See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/350527156

Ηyper-sinh-Convolutional Neural Network for Early Detection of Parkinson’s

Disease from Spiral Drawings

Article in WSEAS TRANSACTIONS ON COMPUTER RESEARCH · March 2021

DOI: 10.37394/232018.2021.9.1

CITATIONS READS

13 496

4 authors, including:

Luca Parisi Renfei Ma

citeline Shenzhen MSU-BIT University
59 PUBLICATIONS 492 CITATIONS 24 PUBLICATIONS 382 CITATIONS

SEE PROFILE SEE PROFILE

Mansour Youseffi
University of Bradford
154 PUBLICATIONS 1,219 CITATIONS

SEE PROFILE

All content following this page was uploaded by Luca Parisi on 03 April 2021.

The user has requested enhancement of the downloaded file.

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

Ηyper-sinh-Convolutional Neural Network for Early Detection of Parkinson’s

Disease from Spiral Drawings

LUCA PARISI, RENFEI MA,

Department of Computer Science, University of Auckland Rehabilitative Technologies
Faculty of Engineering and Informatics, Association (UARTA)
University of Bradford University of Auckland
Bradford, BD7 1DP, United Kingdom Auckland, 1010, NEW ZEALAND
University of Auckland Rehabilitative Technologies [email protected]
Association (UARTA)
University of Auckland
Auckland, 1010, NEW ZEALAND
[email protected]

AMIR ZAERNIA, MANSOUR YOUSEFFI,

Department of Biomedical and Electronics Engineering, Department of Biomedical and Electronics Engineering,
Faculty of Engineering and Informatics, Faculty of Engineering and Informatics,
University of Bradford University of Bradford
Bradford, BD7 1DP, UNITED KINGDOM Bradford, BD7 1DP, UNITED KINGDOM
[email protected] [email protected]

Abstract—Modern developments in the state-of-the-art open-source activation functions for Convolutional Neural
Networks (CNNs) have broadened the selection of benchmark activations for Deep Learning (DL)-aided
classification. Nevertheless, achieving discrimination of non-linear input image data in CNN is still not
straightforward and it is unclear how such novel activation functions can have translational applications with
tangible impact. hyper-sinh, made freely available in TensorFlow and Keras, was demonstrated as a benchmark
activation function on five (N=5) datasets in its ground-breaking paper. Measuring the value from deploying this
activation in a specific application is pivotal to supply the required evidence of its performance on real-life
supervised DL-based image classification tasks. In this study, a CNN was for the first time combined with hyper-
sinh to aid early detection of Parkinson’s Disease (PD) from discriminating pathophysiological patterns extracted
from spiral drawings. Thus, the hyper-sinh activation was deployed to maximise the separability of the input
features from spiral drawings via automated pattern recognition. We demonstrate the accuracy and reliability of
hyper-sinh-CNN to aid early diagnosis of PD, evaluated against other gold standard activation functions, including
the recent Quantum ReLU (QReLU) and the modified Quantum ReLU (m-QReLU) that solved the ‘dying ReLU’
problem for the first time in the literature of DL. Two (N=2) benchmark datasets from the database of the Botucatu
Medical School, São Paulo State University in Brazil, scaled to be in 28 by 28 pixels as the MNIST benchmark
data, were used to discriminate between input image patterns of 158 subjects (53 healthy controls and 105 patients
with PD) from spirals drawn on graphics tablets. Overtraining was avoided via early stopping and the models were
developed and tested in TensorFlow and Keras (Python 3.6). The supervised model (hyper-sinh-CNN) could detect
early Parkinson’s Disease with 81% and 91% classification accuracy from the two datasets respectively (F1-scores:
73% and 91% correspondingly). Furthermore, the model achieved high sensitivity (81% and 91%). Thus, this study
validates the application of hyper-sinh to aid real-life supervised DL-based image classification, in particular early
diagnosis of PD from spiral drawings.

Keywords: Convolutional Neural Network; Parkinson’s Disease; m-arcsinh; hyper-sinh; QReLU; m-QReLU.
Received: February 12, 2021. Revised: March 3, 2021. Accepted: March 19, 2021. Published: March 31, 2021.

