Spiral Drawing Test and Explainable Convolutional Neural Networks for

Parkinson’s Disease Detection

Francesco Mercaldo

1,3

, Luca Brunese

, Mario Cesarelli

, Fabio Martinelli

and Antonella Santone

Department of Medicine and Health Sciences “Vincenzo Tiberio”, University of Molise, Campobasso, Italy

Department of Engineering, University of Sannio, Benevento, Italy

Institute for Informatics and Telematics, National Research Council of Italy (CNR), Pisa, Italy

mcesarelli@unisannio.it

Keywords:

Parkinson, Spiral, Machine Learning, Deep Learning, Explainability.

Abstract:

There is no deﬁnitive test for Parkinson’s disease, and the rate of misdiagnosis, particularly when made by in-

dividuals without specialized training, is signiﬁcantly elevated. The spiral drawing test is a clinical assessment

tool used to evaluate ﬁne motor skills, hand-eye coordination, and tremor in individuals, particularly those

with neurological disorders such as Parkinson’s disease. In this test, a person is typically asked to trace or

draw a spiral pattern on a piece of paper or a digital tablet. The test measures the smoothness and steadiness of

their hand movements. Any irregularities or tremors in the drawn spiral can provide valuable information to

healthcare professionals in diagnosing or monitoring conditions like Parkinson’s disease, essential tremors, or

other movement disorders. In this paper, we provide a method aimed at automatically analyse spiral drawing

tests to understand whether a subject is affected by Parkinson’s disease. We employ two different Convolu-

tional Neural Networks: DenseNet and ResNet50, by obtaining an accuracy equal to 0.96 in the evaluation

of a dataset composed of 3,991 spiral drawing tests, thus showing the effectiveness of the proposed method.

Moreover, with the aim to provide a kind of explainability behind the model prediction, the proposed method

is able to visualise, directly on the spiral drawing test image, the areas of the test image that from the model

point of view are related to Parkinson’s disease.

1 INTRODUCTION AND

RELATED WORK

Parkinson’s disease (PD) is a neurodegenerative dis-

order that primarily affects movement (Balestrino and

Schapira, 2020). It is a chronic and progressive con-

dition that typically develops slowly over time. The

main features of Parkinson’s disease include:

• Tremors. Resting tremors are a common symp-

tom, typically starting in one hand and often de-

scribed as a ”pill-rolling” tremor.

• Bradykinesia. This refers to slowness of move-

ment. People with PD may have difﬁculty initiat-

ing and completing voluntary movements, leading

to a gradual reduction in their ability to perform

everyday tasks.

• Muscle Rigidity. Stiffness of the muscles can

make it difﬁcult for individuals with Parkinson’s

to move smoothly.

• Postural Instability. Balance problems can lead to

a greater risk of falls and other injuries.

In addition to motor symptoms, PD can also cause

a range of non-motor symptoms, including:

• Cognitive Changes. Some individuals may ex-

perience cognitive impairment, which can range

from mild memory problems to more severe is-

sues like dementia.

• Mood Disorders. Depression and anxiety are

common in people with Parkinson’s disease.

• Sleep Disturbances. Sleep problems, such as in-

somnia or excessive daytime sleepiness, can oc-

cur.

• Autonomic Dysfunction. This can lead to is-

sues with blood pressure regulation, digestion,

and other bodily functions.

The exact cause of PD is not fully understood

(Poewe et al., 2017), but it is believed to involve

a combination of genetic and environmental factors.

Mercaldo, F., Brunese, L., Cesarelli, M., Martinelli, F. and Santone, A.

Spiral Drawing Test and Explainable Convolutional Neural Networks for Parkinson’s Disease Detection.

DOI: 10.5220/0012407100003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 2, pages 443-452

ISBN: 978-989-758-680-4; ISSN: 2184-433X

443

There is currently no cure for PD, but there are treat-

ments available to manage its symptoms. Medi-

cations, physical therapy, occupational therapy, and

lifestyle modiﬁcations can all help improve the qual-

ity of life for individuals with Parkinson’s disease.

In some cases, surgical interventions like deep

brain stimulation may be considered to alleviate

symptoms, particularly for those who do not respond

well to medication.

It is important for individuals with PD to work

closely with healthcare professionals, such as neurol-

ogists and physical therapists, to develop a personal-

ized treatment plan and receive the necessary support

for managing the condition.

PD is currently the second most common neurode-

generative disorder after Alzheimer’s disease. Ac-

cording to the Parkinson’s Foundation

, nearly one

million people in the U.S. are living with PD. This

number is expected to rise to 1.2 million by 2030.

Moreover, nearly 90,000 people in the U.S. are di-

agnosed with PD each year.

Its prevalence tends to increase with age. It is esti-

mated that about 1% of the population over the age of

60 is affected by PD. As a matter of fact, while PD can

affect people of all ages, it is most commonly diag-

nosed in people over the age of 60. It is relatively rare

in younger individuals and it is more common in men

than in women. Men are about 1.5 times more likely

to develop the condition. The incidence of new cases

of PD is estimated to be around 20 cases per 100,000

people per year. PD is a progressive condition, and its

rate of progression can vary from person to person.

