.8 [35], on a Lenovo ideapad 300 laptop, with Windows 10 of 64 Bits, Intel Celeron @ 1.60GHz processor, and RAM of 4.00 GB.

4.1 Discussion

The automatic model selection and hyperparameter optimization method after exhaustive experimentation, it allowed determining that the Random Forest algorithm is appropriate for the binary ({control,SA}, {control,HD}, {control,PD}, {SA,HD}, {SA,PD}, {HD,PD}) and multiclass ({SA,HD,PD}, {control,SA,HD,PD}) classification of neurodegenerative diseases. Random Forest showed highly competitive results with a true positive rate, precision, f-measure, and ROC area around 99% . In contrast, the false positive rate was less than 0.005. The results obtained are highly reliable, which makes them suitable for implementation in a computer-aided diagnostic system.

5 Conclusions and future work

Our results showed the effectiveness of Random Forest algorithm for the classification of neurodegenerative diseases based on gait biomarkers. It is evident that when classifying binary and multiclass sets using the criterion proposed by [20, 33] to select the classification algorithm, results close to the optimal 100% are obtained, this is at least corroborated with the NINN-Database and considering the raw data.

Results of this study were obtained from patients suffering from neurodegenerative diseases in different stages, therefore, it is necessary to acquire gait information of patients in early stages.

As a future challenge of this research is considered to develop software so that neurologists can have a second opinion and sharpen the diagnosis as this may be biased due to external factors that directly affect the specialist, such as visual fatigue, emotional tiredness, or other factors of psychological type.

As future work, we propose:

Implement deep learning algorithms in the NINN-Database, in binary and multiclass data sets, to compare and analyze the results.

Use other sensors: gyroscopes or floor sensors, but maintain a low-cost solution. This is to have a greater information spectrum and be able to implement attribute selection algorithms.

Since other diseases present gait disorders, we considered extending the study to others, for example, diabetic neuropathy.

Footnotes

Following the spirit that motivates scientific research, and endorsed by the NINN and, above all, by patients and their relatives, we have proceeded to release the database so that any researcher who wishes to can access to the information:

References

Banaie

, Pooyan

and Mikaili

, Introduction and application ofan automatic gait recognition method to diagnose movement disordersthat arose of similar causes, Expert Systems with Applications38(6) (2011), 7359–7363. ISSN 0957-4174. doi: 10.1016/j.eswa.2010.12.091. URLhttps://dx-doi-org.web.bisu.edu.cn/10.1016/j.eswa.2010.12.091.

Banaie

, Pooyan

and Mikaili

, Introduction and application ofan automatic gait recognition method to diagnose movement disordersthat arose of similar causes, Expert Systems with Applications38(6) (2011), 7359–7363.

Barnes

and Jafari

, Locomotion monitoring using body sensor networks. In Proceedings of the 1st international conference on PErvasive Technologies Related to Assistive Environments (2008), pp. 47. ACM.

Barth

, Sünkel

, Bergner

, Schickhuber

, Winkler

, Klucken

and Eskofier

, Combined analysis of sensor data from hand and gait motor function improves automatic recognition of parkinson’s disease. In 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (2012), pp. 5122–5125. IEEE.

Bilgin

, The impact of feature extraction for the classificationof amyotrophic lateral sclerosis among neurodegenerative diseasesand healthy subjects, Biomedical Signal Processing and Control31 (2017), 288–294.

Breiman

, Random forests, Machine Learning45(1) (2001), 5–32.

Cho

C.-W.

, Chao

W.-H.

, Lin

S.-H.

and Chen

Y.-Y.

, A vision-basedanalysis system for gait recognition in patients with parkinson’sdisease, Expert Systems with Applications36(3) (2009), 7033–7039.

Mourão do Amaral Coutinho

M.P.

, Doença de Machado-Joseph: tentativa de definição. PhD thesis, Universidade do Porto. Reitoria, 1992.

Dawson

, Hendershot

and Fulton

, Aging in the eighties:functional limitations of individuals age 65 years and over, Advance data (133) (1987), 1–12.

10.

Dutta

, Chatterjee

and Munshi

, Hybrid correlation neural network synergy for gait signal classification. In Advances in Heuristic Signal Processing and Applications, (2013), pp. 263–285. Springer.

11.

El Maachi

, Bilodeau

G.-A.

and Bouachir

, Deep 1d-convnet foraccurate parkinson disease detection and severity prediction fromgait, Expert Systems with Applications143 (2020), 113075.

12.

Hossam Elden

, Al-Atabany

and Fathi Ghoneim

, Gait rhythm fluctuations assessment for neurodegenerative patients. In 2018 9th Cairo International Biomedical Engineering Conference (CIBEC) (2018), pp. 9–12. IEEE.

13.

Gaba

, Recent studies on nutrition and parkinson’s diseaseprevention: a systematic review, Open Journal of PreventiveMedicine5 (2015), 197–205.

14.

Grimbergen

Y.A.M.

, Knol

M.J.

, Bloem

B.R.

, Kremer

B.P.H.

, Roosand

R.A.C.

and Munneke

, Falls and gait disturbances in huntington’s disease, Movement Disorders23(7) (2008), 970–976.

15.

Hausdorff

J.M.

, Lertratanakul

, Cudkowicz

M.E.

, Peterson

A.L.

, Kaliton

and Goldberger

A.L.

, Dynamic markers of altered gait rhythmin amyotrophic lateral sclerosis, Journal of Applied Physiology88(6) (2000), 2045–2053.

16.

Iram

, Al-Jumeily

, Fergus

, Randles

and Hussain

, Computational data analysis for movement signals based on statistical pattern recognition techniques for neurodegenerative diseases. In 13th Annual Post Graduate Symposium on the Convergence of Telecommunications, Networking and Broadcasting, (2012), pp. 1–6.

