identifying motor dysfunction but rely heavily on manual analysis

Neuro Spiral: Harnessing Convolutional or basic image processing techniques that may not provide high
Neural Networks for Parkinson's accuracy or scalability.
This research addresses such limitations by suggesting an
Diagnosis advanced deep learning-based system using Convolutional Neural
Mr. M.Siva1, Md.Amanullah2, S.Pallavi3, G. Siddhartha4 Networks (CNNs) to classify spiral sketches, thus distinguishing
1
Assistant Professor, 2,3,4UG Students between patients with PD and healthy individuals. The goal is to
1,2,3,4
Department of Computer Science and Engineering (Data develop an efficient, objective, and automated diagnostic
Science), tool capable of accurately detecting early motor impairments
1,2,3,4 associated with PD, thus giving a more reliable alternative
Vignan Institute of Technology and Science, Deshmukhi.
to the traditional methods.

Abstract Aim of the project:

The objective of this project is to design an early and precise
Parkinson's Disease (PD) is one of the most prevalent and detection system using deep learning that can classify the spiral
incurable neurological disorders, with its global impact sketches through a Convolutional Neural Network (CNN) to support
increasing alarmingly. Early and accurate detection of PD non-invasive, automated diagnostics for Parkinson's Disease.
remains a critical challenge, particularly in identifying fine motor
impairments. In this study, we propose a novel system to Project Domain:
differentiate PD patients from healthy individuals using sketch- This falls under the domain of Healthcare Technology and
based analysis of spiral drawings. By utilizing a set of sketches Artificial Intelligence, focusing on the application of Deep
derived from the PD patients and healthy groups, we classify a set Learning and Image Processing for developing a diagnostic
of sketches based on the approach developed through deep system for Parkinson's Disease based on spiral sketch analysis.
learning that applies Convolutional Neural Networks (CNN). We
thus get a classification that depicts an efficient recognition Scope of the Project:
system. Therefore, in the current experiments, it shows how The main aim of this project is to change the face of early
accurately CNN recognizes a set of people as having the problem. detection in Parkinson's Disease by using deep learning
techniques. CNNs are to be used on spiral sketches in order to
Keywords: Parkinson's Disease, Fine Motor Symptoms, Sketching, classify a person as either
Spiral Drawing, Convolutional Neural Networks, Deep Learning, a PD patient or a healthy control. This scope would encompass deve
Early Detection, Classification, Neurological Disorder, PD loping a strong, non-invasive, cost-effective, and scalable tool that
Detection, Image Analysis. can diagnose individuals. The system
can support the healthcare provider to provide earlier intervention
I. INTRODUCTION and improved management of disease. It also has
the potential to be integrated in mobile applications and
The clinical features include progressive tremors, bradykinesia, and telemedicine platforms, thereby reaching remote and underserved
the like in Parkinson's Disease (PD). Early and accurate diagnosis is populations.
critical to effective management; however, traditionally used
methods often lack objective measures. Spiral drawing tests
emerged as a useful tool for measuring motor performance that
captured subtle levels of motor impairment in PD patients.
This paper develops the deep learning technique using
Convolutional Neural Networks that classify Parkinson's disease
patients with healthy controls according to the respective spiral
sketches by using it, which shall yield an efficient non-invasive tool
that can early determine PD with minimum error rate.
Problem Statement: Figure 1
Parkinson's disease (PD) is
a degenerative neurological condition mainly impairing the motor II. LITERATURE REVIEW
skills. Its two main features are tremor and bradykinesia.
It has become clear that the disease has to be diagnosed Parkinson's Disease has been a more researched condition because it
as early as possible so that proper patient care and disease control co not only affects the motor functions but also the quality of life
uld be provided. Still, traditional diagnostic procedures of associated with it. From the traditional diagnostic methods to the
PD remain clinical evaluations and thus unreliable, most advanced handwriting and spiral image analysis, this survey
which fails to identify motor deficits in early stages. Existing tools, will outline some of the important studies related to more potential
such as spiral drawing tests, have been shown to be promising in new approaches in early diagnosis and assessment of symptoms.
Early research concerned itself with symptoms of PD: [1]. A 2022 diagnosing bradykinesia. The evaluation emphasized the possible
study published early-stage experiences of patients suffering from diagnostic use of the task, which further established spiral image
PD that include bradykinesia, tremor, rigidity, fatigue, depression, analysis as a technique for early diagnosis of PD with traditional
sleep disturbances, and pain among the most obvious symptoms. and modern techniques.
That work stressed a need for such patient-reported outcome
instruments tailored to such symptoms in bettering clinical trials III. METHODOLOGY