Some individuals may experience a relatively slow

progression, while others may progress more quickly.

it can signiﬁcantly impact an individual’s quality of

life and it can lead to difﬁculties in performing daily

activities, increased healthcare costs, and a decreased

ability to work. Moreover, while PD itself is not typ-

ically considered a direct cause of death, complica-

tions related to the condition, such as falls and pneu-

monia, can increase mortality risk. It can also impose

a substantial economic burden on individuals, fami-

lies, and healthcare systems. Costs associated with

medical care, medications, and lost productivity can

be signiﬁcant.

Currently, there is no objective test available for

PD, and the rate of misdiagnosis, especially when

made by non-specialists, is quite high, with the like-

lihood of an incorrect diagnosis reaching up to 20%

(Rizzo et al., 2016). While a careful analysis of pri-

mary symptoms like tremors, bradykinesia, and rigid-

ity can enhance diagnostic accuracy, clinical assess-

ments may still be inﬂuenced by the subjectivity of

https://www.parkinson.org/

the physician. To address this issue, the use of med-

ical decision support tools is of great interest as they

can increase objectivity and aid in early diagnosis.

This early diagnosis is crucial as it paves the way

for the development of tailored treatments for PD-

affected patients (Ammenwerth et al., 2013; Dreiseitl

and Binder, 2005). An important objective in neu-

rodegenerative disease research is the identiﬁcation of

precise biomarkers (Mattison et al., 2012).

In the scientiﬁc literature, there is a wide array of

studies dedicated to PD detection through speech pro-

cessing (Lahmiri and Shmuel, 2019; G

omez-Garc

ıa

et al., 2019), where the diagnosis is based on sus-

tained vowels and natural speech. Additionally, mo-

tor symptoms can be detected and monitored by mod-

eling patient movements and gait (Viteckova et al.,

2018; San-Segundo et al., 2019).

One of the initial signs often observed in PD is

alterations in the kinematics of handwriting. McLen-

nan et al. (Letanneux et al., 2014) found that around

5% of PD patients exhibited micrographia (abnor-

mally small letter size), and 30% reported deterio-

rating handwriting before the onset of motor symp-

toms. The motor symptoms associated with PD, such

as stiffness, bradykinesia, and tremors, result in three

primary changes in writing (Zham et al., 2017a): the

size of writing (Potgieser et al., 2015) (micrographia

(Drot

ar et al., 2016)), pen-pressure (Letanneux et al.,

2014), and kinematics. Various tools have been devel-

oped for the analysis of handwriting related to PD pa-

tients (Chatterjee and Kordower, 2019). It is not just

the static aspects but also the dynamic ones that are

of interest, including speed and pen-pressure reduc-

tion during writing (Drot

ar et al., 2016; Rosenblum

et al., 2013).

Several recent review papers on this topic have

been published (Impedovo and Pirlo, 2019; Impedovo

and Pirlo, 2018). It is important to note that an in-

dividual’s handwriting can be inﬂuenced by their vi-

sual capability (Potgieser et al., 2015), writing style,

or language skills, leading to signiﬁcant variability

among individuals. An alternative to handwriting

analysis is the use of drawings, for instance, the spiral

drawing test.

In this direction, Kotsavasiloglou et al. (Kot-

savasiloglou et al., 2017) employed a pen-and-tablet

device to examine differences in hand movement and

muscle coordination between healthy individuals and

PD patients. They used ﬁve metrics, including mean

horizontal velocity, normalized velocity variability (in

units per second), standard deviation of horizontal ve-

locity, and entropies of horizontal and vertical signal

components. Their evaluation of various classiﬁca-

tion algorithms resulted in the best accuracy equal to

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

444

0.88. Zham et al. (Zham et al., 2017a) assessed 10

features, combining static and dynamic information,

using the Na

ıve Bayes algorithm for classiﬁcation,

achieving an accuracy of 0.83. Both of these previ-

ous papers (Zham et al., 2017a; Kotsavasiloglou et al.,

2017) made decisions at intervals of approximately 2

seconds for each drawing.

Gallicchio et al. (Gallicchio et al., 2018) proposed

the utilization of DeepESNs, achieving an accuracy

of 0.89. Meanwhile, Khatamino et al. (Khatamino

et al., 2018) employed a Convolutional Neural Net-

work (CNN) inspired by the AlexNet architecture

(Krizhevsky et al., 2012), which consisted of two

main components (convolutional layers for feature ex-

traction and fully connected layers for classiﬁcation).

A test that involves drawing a spiral on a sheet of

paper could be used to diagnose early PD i.e., the so-

called Spiral Drawing Test (Chakraborty et al., 2020).

The Spiral Drawing Test is a neuropsychological

test used to assess ﬁne motor skills, coordination, and

dexterity, particularly in the context of neurological

and motor function evaluations(Stanley et al., 2010).