17.

Iram

, Al-Jumeily

, Fergus

, Randles

and Hussain

, Computational Data Analysis for Movement Signals Based on Statistical Pattern Recognition Techniques for Neurodegenerative Diseases. In Proceedings of the 13th Annual Post Graduate Symposium on the Convergence of Telecommunications, Networking and Broadcasting (PGNet2012), 2012. ISBN 978-1-902560-26-7.

18.

Khan

, Westin

and Dougherty

, Motion cue analysis for parkinsonian gait recognition, The Open Biomedical Engineering Journal7 (2013), 1.

19.

Koller

W.C.

and Trimble

, The gait abnormality of huntington’sdisease, Neurology35(10) (1985), 1450–1450.

20.

Kotthoff

, Thornton

, Hoos

H.H.

, Hutter

and Leyton-Brown

, Auto-weka 2.0: Automatic model selection and hyperparameteroptimization in weka, Journal of Machine Learning Research17 (2016), 1–5.

21.

, Wang

and Wang

, A novel gait recognition analysis systembased on body sensor networks for patients with parkinson’s disease, International Journal of Communication Networks and DistributedSystems7(3-4) (2011), 262–274.

22.

Merory

J.R.

, Wittwer

J.E.

, Rowe

C.C.

and Webster

K.E.

, Quantitativegait analysis in patients with dementia with lewy bodies andalzheimer’s disease, Gait & Posture26(3) (2007), 414–419.

23.

Mielke

M.M.

, Roberts

R.O.

, Savica

, Cha

, Christianson

, Pankratz

V.S.

, Geda

Y.E.

, Machulda

M.M.

, Ivnik

R.J.

, Drubach

D.I.

, et al., Assessing the temporal relationship between cognition andgait: slow gait predicts cognitive decline in the mayo clinic studyof aging, Journals of Gerontology Series A: Biomedical Sciencesand Medical Sciences68(8) (2012), 929–937.

24.

Nakamura

, Meguro

, Yamazaki

, Okuzumi

, Tanaka

, Horikawa

, Yamaguchi

, Katsuyama

, Nakano

, Arai

, et al., Postural and gait disturbance correlated with decreased frontalcerebral blood flow in alzheimer disease, Alzheimer Disease andAssociated Disorders11(3) (1997), 132–139.

25.

Paulson

H.L.

, Dominantly inherited ataxias: lessons learned from machado-joseph disease/spinocerebellar ataxia type 3. In Seminars in Neurology27 pp. 133–142. Copyright© 2007 by Thieme Medical Publishers, Inc., 333 Seventh Avenue, New..., 2007.

26.

Ren

, Tang

, Fang

, Luo

, Xu

, Bringas-Vega

M.L.

, Yao

, Kendrick

K.M.

and Valdes-Sosa

P.A.

, Gait rhythm fluctuation analysisfor neurodegenerative diseases by empirical mode decomposition, IEEE Transactions on Biomedical Engineering64(1) (2017), 52–60.

27.

Rodríguez-Violante

, Velásquez-Pérez

and Cervantes-Arriaga

, Incidence rates of parkinson’s disease in mexico:Analysis of -statistics, Revista Mexicana deNeurociencia20(3) (2019), 136–140.

28.

Shalev-Shwartz

and Ben-David

, Understanding machine learning: From theory to algorithms. Cambridge university press, 2014.

29.

Smith

E.R.

, Perrin

P.B.

, Tyler

C.M.

, Lageman

S.K.

and Villaseñor

, Cross-cultural differences in parkinson’s diseasecaregiving and burden between the united states and mexico, Brain and Behavior10(9) (2020), e01753.

30.

Sudarsky

, Geriatrics: gait disorders in the elderly, The NewEngland Journal of Medicine322(20) (1990), 1441–1446.

31.

Sugavaneswaran

, Umapathy

and Krishnan

, Discriminative time-frequency kernels for gait analysis for amyotrophic lateral sclerosis. In Engineering in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the IEEE, (2011), pp. 2683–2686. IEEE.

32.

Theill

, Martin

, Schumacher

, Bridenbaugh

S.A.

and Kressig

R.W.

, Simultaneously measuring gait and cognitive performance incognitively healthy and cognitively impaired older adults: The basel motor–cognition dual-task paradigm, Journal of theAmerican Geriatrics Society59(6) (2011), 1012–1018.

33.

Thornton

, Hutter

, Hoos

H.H.

and Leyton-Brown

, Auto-weka: Combined selection and hyperparameter optimization of classification algorithms. In Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining (2013), pp. 847–855. ACM.

34.

Verghese

, Lipton

R.B.

, Hall

C.B.

, Kuslansky

, Katz

M.J.

and Buschke

, Abnormality of gait as a predictor of non-alzheimer’s dementia, New England Journal of Medicine347(22) (2002), 1761–1768.

35.

Waikato. Weka software. Available from: http://www.cs.waikato.ac.nz/ml/weka/. Online; accessed 2017.

36.

Witten

I.H.

, Frank

, Hall

M.A.

and Pal

C.J.

, Data Mining: Practical machine learning tools and techniques, Morgan Kaufmann 2016.

37.

and Chun Ng

, A pdf-based classification of gait cadence patterns in patients with amyotrophic lateral sclerosis. In Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE (2010), pp. 1304–1307. IEEE.

38.

Xia

, Gao

and Ye

, Classification of gait rhythm signals between patients with neuro-degenerative disdiseases and normal subjects: Experiments with statistical features and different classification models, Biomedical Signal Processing and Control18 (2015), 254–262.

39.

Yang

, Zheng