E-ISSN: 2415-1521 1 Volume 9, 2021

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

In these scenarios, failure of the activation functions in

the CNN to converge may lead to misclassifications and,
1. Introduction thus, result in misdiagnoses when deployed in a clinical
Millions of patients live with Parkinson’s disease (PD) setting. Convergence during training can be aided by
worldwide, which is a long-term neurodegenerative selecting an optimal activation function to facilitate
disorder resulting from damage to neurons that synthesise learning from non-linear data to ensure they can be linearly
dopamine. As PD develops, tremor in one hand and bodily separable [2, 7, 10-14]. As the traditional activation
stiffness may be experienced; nevertheless, the main four function involved in the training of CNN, i.e., the Rectified
symptoms are: tremor, stiffness in the limbs, decreased Linear Unit (‘ReLU’), is affected by the ‘dying ReLU’
movement and coordination, difficulty with fine problem [15, 16], which, despite appropriate He
movements to accomplish the activities of daily living, initialisation [17], is characterised by negative summed
such as eating and drinking. Whilst the underlying inputs with large weight updates. Recently, the Quantum
aetiology of PD remains unknown, there is no consensus ReLU (‘QReLU’) and modified Quantum ReLU (‘m-
on which biomarkers could be leveraged to provide a QReLU’) were developed in TensorFlow and Keras [18]
definite diagnosis of PD. To improve function, alternate and validated against nine ReLU-based AFs on seven
treatments, such as exercise and full body vibration using benchmark datasets as the first activation functions to
painless vibrating machines to stabilise multiple muscles resolve this long-standing issue in the literature of DL [16].
and posture have been investigated with promising results In fact, both the QReLU and the m-QReLU were
[1]. Advanced computational techniques have leveraged demonstrated to aid complex medical diagnostics, such as
feature-driven Machine Learning (ML) to accomplish early detecting COVID-19 from ultrasound images and early PD
detection of PD from patterns used as indicators of its from spiral drawings [16].
related speech impairments [2]. As an alternative approach, the modified arcsinh (‘m-
Impaired handwriting in PD patients may appear years arcsinh’) developed as a generalised kernel and activation
before a clinical diagnosis is made and, thus, could be the function in scikit-learn [19] was validated on 15
first signs of possible PD [3]. By assessing handwriting via benchmark datasets from scikit-learn and the University
a digitiser, which measures the mean pressure and mean California Irvine (UCI) ML repository [10]. Building on
velocity, as well as the spatial and temporal characteristics the m-arcsinh, the hyperbolic sinh or ‘hyper-sinh’ was
of each stroke, previous research showed that 97.5% of the developed to generalise its formulation to be accurate and
patients at risk of PD had significantly different writing reliable from shallow to deep neural networks in
with smaller letters (‘micrographia’), less force being TensorFlow and Keras [20] and validated on 5 benchmark
applied and a significantly longer writing time [4]. More datasets, involving both image and text classification tasks
recent studies have stated the impairment in writing can be [7].
analysed easier via a computerised analysis of spirals [5, This study’s novelty lies in assessing the replicability
6]. of the hyper-sinh in a clinical application, along with
The development of the open-source kernel function assessing its performance when coupled with a CNN
hyperbolic sinh or ‘hyper-sinh’ has provided a further against gold standard activation functions, including the
benchmark kernel for Deep Learning (DL)-aided recent QReLU and m-QReLU [16]. In this research work,
classification [7]. When coupled with the shallow ‘Fully- such a supervised DL technique was leveraged to perform
Connected Neural Network’ (FC-NN) and the deep a classification task that involves noisy data with
‘Convolutional Neural Network’ (CNN), its reliability and overlapping features, i.e., image data of spirals drawn on a
computational efficiency have been demonstrated on five graphics tablet to discern patterns as indicators of healthy
benchmark datasets for aiding both image and text and PD [16], to aid diagnosis of PD from the underlying
classification tasks [7]. However, besides its image-based patterns associated with its characteristic
reproducibility, additional studies need to verify its subtle tremor at early stages.
replicability for other applications. In fact, CNN may be
unable to achieve linear separability of noisy image data,
due to activation functions incurring vanishing [8] or
exploding [9] gradients leading to poor or lack of
convergence.