It is a relatively simple test that involves asking an

individual to draw a spiral or a series of spirals on a

piece of paper. The evaluator may provide speciﬁc in-

structions, such as starting from the center and work-

ing outward or starting at a particular point on the pa-

per (San Luciano et al., 2016).

The Spiral Drawing Test can be used in various

clinical settings, including assessing neurological dis-

orders such as Parkinson’s disease. In such cases, it

can help evaluate ﬁne motor control and detect any

abnormalities or tremors in the drawing pattern. For

example, people with Parkinson’s disease may pro-

duce spirals with more visible tremors or irregulari-

ties, which can be indicative of their motor control

issues (Kamble et al., 2021).

This test is also used in research and clinical as-

sessments to evaluate other conditions, such as essen-

tial tremor, multiple sclerosis, and stroke, which can

affect motor skills and coordination.

The Spiral Drawing Test is just one tool in a bat-

tery of assessments used by healthcare profession-

als to diagnose and monitor neurological and motor

function disorders. When used alongside other clin-

ical evaluations, medical history, and imaging tests,

it can provide valuable insights into a patient’s con-

dition and aid in treatment planning and monitoring

(Gil-Mart

ın et al., 2019).

In recent years, deep learning has demonstrated

remarkable capabilities in the ﬁeld of image classiﬁ-

cation(Cimitile et al., 2017; Bacci et al., 2018; Huang

et al., 2023; Mercaldo and Santone, 2021). CNNs are

a speciﬁc class of deep learning models that have been

particularly successful in biomedical image classiﬁca-

tion tasks (Huang et al., 2022). Despite the possibil-

ity of deep learning being highly successful in image

classiﬁcation, it is important to note that it requires

substantial computational resources, large datasets,

and specialized hardware (such as GPUs) for train-

ing. Additionally, the explainability of deep learning

models can be challenging, which is a topic of ongo-

ing research and development that limits the applica-

tion of deep learning in the real-world domain, with

particular regard to healthcare (Huang et al., 2021).

For these reasons, in this paper, we propose an ex-

plainable method aimed at detecting whether a patient

is affected by PD by analysing the Spiral Drawing

Test. We consider several CNNs for the classiﬁca-

tion task (aimed to discriminate between healthy and

unhealthy patients).

The paper proceeds as follows: in the next sec-

tion, we present the proposed method for explainable

PD detection from the spiral drawing test, in Section

3 the experimental analysis results are presented and,

ﬁnally, in the last section, conclusion and future re-

search directions are drawn.

2 THE METHOD

As mentioned in the previous section, the pro-

posed approach employs supervised machine learn-

ing, speciﬁcally delving into deep learning techniques

through the utilization of CNNs. CNNs stand out as

a type of artiﬁcial neural network exceptionally well-

suited for tasks involving image classiﬁcation, mak-

ing them particularly relevant in the realm of diag-

nosing brain cancer.

In this approach, CNNs are subjected to training

using a labeled dataset encompassing images of both

spiral drawing tests drawn by healthy subjects and

from subjects affected by PD. The network gleans

knowledge from these instances and distills meaning-

ful characteristics from the images to discern between

patterns associated with healthy subjects and those in-

dicative of PD. The training procedure involves iter-

atively ﬁne-tuning the network’s parameters to mini-

mize classiﬁcation errors and enhance its precision in

distinguishing between the two categories.

Once the CNN has been sufﬁciently trained, it can

be applied to categorize new, previously unseen brain

images as either healthy or cancer-affected, based on

the knowledge it has acquired. The network assesses

the input image using its learned ﬁlters and identiﬁes

pertinent features to make a prediction.

It is essential to acknowledge that the effective-

ness and accuracy of the proposed approach depend

Spiral Drawing Test and Explainable Convolutional Neural Networks for Parkinson’s Disease Detection

445

Figure 1: The workﬂow of the proposed method for explain-

able PD detection through spiral draw testing.

on various factors, including the quality and repre-

sentativeness of the training data, the architecture

and conﬁgurations of the CNN, and the methods em-

ployed for validation and testing.

The workﬂow of the proposed method is shown in

Figure 1.

The proposed method comprises ﬁve distinct

phases:

1. Dataset. An essential element in developing an

effective model for brain cancer diagnosis is the

dataset used in machine learning. It is crucial

to possess a meticulously annotated dataset that

encompasses both images of spiral drawing tests

obtained from healthy subjects and subjects with

PD. Ensuring the model’s robustness and general-

ization capability requires a diverse and represen-

tative dataset.

2. Data Augmentation. Once the dataset is obtained,

it is crucial to preprocess the images to stan-

dardize and remove biases introduced by various

imaging settings. Common preprocessing tech-

niques include brightness adjustment during train-

ing, which mitigates intensity variations. In the

proposed paper we consider the following data

augmentation techniques: random rotations, hori-

zontal and vertical reﬂections, zoom in and zoom

out, random cropping, scaling, contrast, bright-

ness, and saturation adjustments. As a matter

of fact, data Augmentation, especially for small-

sized datasets, can generate new data examples

from existing ones, preserving the same general

characteristics but with random variations that can

enhance the model’s generalization ability. These

techniques aim to enhance data consistency and

algorithm effectiveness.