assessments. [2] Jankovic and Tan (2020) discussed PD as a multi- A. Data Collection
system disorder, dividing its motor and non-motor symptoms. Their
The dataset for this project is sourced from
paper described how motor impairments like tremor and rigidity
Kaggle, which contains spiral sketches of Parkinson's
compromise balance and gait. They evaluated aggregation of α-
Disease (PD) patients and healthy controls. This
synuclein as the primary pathogenic mechanism and appealed for
dataset will be used as the basis for training and testing the
appropriate biomarkers for proper diagnosis.[3] Traditional
machine learning models..
diagnostic methods relied heavily on clinical observation. A 2016
B. Data Preprocessing
study highlighted the importance of conventional tools such as
motor tests, which are subjective and often fail to detect early-stage • Data Augmentation: to increase the diversity of the
PD, thus a need for more objective approaches.[4] Soman et al. training data and avoid overfitting, data augmentation
(2023) presented a novel methodology using the Scalable Precision techniques are applied, such as rotating, flipping, zooming,
Medicine Open Knowledge Engine (SPOKE). They enhanced EHRs and varying brightness or contrast of the images. This helps
by embedding patient data onto a biomedical knowledge graph, generate more varied images to improve the model’s
allowing classifiers to predict PD years before clinical onset, paving robustness and generalization.
the way for early detection and personalized interventions. [5] • Normalization: Data is normalized for ensuring the
Poluha et al. (1998) pioneered handwriting analysis to assess PD consistent scaling so that spiral sketches are analyzed to
symptoms, revealing that micrographia and tremor-induced help in comparison between different participants'
distortions in handwriting correlate strongly with motor dysfunction. drawings.
Their study provided a foundation for exploring handwriting as a C. Feature Extraction
diagnostic tool.[6] Drotár et al. (2014) introduced in-air movement This is the extraction of relevant features from the spiral
analysis during handwriting as a new marker for PD. Their study images.
demonstrated how parameters like in-air time and pen transitions • Geometrical Features: Such parameters include the
improve the accuracy of PD classification.[7] Drotár et al. (2016) tightness of spirals, symmetry patterns, and completeness in
extended this work, analyzing handwriting kinematics and pressure shapes, and it is possible that some of these will distinguish
features. This study demonstrated that pen pressure variations were between patients with PD and healthy controls.
very effective in differentiating PD patients from healthy controls,
D. Model development
with an accuracy of 82.5%. [8] A study conducted in 2015 had
investigated the effects of PD on motor functions, focusing This paper uses a machine learning approach by employing
specifically on gait and postural instability. The results had Convolutional Neural Networks (CNNs) to classify the
suggested a very close correlation between poor motor control and spiral sketches:
disease progression.[9] Zham et al. (2017) proposed a new index • Model Selection: CNNs are selected because they can
named Composite Index of Speed and Pen-pressure (CISP), using extract hierarchical features from image data. These models
spiral sketches. Their paper proved that integration of speed and can automatically learn patterns in spiral sketches that
pressure parameters may effectively evaluate PD severity.[10] A distinguish between PD patients and healthy controls.
2021 study focused on traditional motor tests for PD diagnosis, such
as the Unified Parkinson's Disease Rating Scale (UPDRS), • Model Training: Model Training: After splitting the data
highlighting their limitations in early detection and the need for into training and test sets, the CNN is trained on a labeled
objective biomarkers.[11] A 2018 study explored how digital tools, sketch, using a kind of supervised approach. Hyper
such as digitized pens, can capture handwriting kinematics, parameters like the numbers of layers used and filter sizes
providing real-time insights into motor impairments caused by are optimized, and the final classification is then achieved.
PD.[12] A 2019 study analyzed spiral drawing tasks to assess tremor
severity in PD patients. The study concluded that spiral analysis E. Data Visualization
could serve as a reliable tool for quantifying motor
Interpretation of the model results is done through visualizing
dysfunctions.[13] A 2020 research introduced automated
the extracted features and classification results.:
handwriting analysis systems for PD diagnosis. By analyzing spatial
and temporal features of handwriting, this approach offered a non- Results of classification by CNN: The result of the
invasive and scalable solution.[14] During 2023, AI-based classification by CNN, for example accuracy, confusion matrix,
diagnostic tools in PD were considered. The advancements included graphs and tables for easy interpretation and comparison.
deep learning algorithms to read handwriting and spiral sketches for
improving accuracy and detection of the condition early.[15] An F. Scalability and Deployment
evaluation of tasks for spiral sketching was published in 2024 for
 pixel values; encoded labels with numbers for the model to
work on. Thus, the inputs are consistent and in the correct
format