E-ISSN: 2415-1521 2 Volume 9, 2021

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

image data based on the activation function deployed [16,

2. Automated detection of 20, 23]. These input data must be labelled, since the CNN
parkinson’s disease from image is a supervised DL-based learning classifier, thus the class
which each input image belongs to is preliminarily known.
data
The choice of which activation function to use within
convolutional layers impacts the accuracy and reliability of
2.1 Data and Software for Modelling the classification of such input data into the target classes
[16, 20, 24]. Thus, in this study, CNN was assessed with
various activation functions, including the recently
To demonstrate the accuracy and reliability of early developed hyper-sinh [7], which scales from shallow to
detection of PD from image patterns from spiral drawings deep neural networks, such as the CNN.
on graphics tablets via the proposed algorithm, the
To avoid overtraining or overfitting, thus preventing the
following two datasets with image features from 53 healthy
model from ‘remembering’ the features seen during
subjects and 105 patients with PD were obtained from the
training and enabling learning that yields generalisation,
database of the Botucatu Medical School, São Paulo State
early stopping was applied [23, 24]. The Adam optimiser
University in Brazil:
was used for training, whilst the loss or probability error to
 Spiral HandPD dataset (data format: .jpg) [21] on minimise was the sparse softmax cross-entropy between
92 subjects (74 PD patients, 18 healthy controls). logits and labels, as suitable in the discrete classification
 NewHandPD dataset (data format: .jpg) [22] on task of interest in which the classes (healthy and PD-related
66 additional subjects (31 PD patients, 35 healthy spiral drawings) are mutually exclusive (each image is
controls). either characteristic of healthy or PD-related patterns).

Both datasets were represented as one-dimensional Given its ability to cope with non-linearly separable
(1D) vectors of 784 features for each image (28*28 pixels); data, the hyper-sinh (Fig. 1) activation function was
training and testing datasets were used as per the coupled with CNN and evaluated against other gold
proportions already established from the two benchmark standard activations (ReLU, Leaky ReLU, sigmoid, tanh,
datasets used. QReLU and m-QReLU [16]).

The data were reshaped to match the picture format The equations for the hyper-sinh activation function for
(height * width * channel) and then, to a four-dimensional positive values of input (x) data, and negative values and
(4D) tensor input (batch size, height, width, channel). The zero, are available in (1) and (2) respectively.
image datasets underwent min-max normalisation to
prepare them for the subsequent CNN-based classification.
sinh 𝑥 (1)
The CNN model was developed and evaluated via the 𝑦 = 𝑖𝑓 x > 0 ,
Keras library in Python, whilst the hyper-sinh [7]
3
activation function was developed in both TensorFlow and 𝑤ℎ𝑒𝑟𝑒 𝑥 𝑖𝑠 𝑡ℎ𝑒 𝑖𝑛𝑝𝑢𝑡 𝑑𝑎𝑡𝑎
Keras [20]. All other activation functions evaluated
(ReLU, Leaky ReLU, sigmoid and tanh) are available in
Keras, except for the custom QReLU and m-QReLU [16]
also developed and available in both TensorFlow and
Keras [18]. 𝑥3
𝑦 = 4
𝑖𝑓 x ≤ 0 (2)

2.2 hyper-sinh-based CNN Algorithm A flattening layer was applied to flatten the input data
to a one-dimensional (1D) vector for the fully connected
layer. The output layer had two neurons, as per the number
CNN is a deep neural network-based classifier that
of classes to discriminate (healthy vs PD-related image
leverages convolutional layers and filters to achieve
patterns). In Table I, these and other hyperparameters used
automated pattern recognition without the need for
in the CNN model tested in this study are listed, which
extrinsic feature engineering and separability of the input

E-ISSN: 2415-1521 3 Volume 9, 2021

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

were leveraged to classify both benchmark datasets, In binary classification problems, e.g., discriminating
regardless of the activation functions evaluated. between healthy and PD-related image patterns, classes can
be defined as ‘true’ or ‘false’, resulting in four scenarios
TABLE I. LEARNING HYPERPARAMETERS FOR THE CNN
MODEL USED FOR BOTH BENCHMARK DATASETS AND (Table II):
REGARDLESS OF THE ACTIVATION FUNCTIONS DEPLOYED.

Hyperparameter Value
 true positive (TP), the occurrence of
Learning rate 0.001 pathophysiological image patterns correctly
Maximum number of epochs or 15
detected (e.g., due to PD).
training iterations  true negative (TN), physiological/healthy speech
image correctly classified.
Batch size 128
 false positive (FP), physiological/healthy image
Dropout rate prior to output layer 0.25 patterns misclassified as if they were
pathophysiological instead.
Number of convolutional layers 2
 false negative (FN), the occurrence of a
Number of max pooling (down- 2 pathophysiological image patterns misclassified
sampling) layers as if they were physiological/healthy instead.
Number of convolutional filters in first 32
convolutional layer
TABLE II. CONFUSION MATRICES RELATING THE ACTUAL
Kernel size of first convolutional layer 5
AND PREDICTED IMAGE PATTERNS AS TO WHETHER THEY
Number of convolutional filters in 64 ARE INDICATORS OF PATHOPHYSIOLOGICAL TREMOR (E.G.,
second convolutional layer DUE TO EARLY PARKINSON’S DISEASE).