3. CNN Model. After data collection and augmenta-

tion, the next step involves selecting deep learning

models. We consider a binary classiﬁcation task,

where a spiral drawing test can be classiﬁed as be-

longing to a healthy patient or to a patient affected

by PD. Consideration extends beyond accuracy:

explainability is essential, particularly in medical

applications. Choosing appropriate hyperparame-

ters, such as the number of epochs, batch size, and

learning rate, requires careful consideration and

experimentation. In this paper, two of the most

widespread deep learning architectures based on

CNNs are exploited: Densely Connected Con-

volutional Networks (DenseNet) (Huang et al.,

2019) and Residual Networks (ResNet) (He et al.,

2016). Both of which have made signiﬁcant con-

tributions to the ﬁeld of computer vision and im-

age classiﬁcation (Zhang et al., 2021; Li et al.,

2020; Liu et al., 2021; Chen et al., 2022). The fol-

lowing is a brief description of each architecture

we exploited in this paper:

• The key innovation in DenseNet is its dense

connectivity pattern. In traditional CNN ar-

chitectures, each layer typically receives inputs

only from the previous layer. In DenseNet,

each layer is connected to all subsequent lay-

ers. This dense connectivity encourages fea-

ture reuse and gradient ﬂow throughout the net-

work, making it more efﬁcient and reducing the

risk of vanishing gradients. DenseNet is com-

posed of dense blocks, each containing a se-

ries of convolutional layers, batch normaliza-

tion, and non-linear activation functions. Skip

connections from earlier layers are concate-

nated with the feature maps in the current layer.

The dense connectivity allows for more efﬁ-

cient parameter usage, enabling the construc-

tion of deep networks with relatively fewer pa-

rameters.

• The central idea behind ResNet is the use of

residual connections, which allow for very deep

networks to be trained effectively. A residual

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

446

block contains a shortcut connection (skip con-

nection) that bypasses one or more convolu-

tional layers. The residual block reformulates

the learning problem as learning the residual of

the identity mapping. This makes it easier to

train deep networks, as the network can learn to

adjust the output from a previous layer. ResNet

architectures come in various depths, such as

ResNet18, ResNet50, ResNet101, and more,

where the numbers indicate the number of lay-

ers in the network. In this paper, we experi-

ment with the ResNet50 architecture. ResNet’s

use of residual connections has been instru-

mental in enabling the training of extremely

deep networks, leading to signiﬁcant improve-

ments in performance on image classiﬁcation

tasks. ResNet has also been widely adopted

and adapted for various computer vision tasks,

including object detection and image segmen-

tation.

Both DenseNet and ResNet are deep neural net-

work architectures that have had a profound im-

pact on the ﬁeld of computer vision. DenseNet

focuses on dense connectivity, while ResNet uti-

lizes residual connections to enable the training of

very deep networks, leading to improved perfor-

mance in various image analysis tasks: these are

the reasons why we experiment with both of them

in this paper.

4. Training-Test. in this step, we consider Model

training and testing, involving computing met-

rics like Accuracy, Precision, and Recall to as-

sess prediction efﬁciency. If results are unsatis-

factory, different combinations of hyperparame-

ters are considered to achieve desired outcomes

(Mercaldo and Santone, 2021).

5. Grad-CAM. The Grad-CAM (Gradient-weighted

Class Activation Mapping) algorithm is utilized to

create heatmaps, offering visual explanations for

model predictions. Beyond prediction accuracy,

the model’s ability to highlight areas in the input

image that inﬂuenced the classiﬁcation is evalu-

ated. The aim of the Grad-CAM is to extract

gradients from the model’s convolutional layers

to provide these visual explanations, offering in-

sights into the decision-making process.

3 EXPERIMENTAL EVALUATION

With the aim of validating the proposed method, we

exploit a dataset containing 3,991 images related to

spiral drawing tests, with 1,995 images correspond-

Figure 2: Spiral drawing tests related to a healthy (on the

left) and an unhealthy patient (on the right).

ing to patients diagnosed with PD and the remain-

ing 1,996 images representing individuals without the

disease condition (Zham et al., 2017b). This dataset,

referred to as ”Parkinson’s Drawings” is accessible

for research purposes on the Kaggle website and can

be freely obtained

. As discussed in the previous sec-

tion, we consider a binary classiﬁcation task, where

each image obtained from a spiral drawing test can be

classiﬁed as belonging to a PD-affected patient or to

a healthy one.

Figure 2 shows two examples of spiral drawing

tests (belonging to the analysed dataset): the test on

the left is related to a healthy patient, while the one

on the right is drawn by a PD-affected subject.

As shown in Figure 2, we can note that in the case

of the patient PD-affected on the left, the spiral lines

are less linear than in the case on the left of the healthy

patient, due to the trembling caused by the disease.

The dataset is later divided into an 80:10:10 ratio

for the training, validation, and testing datasets, re-

sulting in the selection of 1,579 images for training

(790 associated with healthy patients and the remain-

ing 789 with PD patients), 200 images for validation

(100 from each group), and the remaining 200 images

for testing (100 from each group).