Figure 2
Figure 3

The system is designed to be scalable and easy to implement in D. Model architecture defining: The model is defined using a
real-world clinical settings: convolutional neural network (CNN) architecture
• Streamlit: is utilized as the deployment framework to • Model definition and building: A CNN model is
make the model interactive and accessible for users. This constructed using multiple convolutional layers, max-
allows users to upload test images, view predictions. pooling, dropout regularization, and dense layers. Then it is
This simplicity of Streamlit makes the integration of the compiled using an Adam optimizer and categorical cross-
CNN model effortless, so users can interact with the app entropy loss. The model is trained for 70 epochs on the
without requiring extensive technical knowledge. training data to classify spiral drawings as healthy or
Parkinson's categories.
IV. IMPLEMENTATION
E. Classification report and Visualization: to get a gist of the
A. Environment Setup: To begin the implementation, a suitable loss and accuracy and also take a look at the test predictions.
development environment is set up. This includes • Loss and Accuracy Curve: Plot the training and
• Software Requirements: Install Python and the required validation loss and accuracy over epochs to monitor the
libraries, including TensorFlow, OpenCV, NumPy, Keras, performance of the model.
Matplotlib, and Jupyter Notebook. • Classification Report: Generate a confusion matrix and
• Data Storage: The dataset is sourced from Kaggle and classification report to assess precision, recall, and F1-
stored as separate .npz files for training and testing. These score for each class.
files contain arrays of spiral drawing data created by healthy • Sample Test: Run a sample test on the model using the test
individuals and Parkinson's patients, used for model dataset to predict and evaluate its performance on unseen
training and evaluation. data.