Kernel size of second convolutional 3 Actual

layer
True False
Number of strides in max pooling layer 2
(pathophysiological (physiological
Kernel size of max pooling layer 2 image patterns) image patterns)

Number of neurons in fully 1,024 Positive

connected/dense layer TP FP
(pathophysiological
Predicted

Number of neurons in final output layer 2 image patterns)

Negative

(physiological FN TN
2.3 Performance Measures to Validate image patterns)
Model Accuracy and Reliability

Besides the classification accuracy, the classification

The classification performance of the CNN was performance of the CNN with the various activation
evaluated quantitatively based on the trade-off between functions tested were assessed via two further performance
accuracy and reliability on the test sets [7, 10-14, 16, 20] metrics as indicators of reliability, i.e., precision and recall
whilst applying early stopping to avoid overfitting or (Tables III and IV).
overtraining; the latter metric was assessed via multiple
quantifiers, i.e., precision, recall and F1-score as their As a more comprehensive quantifier of the model
harmonic mean, thus representing their trade-off. reliability, the F1-score was computed too (Tables III and
IV), which corroborates or disproves the test classification-
related outcomes.

E-ISSN: 2415-1521 4 Volume 9, 2021

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

a) b)

Figure 1. The hyper-sinh activation function for positive values of input data (a)) and for negative values or zero of input data (b)).

(0.80-0.88) (0.85-0.90) (0.82- (0.83-

0.87) 0.88)
III. HYPER-SINH WITH A CONVOLUTIONAL NEURAL
Tanh 0.82 0.92 0.82 0.86
NETWORK FOR SUPERVISED DEEP LEARNING-BASED
AUTOMATED DETECTION OF PARKINSON’S DISEASE (0.81-0.85) (0.90-0.94) (0.79- (0.82-
0.84) 0.88)

QReLU 0.88 0.95 0.88 0.90

A quantitative analysis of classification accuracy and
(0.87-0.90) (0.93-0.97) (0.87- (0.88-
reliability was performed for automated early detection of
0.90) 0.92)
PD from image patterns from spiral drawings when
considering both benchmark datasets (Tables III and IV): m-QReLU 0.92 0.95 0.92 0.93

(0.90-0.95) (0.94-0.96) (0.90- (0.91-

Both accuracy and reliability metrics have been
0.94) 0.95)
reported in Tables III and IV, as well as their 95%
confidence intervals. hyper-sinh 0.81 0.66 0.81 0.73

(0.78-0.83) (0.65-0.69) (0.79- (0.72-

0.84) 0.75)
TABLE III. COMPARISON OF CLASSIFICATION
QReLU: Quantum ReLU; m-QReLU: modified Quantum ReLU; hyper-
PERFORMANCE OF THE CONVOLUTIONAL NEURAL
sinh: hyperbolic sinh.
NETWORK WITH ALL ACTIVATION FUNCTIONS TESTED,
EVALUATED ON THE SPIRAL HANDPD BENCHMARK
DATASET [21].
TABLE IV. COMPARISON OF CLASSIFICATION
Activation Accuracy Precision Recall F1- PERFORMANCE OF THE CONVOLUTIONAL NEURAL
function score NETWORK WITH ALL ACTIVATION FUNCTIONS TESTED,
(0-1) (0-1) (0-1) EVALUATED ON THE NEWHANDPD BENCHMARK DATASET
(0-1) [22].
ReLU 0.81 1.00 0.81 0.90
Activation Accuracy Precision Recall F1-
(0.76-0.87) (0.98-1.00) (0.78- (0.85- function (0-1) score
(0-1) (0-1)
0.83) 0.93)
(0-1)
Leaky ReLU 0.82 0.83 0.82 0.83
ReLU 0.75 0.76 0.75 0.76
(0.79-0.85) (0.81-0.85) (0.79- (0.80-
(0.73- (0.74- (0.73- (0.74-
0.84) 0.84)
0.76) 0.78) 0.77) 0.77)
Sigmoid 0.85 0.88 0.85 0.86