With the aim of understanding the effectiveness of

the models in the discrimination between spiral draw-

ing tests related to healthy and PD-affected patients,

the following metrics are computed:

• Loss. The loss, also known as the cost or ob-

jective function, quantiﬁes the error between the

predicted values generated by a machine learn-

ing model and the actual ground truth values in

the dataset. The goal is to minimize this error

during training. Various loss functions, such as

mean squared error (MSE), cross-entropy loss,

and hinge loss, are used depending on the prob-

lem type (e.g., regression or classiﬁcation).

• Accuracy. Accuracy is a classiﬁcation metric that

measures the proportion of correctly predicted in-

https://www.kaggle.com/datasets/kmader/parkinson

s-drawings

Spiral Drawing Test and Explainable Convolutional Neural Networks for Parkinson’s Disease Detection

447

stances in a dataset. It’s calculated as the number

of correct predictions divided by the total num-

ber of predictions. High accuracy indicates that a

model is making correct predictions for most of

the data.

• Precision. Precision is a metric used in binary

classiﬁcation. It measures the accuracy of positive

predictions made by a model. It is calculated as

the number of true positives divided by the sum of

true positives and false positives. Precision helps

assess the model’s ability to avoid false positives.

• Recall. Recall, also known as sensitivity or true

positive rate, measures the ability of a model to

correctly identify positive instances from the to-

tal number of actual positives. It is calculated as

the number of true positives divided by the sum of

true positives and false negatives. Recall is essen-

tial for understanding how well a model captures

all relevant instances.

• F-Measure. The F-Measure is a metric that com-

bines both precision and recall into a single value.

It is the harmonic mean of precision and recall and

provides a balance between them. A higher F1

Score indicates a model that has good precision

and recall.

• Area Under the ROC Curve (AUC). AUC is a met-

ric used to evaluate the performance of a binary

classiﬁcation model. It measures the ability of the

model to distinguish between positive and nega-

tive classes across different classiﬁcation thresh-

olds. The ROC (Receiver Operating Character-

istic) curve is a graphical representation of the

trade-off between true positive rate (recall) and

false positive rate as the decision threshold varies.

The AUC is the area under this ROC curve. A

higher AUC signiﬁes better model performance,

with a value of 1 indicating perfect classiﬁcation.

Next, we proceed with the training of the deep

learning models. To ensure replicability, in the fol-

lowing, we show the hyperparameters we used for

training the DenseNet and MobileNet models in Ta-

ble 1.

Table 1: Hyper-parameters setting.

Batch size Learning rate Image size

32 0.01 224x224x3

The results of the experimental analysis are dis-

played in Table 2.

As shown in Table 2 we ran several experiments:

the ﬁrst ones with a number of epochs equal to 20:

with this number of epochs we obtained an accuracy

equal to 0.86 with the DenseNet network and an accu-

racy of 0.75 by exploiting the ResNet50 one. Consid-

ering that from these two experiments, the DenseNet

network demonstrated better performances in the dis-

crimination of PD-affected patients and healthy ones,

we consider the DenseNet model for additional exper-

iments with a different number of epochs: 50 and 100.

With a number of epochs of 50 we obtain an accuracy

equal to 0.96, while when the number of epochs is

increased to 100, the accuracy obtained is 0.89.

Thus, from the results shown in Table 2, it

emerges that the model obtaining the best perfor-

mances in PD-affected patients is the DenseNet one,

trained with a number of epochs equal to 50, with an

accuracy, a precision and a recall equal to 0.96.

With the aim to show how the proposed method

can provide a kind of explainability behind the model

prediction, we show four different examples of Grad-

CAM applications (in Figures 3, 4, 5 and 6), belong-

ing to two healthy patients (Figures 3, 4) and to two

subjects PD affected (Figures 5, 6), obtained with the

best model i.e., the DenseNet one, trained with for 50

epochs.

For instance, in Figure 3 there is an example of

localization provided by the Grad-CAM algorithm re-

lated to a healthy patient (correctly predicted with a

percentage equal to 92.3%). In particular, the pixels

highlighted in yellow relate to areas that were of par-

ticular interest for the model to make the predictions,

the green areas are areas of interest for the model, but

of lesser interest than the areas highlighted in yellow.

The areas in purple are the areas that are not of in-

terest to the model for the purposes of the prediction

made.

For this reason, in Figure 3 we can note that the

areas of interest from the model point of view are the

ones related to the smallest part of the spiral and some

areas on the left of the image (as we can note from the

yellow areas in the overlay with heatmap image).

Figure 4 shows another example of prediction re-

lated to a healthy patient.

In the Grad-CAM generated for the spiral draw-

ing test shown in Figure 4 we can note that the areas

of interest from the model point of view are several

areas, all focused on the outermost spiral curves com-

pared to the previous example. The fact that different

areas of the spiral are highlighted on a different image

is a symptom that the model has learned to generalize

the distinctive features of a spiral written by a healthy

user, regardless of the area of the spiral or sheet where

it was written, demonstrating, therefore, a good model

generalization ability, fundamental in machine learn-

ing problems to avoid models capable of identifying

only certain types of cases (and therefore poorly gen-

eralizable).