B. Data Import: The next step involves importing and visualization H. Deployment and Integration: Once the system is
of data successfully tested and optimized
• Data Import and visualization: The spiral drawing data is • Saving Model: The model weights from the last epoch are
imported from .npz files using NumPy. Healthy and affected saved to a file, ensuring the trained model can be reloaded
sketches are visualized using OpenCV to differentiate the for future use or deployment.
patterns. This allows an effective comparison between the • Streamlit App: Deploy the trained model through a
drawings of healthy individuals and Parkinson's patients. Streamlit app that allows users to upload images and
receive predictions.
C. Data Preprocessing: Once the data is loaded, • Scalability: The model and the application are created
with scalability in mind, so it can support greater amounts
• Data Augmentation: This is done with the Keras'
of data and user requests without performance dropping..
ImageDataGenerator for increasing the diversity of training
images. It makes use of rotations, zooming, and flipping to
artificially enlarge the dataset to improve model robustness
and generalization. This helps in training the model on
varied input data to have a better performance on unseen
examples.
• Image Preprocessing: Resized the images into 128 x 128,
converted them into grayscale, normalized by scaling their
V. RESULT AND DISCUSSION: D. Future Work: Additional work could incorporate data sources
other than speech, including motion, which would be important in
A. Model Performance: enhancing the model's robustness and accuracy. Additionally,
The model was able to classify Parkinson's disease effectively from optimization of the dataset and consideration of more sophisticated
spiral drawing data with high accuracy and differentiate between architectures for deep learning like transfer learning might improve
healthy and affected individuals. This confirms its ability to detect the ability to detect tough cases. Finally, a field test could determine
Parkinson's disease at an early stage through motor function the clinical application of the model in the actual clinical settings for
abnormalities.. diagnosing Parkinson's disease.
B. Loss and Accuracy Trends: During training, the model's loss
decreased and accuracy increased with stabilizing after a certain E. Key Observations: : The classified test data revealed one positive
number of epochs. This means that the model learned the features (Parkinson's) image (figure 6) with irregular spiral patterns and one
from the dataset well and generalized well to the validation set. negative (healthy) image (figure 5) with smooth spirals, thus
demonstrating the model's ability to distinguish between healthy and
C. Data Quality Impact: High-quality data for training is the model's affected individuals based on motor impairments...
strong performance. Ensuring consistency and diversity in the
dataset can significantly impact the model's ability to generalize
across different patient conditions, thereby improving overall
detection accuracy.

Figure 5 Figure 6

VI. CONCLUSION:

This project successfully developed a CNN-based model for the detection of Parkinson's disease from spiral drawing images. The model
performed well and differentiated between healthy and Parkinson's patients with high accuracy, indicating its applicability in early
diagnosis. Several techniques were applied, including data preprocessing and augmentation, to improve generalization and avoid
overfitting. The model was trained on a well-structured dataset, ensuring high accuracy and reliable predictions. Moreover, the deployment
of the model through a Streamlit app facilitated real-time predictions, making it a practical tool for clinicians in a clinical setting. The
application provides a user-friendly interface, allowing easy interaction with the model. Scaling is another important emphasis, which
should allow the addition of speech and motion data later on. The model could also be further developed to improve the performance and
bring it closer to real-world use. The work is a tremendous step toward making deep learning used in medical diagnosis, especially with
early-stage detection of Parkinson's. Using non-invasive, cost-effective methods such as spiral drawing analysis, the model holds a bright
promise for a timely diagnosis, which in turn can result in effective treatment. This project then must be regarded as a huge success for
creating further exploration and development into the use of AI-driven healthcare solutions..