E-ISSN: 2415-1521 5 Volume 9, 2021

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

Leaky 0.83 0.83 0.83 0.83 Future work is underway to understand if

ReLU
(0.82- (0.81- (0.81- (0.81- hyperparameter optimisation can level out differences in
0.84) 0.84) 0.84) 0.84) overall classification performance of the CNN when
leveraging the activation functions ‘hyper-sinh’, ‘QReLU’
Sigmoid 0.57 1.00 0.57 0.72
and ‘m-QReLU’. Additional training data will be required
(0.53- (0.73- (0.54- (0.69- minimise the overlapping regions of the transformed data
0.59) 1.00) 0.59) 0.75) associated with physiological and pathophysiological
Tanh 0.94 0.94 0.94 0.94 tremor more reliably.

(0.93- (0.92- (0.92- (0.92-

0.95) 0.95) 0.95) 0.95)

QReLU 0.83 0.83 0.83 0.83 Acknowledgment

(0.82- (0.81- (0.82- (0.82-
The authors would like to thank the University of
0.85) 0.86) 0.84) 0.85) Bradford (Faculty of Engineering & Informatics) and the
University of Auckland Rehabilitative Technologies
m-QReLU 0.79 0.79 0.79 0.79
Association (UARTA) for giving them the opportunity to
(0.78- (0.77- (0.78- (0.78- develop this collaborative research work. This research did
0.81) 0.82) 0.80) 0.81) not receive any specific grant from funding agencies in the
hyper- 0.91 0.91 0.91 0.91 public, commercial, or not-for-profit sectors.
sinh
(0.89- (0.88- (0.89- (0.89-
0.94) 0.94) 0.94) 0.94)

QReLU: Quantum ReLU; m-QReLU: modified Quantum ReLU; hyper-

sinh: hyperbolic sinh.

3. Discussion and Conclusion

The supervised DL-based algorithm alone (CNN) could

detect early PD with 81% and 91% classification accuracy,
and 73% and 91% F1-score (Tables III and IV) via the
recent hyper-sinh activation function on the PD-related
Spiral HandPD [21] and the NewHandPD [22] image
datasets, respectively. Thus, the hyper-sinh [7], when
coupled with a CNN, led to the highest classification
performance (Tables III and IV). The QReLU and m-
QReLU also performed better than the other activation
functions when classifying both datasets (88% and 83%
classification accuracy, and 90% and 83% F1-score,
Tables III and IV), except for the tanh function for the
second dataset. The sigmoid activation functions led to the
worst classification performance on the NewHandPD
dataset (57% classification accuracy, Table IV), despite a
good reliability (72% F1-score, Table IV).

This study validates the use of hyper-sinh when

coupled with a CNN for supervised DL-based
classification to aid early detection of PD from image
features of spiral drawings from graphics tablets, which is
an inexpensive and non-invasive tool to facilitate its early
diagnosis.

E-ISSN: 2415-1521 6 Volume 9, 2021

 WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2021.9.1 Luca Parisi, Amir Zaernia, Renfei Ma, Mansour Youseffi

References
[1] Chakravarthy, K. V., Chaturvedi, R., Agari, T., Iwamuro, H., [14] Parisi, L., RaviChandran, N., & Manaog, M. L. (2018).
Reddy, R., & Matsui, A. (2020). Single arm prospective Decision support system to improve postoperative
multicenter case series on the use of burst stimulation to discharge: A novel multi-class classification
improve pain and motor symptoms in Parkinson’s approach. Knowledge-Based Systems, 152, 1-10.
disease. Bioelectronic medicine, 6(1), 1-8.
[15] Dahl, G. E., Sainath, T. N., & Hinton, G. E. (2013, May).
[2] Parisi, L., RaviChandran, N., & Manaog, M. L. (2018). Improving deep neural networks for LVCSR using rectified
Feature-driven machine learning to improve early diagnosis linear units and dropout. In 2013 IEEE international
of Parkinson's disease. Expert Systems with conference on acoustics, speech, and signal processing (pp.
Applications, 110, 182-190. 8609-8613). IEEE.

[3] Carmeli, E., Patish, H., & Coleman, R. (2003). The aging [16] Parisi, L., Neagu, D., Ma, R., & Campean, F. (2020). QReLU
hand. The Journals of Gerontology Series A: Biological and m-QReLU: Two novel quantum activation functions to
Sciences and Medical Sciences, 58(2), M146-M152. aid medical diagnostics. arXiv preprint arXiv:2010.08031.