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

448

Table 2: The experimental analysis results obtained with the DenseNet and the ResNet50 models.

Model Epochs Loss Accuracy Precision Recall F-Measure AUC

DenseNet 20 0.34 0.86 0.86 0.86 0.86 0.93

DenseNet 50 0.17 0.96 0.96 0.96 0.96 0.98

DenseNet 100 0.51 0.89 0.89 0.89 0.89 0.94

ResNet50 20 0.48 0.75 0.75 0.75 0.75 0.84

Figure 3: An example of localization provided by the Grad-CAM algorithm related to a healthy patient predicted with a

percentage equal to 92.3%.

Figure 4: An example of localization provided by the Grad-CAM algorithm related to a healthy patient predicted with a

percentage equal to 85.1%.

Figure 5 shows an example of prediction related to

an unhealthy patient (i.e., PD-affected) predicted with

a percentage equal to 94.9%.

In this prediction example, we can see that the

area of interest of the model (highlighted by the Grad-

CAM in yellow) is related to different areas of the spi-

ral, therefore the model recognized different areas of

the spiral as symptomatic of the presence of PD.

Figure 6 shows another example of a (correct) pre-

diction of a patient PD-affected, detected with a per-

centage equal to 100%.

In this last case, we can see how the spiral is de-

cidedly different in its features compared to the ex-

amples of healthy patients (shown in Figures 3 and

4) but also compared to the PD-affected patient ana-

lyzed in the previous example (i.e., in Figure 5): as

a matter of fact, the features of the spiral are decid-

edly less linear, and this is symptomatic of a patient

who is evidently suffering from severe trembling, one

of the main symptoms of PD. In this case, the areas

of interest for Grad-CAM are those in which the non-

linearity of the spiral curves is greater, therefore in

the innermost part of the spiral we can see how there

is a concentration of yellow areas. The model pre-

dicted the test with a percentage equal to 100% as the

presence of the tremor appears decidedly more evi-

dent than in the case of the PD patient analyzed pre-

viously.

Also in this case the model highlighted different

areas of the image, demonstrating that the distinc-

tive features of the disease were identiﬁed regardless

of the area of the image in which they were present,

showing the generability of the model trained through

the DenseNet network.

Spiral Drawing Test and Explainable Convolutional Neural Networks for Parkinson’s Disease Detection

449

Figure 5: An example of localization provided by the Grad-CAM algorithm related to an unhealthy patient (i.e., PD-affected)

predicted with a percentage equal to 94.9%.

Figure 6: An example of localization provided by the Grad-CAM algorithm related to an unhealthy patient (i.e., PD-affected)

correctly predicted with a percentage equal to 100%.

4 CONCLUSIONS AND FUTURE

WORK

Considering that there is no deﬁnitive diagnostic test

available for PD diagnosis, and the likelihood of mis-

diagnosis is notably higher, particularly when the

diagnosis is made by individuals lacking special-

ized training, in this paper we proposed a method

aimed to discriminate between PD-affected patients

and healthy subjects. For this task, we consider differ-

ent CNN models trained by exploiting images related

to spiral drawing tests. Two different CNNs are con-

sidered: ResNet50 and DenseNet. The experimen-

tal results analysis demonstrated that the DenseNet

model is able to obtain better performances if com-

pared with the ResNet50 one: as a matter of fact,

accuracy, precision, recall, and F-Measure equal to

0.96 are obtained, in the evaluation of a dataset re-

lated to 3,991 different images related to spiral draw-

ing tests, with 1,995 images corresponding to patients

PD-affected and the remaining 1,996 images repre-

senting individuals without the PD condition (i.e.,

healthy subjects). Moreover, with the aim to pro-

vide prediction explainability, we take into account

the Grad-CAM algorithm, able to highlight (with a

heatmap) the areas on the spiral drawing test image

symptomatic of a certain prediction, thus providing a

kind of rationale behind the model decision, by show-

ing the areas on the image responsible for the predic-

tion. In this way, we think that CNNs can be really

employed in real-world clinical decisions, as a mat-

ter of fact, due to the lack of prediction explainability

medical staff do not have a strong trust in the predic-

tion provided by machine learning models.

In future work, we plan to consider additional

CNN-based architectures, with the aim of improving

the obtained performances. Moreover, we will con-

sider not only images related to spiral drawing tests

but also images related to wave tests, another kind of

geometric drawing used to understand whether there

is the presence of PD. We will investigate to under-

stand if by exploiting wave images we are able to

obtain better PD detection than spiral ones, or if the

combination of spiral and wave images can help to

obtain better performances.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

450

ACKNOWLEDGEMENTS

This work has been partially supported by EU DUCA,

EU CyberSecPro, SYNAPSE, PTR 22-24 P2.01 (Cy-

bersecurity) and SERICS (PE00000014) under the

MUR National Recovery and Resilience Plan funded

by the EU - NextGenerationEU projects.