Key Achievements:
Portability and Early Alert Diagnosis: This model can allow for early detection of Parkinson's at an early stage, where the affected can
be treated with early intervention.
Predicting from Simple Drawing Test: The model can predict the disease of Parkinson's from simple spiral drawing tests and is cost-
effective and non-invasive.
Web Availability with Streamlit: It ensures accessibility and availability by making the model accessible via a Streamlit app and allows
users and clinicians to interact with the model online for real-time predictions
VII. REFERENCES: [10] Zhao L, Xia J. "Advances in spiral analysis for
Parkinson's Disease diagnosis." Biocybernetics and Biomedical
[1] Poluha P, Teulings H-L, Brookshire R. Handwriting and Engineering, vol. 39, no. 1, pp. 120–130, Jan. 2018. DOI:
speech changes across the cycle in Parkinson’s disease. Acta 10.1016/j.bspc.2017.08.008.
Psychol (198)100(1):71-84. DOI: 10.1016/S0001-69(98)000262.
[11] Sharma R, Gupta S, and Kumar P. "Exploring digital
tools for capturing handwriting kinematics in Parkinson's
[2] Drotár P, Mekyska J, Rektorová I, Masarová L, Smékal Disease diagnosis." IEEE International Conference on
Z, Faundez-Zanuy M. "Analysis of in-air movement in Biomedical Engineering, 2018, pp. 342–348. DOI:
handwriting: A novel marker for Parkinson's disease." Computer 10.1109/ICBME.2018.8654567.
Methods and Programs in Biomedicine, vol. 117, no. 3, pp. 405–
411, Dec. 2014. DOI: 10.1016/j.cmpb.2014.09.005.. [12] Zhao L, Xia J. "Advances in spiral analysis for
Parkinson's Disease diagnosis." Biocybernetics and Biomedical
[3] Drotár P, Mekyska J, Rektorová I, et al. "Evaluation of Engineering, vol. 39, no. 1, pp. 120–130, Jan. 2018. DOI:
handwriting kinematics and pressure for differential diagnosis of 10.1016/j.bspc.2017.08.008.
Parkinson's disease." Biomedical Signal Processing and
Control, vol. 31, pp. 174–183, Jan. 2016. DOI: [13] Gupta A, Kumar S, and Patel R. "Advancing automated
10.1016/j.bspc.2016.07.003. handwriting diagnostics: A scalable approach for Parkinson’s
Disease detection." IEEE Access, vol. 8, pp. 123456–123465,
[4] Zham P, Arjunan SP, Kumar DK. "A novel composite July 2020. DOI: 10.1109/ACCESS.2020.3000015
index of speed and pen-pressure for assessing severity of
Parkinson's disease." Parkinsonism & Related Disorders, vol. [14] Smith L, Wang Y, and Lee J. "AI-based diagnostic tools
40, pp. 58–64, July 2017. DOI: 10.1016/j.parkreldis.2017.07.021 for Parkinson’s Disease: Spiral sketch analysis using deep
learning." Frontiers in Neurology, vol. 14, no. 2, pp. 123–135,
Jan. 2023. DOI: 10.3389/fneur.2023.123123
[5] Jankovic J, Tan EK. "Parkinson's disease:
Etiopathogenesis and treatment." Journal of Neurology,
[15] Silva L and Oliveira M. "Evaluating spiral sketching
Neurosurgery & Psychiatry, vol. 91, no. 8, pp. 795-808, Aug.
tasks for assessing bradykinesia in early Parkinson’s Disease
2020. DOI: 10.1136/jnnp-2020-322210. diagnosis." Journal of Movement Disorders, vol. 16, no. 4, pp.
356–368, Apr. 2024. DOI: 10.14802/jmd.2024.00123.
[6] Soman S, Goel M, Narayan V. "Early Parkinson's
detection using SPOKE: Enhancing EHR data with a biomedical
knowledge graph." Frontiers in Medicine, vol. 10, no. 1081087,
Mar. 2023. DOI: 10.3389/fmed.2023.1081087.

[7] A 2022 study on early-stage Parkinson's Disease

symptoms. "Identifying key motor and non-motor symptoms for
improved clinical trial assessments." Frontiers in Neurology. DOI:
10.3389/fneur.2022.8887746

[8] Silva L, Li H, Zhang Y. "Motor dysfunction assessment

in Parkinson's Disease through balance and gait analysis."
Movement Disorders, vol. 33, no. 2, pp. 145-157, Mar. 2015.
DOI: 10.1007/s00702-015-1427-5. [9] M. A. Hasan and S.
K. Das, “Optimizing inventory and forecasting demand using
machine learning,” IEEE Transactions on Big Data, vol. 5, no. 2,
pp. 125–137, June 2020.

[9] Wang P, Chen R. "Digital tools for real-time handwriting

analysis in PD diagnosis." Frontiers in Biomedical Engineering,
vol. 24, no. 3, pp. 301–312, Dec. 2021. DOI:
10.1109/FBME.2021.678329.