[4] Rosenblum, S., Samuel, M., Zlotnik, S., Erikh, I., & [17] Glorot, X., Bordes, A., & Bengio, Y. (2011, June). Deep
Schlesinger, I. (2013). Handwriting as an objective tool for sparse rectifier neural networks. In Proceedings of the
Parkinson’s disease diagnosis. Journal of neurology, 260(9), fourteenth international conference on artificial intelligence
2357-2361. and statistics (pp. 315-323). JMLR Workshop and
Conference Proceedings.
[5] Drotár, P., Mekyska, J., Rektorová, I., Masarová, L.,
Smékal, Z., & Faundez-Zanuy, M. (2016). Evaluation of [18] Parisi, L. (2020). QReLU and m-QReLU in TensorFlow and
handwriting kinematics and pressure for differential Keras. GitHub repository. https://github.com/luca-
diagnosis of Parkinson's disease. Artificial intelligence in parisi/QReLU_m-QReLU_TensorFlow_Keras.
Medicine, 67, 39-46.
[19] Parisi, L. (2020). m-arcsinh in scikit-learn. GitHub
[6] San Luciano, M., Wang, C., Ortega, R. A., Yu, Q., Boschung, repository. https://github.com/luca-parisi/m-arcsinh_scikit-
S., Soto-Valencia, J., ... & Saunders-Pullman, R. (2016). learn.
Digitized spiral drawing: a possible biomarker for early
Parkinson’s disease. PloS one, 11(10), e0162799. [20] Parisi, L. (2020). hyper-sinh in TensorFlow and
Keras. GitHub repository. https://github.com/luca-
[7] Parisi, L., Ma, R., RaviChandran, N., & Lanzillotta, M. parisi/hyper-sinh_TensorFlow_Keras.
(2020). hyper-sinh: An Accurate and Reliable Function from
Shallow to Deep Learning in TensorFlow and Keras. arXiv [21] Pereira, C. R., Pereira, D. R., Silva, F. A., Masieiro, J. P.,
preprint arXiv:2011.07661. Weber, S. A., Hook, C., & Papa, J. P. (2016). A new
computer vision-based approach to aid the diagnosis of
[8] Clevert, D. A., Unterthiner, T., & Hochreiter, S. (2015). Fast Parkinson's disease. Computer Methods and Programs in
and accurate deep network learning by exponential linear Biomedicine, 136, 79-88.
units (elus). arXiv preprint arXiv:1511.07289.
[22] Pereira, C. R., Weber, S. A., Hook, C., Rosa, G. H., & Papa,
[9] Pascanu, R., Mikolov, T., & Bengio, Y. (2012). J. P. (2016). Deep learning-aided Parkinson's disease
Understanding the exploding gradient problem. CoRR, diagnosis from handwritten dynamics. In 2016 29th
abs/1211.5063, 2(417), 1. SIBGRAPI Conference on Graphics, Patterns, and Images
(SIBGRAPI), 340-346. IEEE.
[10] Parisi, L. (2020). m-arcsinh: An Efficient and Reliable
Function for SVM and MLP in scikit-learn. arXiv preprint [23] LeCun, Y., & Bengio, Y. (1995). Convolutional networks for
arXiv:2009.07530. images, speech, and time series. The handbook of brain
theory and neural networks, 3361(10), 1995.
[11] Parisi, L., RaviChandran, N., & Manaog, M. L. (2020). A
novel hybrid algorithm for aiding prediction of prognosis in [24] Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai, B.,
patients with hepatitis. Neural Computing and & Chen, T. (2018). Recent advances in convolutional neural
Applications, 32(8), 3839-3852. networks. Pattern Recognition, 77, 354-377.
[12] Parisi, L., & RaviChandran, N. (2020). Evolutionary
Denoising-Based Machine Learning for Detecting Knee Creative Commons Attribution License 4.0
Disorders. Neural Processing Letters, 52(3), 2565-2581.
(Attribution 4.0 International, CC BY 4.0)
[13] Parisi, L., & RaviChandran, N. (2020). Evolutionary feature
transformation to improve prognostic prediction of This article is published under the terms of the Creative
hepatitis. Knowledge-Based Systems, 200, 106012. Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US

E-ISSN: 2415-1521 7 Volume 9, 2021

View publication stats