REFERENCES

Ammenwerth, E., Nyk

anen, P., Rigby, M., and de Keizer,

N. (2013). Clinical decision support systems: need

for evidence, need for evaluation.

Bacci, A., Bartoli, A., Martinelli, F., Medvet, E., and Mer-

caldo, F. (2018). Detection of obfuscation techniques

in android applications. In Proceedings of the 13th In-

ternational Conference on Availability, Reliability and

Security, pages 1–9.

Balestrino, R. and Schapira, A. (2020). Parkinson disease.

European journal of neurology, 27(1):27–42.

Chakraborty, S., Aich, S., Han, E., Park, J., Kim, H.-C.,

et al. (2020). Parkinson’s disease detection from spi-

ral and wave drawings using convolutional neural net-

works: A multistage classiﬁer approach. In 2020 22nd

International Conference on Advanced Communica-

tion Technology (ICACT), pages 298–303. IEEE.

Chatterjee, D. and Kordower, J. H. (2019). Immunother-

apy in parkinson’s disease: Current status and future

directions. Neurobiology of disease, 132:104587.

Chen, H., Liu, J., Hua, C., Feng, J., Pang, B., Cao, D., and

Li, C. (2022). Accurate classiﬁcation of white blood

cells by coupling pre-trained resnet and densenet with

scam mechanism. BMC bioinformatics, 23(1):1–20.

Cimitile, A., Martinelli, F., Mercaldo, F., et al. (2017). Ma-

chine learning meets ios malware: Identifying mali-

cious applications on apple environment. In ICISSP,

pages 487–492.

Dreiseitl, S. and Binder, M. (2005). Do physicians value de-

cision support? a look at the effect of decision support

systems on physician opinion. Artiﬁcial intelligence

in medicine, 33(1):25–30.

Drot

ar, P., Mekyska, J., Rektorov

a, I., Masarov

a, L.,

ekal, Z., and Faundez-Zanuy, M. (2016). Evalu-

ation of handwriting kinematics and pressure for dif-

ferential diagnosis of parkinson’s disease. Artiﬁcial

intelligence in Medicine, 67:39–46.

Gallicchio, C., Micheli, A., and Pedrelli, L. (2018). Deep

echo state networks for diagnosis of parkinson’s dis-

ease. arXiv preprint arXiv:1802.06708.

Gil-Mart

ın, M., Montero, J. M., and San-Segundo, R.

(2019). Parkinson’s disease detection from draw-

ing movements using convolutional neural networks.

Electronics, 8(8):907.

omez-Garc

ıa, J. A., Moro-Vel

azquez, L., and Godino-

Llorente, J. I. (2019). On the design of automatic

voice condition analysis systems. part ii: Review of

speaker recognition techniques and study on the ef-

fects of different variability factors. Biomedical Sig-

nal Processing and Control, 48:128–143.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Huang, G., Liu, Z., Pleiss, G., Van Der Maaten, L., and

Weinberger, K. Q. (2019). Convolutional networks

with dense connectivity. IEEE transactions on pat-

tern analysis and machine intelligence, 44(12):8704–

8716.

Huang, P., He, P., Tian, S., Ma, M., Feng, P., Xiao, H.,

Mercaldo, F., Santone, A., and Qin, J. (2022). A vit-

amc network with adaptive model fusion and multiob-

jective optimization for interpretable laryngeal tumor

grading from histopathological images. IEEE Trans-

actions on Medical Imaging, 42(1):15–28.

Huang, P., Tan, X., Zhou, X., Liu, S., Mercaldo, F., and

Santone, A. (2021). Fabnet: fusion attention block

and transfer learning for laryngeal cancer tumor grad-

ing in p63 ihc histopathology images. IEEE Journal

of Biomedical and Health Informatics, 26(4):1696–

1707.

Huang, P., Zhou, X., He, P., Feng, P., Tian, S., Sun, Y., Mer-

caldo, F., Santone, A., Qin, J., and Xiao, H. (2023).

Interpretable laryngeal tumor grading of histopatho-

logical images via depth domain adaptive network

with integration gradient cam and priori experience-

guided attention. Computers in Biology and Medicine,

154:106447.

Impedovo, D. and Pirlo, G. (2018). Dynamic handwriting

analysis for the assessment of neurodegenerative dis-

eases: a pattern recognition perspective. IEEE reviews

in biomedical engineering, 12:209–220.

Impedovo, D. and Pirlo, G. (2019). Online handwriting

analysis for the assessment of alzheimer’s disease and

parkinson’s disease: overview and experimental in-

vestigation. Frontiers In Pattern Recognition And Ar-

tiﬁcial Intelligence, pages 113–128.

Kamble, M., Shrivastava, P., and Jain, M. (2021). Digi-

tized spiral drawing classiﬁcation for parkinson’s dis-

ease diagnosis. Measurement: Sensors, 16:100047.

Khatamino, P., Cant

urk,

I., and

Ozyılmaz, L. (2018). A

deep learning-cnn based system for medical diagno-

sis: An application on parkinson’s disease handwrit-

ing drawings. In 2018 6th International Conference

on Control Engineering & Information Technology

(CEIT), pages 1–6. IEEE.

Kotsavasiloglou, C., Kostikis, N., Hristu-Varsakelis, D., and

Arnaoutoglou, M. (2017). Machine learning-based

classiﬁcation of simple drawing movements in parkin-

son’s disease. Biomedical Signal Processing and Con-

trol, 31:174–180.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. Advances in neural information processing

systems, 25.

Lahmiri, S. and Shmuel, A. (2019). Detection of parkin-

son’s disease based on voice patterns ranking and op-

Spiral Drawing Test and Explainable Convolutional Neural Networks for Parkinson’s Disease Detection

451

timized support vector machine. Biomedical Signal

Processing and Control, 49:427–433.

Letanneux, A., Danna, J., Velay, J.-L., Viallet, F., and Pinto,

S. (2014). From micrographia to parkinson’s disease

dysgraphia. Movement Disorders, 29(12):1467–1475.

Li, Z., Lin, Y., Elofsson, A., Yao, Y., et al. (2020). Protein

contact map prediction based on resnet and densenet.

BioMed research international, 2020.

Liu, T., Chen, T., Niu, R., and Plaza, A. (2021). Land-

slide detection mapping employing cnn, resnet, and

densenet in the three gorges reservoir, china. IEEE

Journal of Selected Topics in Applied Earth Observa-

tions and Remote Sensing, 14:11417–11428.

Mattison, H. A., Stewart, T., and Zhang, J. (2012). Apply-

ing bioinformatics to proteomics: Is machine learning

the answer to biomarker discovery for pd and msa?

Movement Disorders, 27(13):1595–1597.

Mercaldo, F. and Santone, A. (2021). Transfer learning for

mobile real-time face mask detection and localization.

Journal of the American Medical Informatics Associ-

ation, 28(7):1548–1554.

Poewe, W., Seppi, K., Tanner, C. M., Halliday, G. M.,

Brundin, P., Volkmann, J., Schrag, A.-E., and Lang,

A. E. (2017). Parkinson disease. Nature reviews Dis-

ease primers, 3(1):1–21.

Potgieser, A. R., Roosma, E., Beudel, M., de Jong, B. M.,

et al. (2015). The effect of visual feedback on writ-

ing size in parkinson’s disease. Parkinson’s Disease,

2015.

Rizzo, G., Copetti, M., Arcuti, S., Martino, D., Fontana,

A., and Logroscino, G. (2016). Accuracy of clinical

diagnosis of parkinson disease: a systematic review

and meta-analysis. Neurology, 86(6):566–576.

Rosenblum, S., Samuel, M., Zlotnik, S., Erikh, I., and

Schlesinger, I. (2013). Handwriting as an objective

tool for parkinson’s disease diagnosis. Journal of neu-

rology, 260:2357–2361.

San Luciano, M., Wang, C., Ortega, R. A., Yu, Q.,

Boschung, S., Soto-Valencia, J., Bressman, S. B.,

Lipton, R. B., Pullman, S., and Saunders-Pullman,

R. (2016). Digitized spiral drawing: A possible

biomarker for early parkinson’s disease. PloS one,

11(10):e0162799.

San-Segundo, R., Navarro-Hell

ın, H., Torres-S

anchez, R.,

Hodgins, J., and De la Torre, F. (2019). Increasing ro-

bustness in the detection of freezing of gait in parkin-

son’s disease. Electronics, 8(2):119.

Stanley, K., Hagenah, J., Br

uggemann, N., Reetz, K., Sev-

ert, L., Klein, C., Yu, Q., Derby, C., Pullman, S.,

and Saunders-Pullman, R. (2010). Digitized spi-

ral analysis is a promising early motor marker for

parkinson disease. Parkinsonism & related disorders,

16(3):233–234.

Viteckova, S., Kutilek, P., Svoboda, Z., Krupicka, R.,

Kauler, J., and Szabo, Z. (2018). Gait symmetry mea-

sures: A review of current and prospective methods.

Biomedical Signal Processing and Control, 42:89–

100.

Zham, P., Arjunan, S. P., Raghav, S., and Kumar, D. K.

(2017a). Efﬁcacy of guided spiral drawing in the

classiﬁcation of parkinson’s disease. IEEE journal of

biomedical and health informatics, 22(5):1648–1652.

Zham, P., Kumar, D. K., Dabnichki, P., Poosapadi Arjunan,

S., and Raghav, S. (2017b). Distinguishing different

stages of parkinson’s disease using composite index of

speed and pen-pressure of sketching a spiral. Frontiers

in neurology, page 435.

Zhang, C., Benz, P., Argaw, D. M., Lee, S., Kim, J.,

Rameau, F., Bazin, J.-C., and Kweon, I. S. (2021).

Resnet or densenet? introducing dense shortcuts to

resnet. In Proceedings of the IEEE/CVF winter con-

ference on applications of computer vision, pages

3550–3559.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

452