CHAPTER - 2

LITERATURE SURVEY & EXISTING SYSTEM

2.1     INTRODUCTION

Machine learning techniques have shown promise in the early detection of PD. In recent years, several studies have been conducted to develop machine learning models for the detection of PD using various types of data, such as voice, gait, and handwriting, among others.

In Aleksandr Talitcki et al target the recognition of exercises enabling precise PD detection. The author created a collection of 15 tasks to help individuals perceive the symptoms that are uniquely linked to PD.The exercises are grouped into 4 classes they are Gross Motor (GM), Clinical Evaluation (CE), Fine Motor (FM) and Tremor at Rest. The first of the Gross Motor (GM) is used to measure the subject's movements, which are frequently employed in daily activities. The clinical evaluation (CE) entails customary techniques that neurologists employ to identify the tremor. For the purpose of evaluating the subject's fine alignment, fine motor (FM) tests are required. While the body is restrained, Tremor at Rest is employed to help expose the tremor.

Each measurement was carried out at 100 Hz and transmitted through BLE to an Intel Next Unit of Computing (NUC) computer that was running by the specialized Python software. Although it is advised, applying the entire exercises can be disruption and long delayed for both the patient and the practitioner. The primary limitations of this study are wireless communication and the possibility that people with the fourth stage of PD might find it difficult to wear even the smallest sensing device. The wireless transmitter is also the device on board that consumes the more current, and the battery can run out in about an hour.

In  Juan C. Pérez-Ibarra et al proposed two online causal algorithms for gait event detection and phase segmentation. A simple strategy is to split the gait into two things: Position and Motion. The author suggestion is to alternate between various linear classifiers, a single one for each stage, enabling the classifier changes whenever the FSM transitions to a new state. The author employed a mix of the Simulated Annealing (SA) and Genetic meta-heuristics techniques (GA). The author has created a dyad of wearable gait

analysis gadgets. They were created to be worn over the subjects' normal footwear for regular activities. They created an upgraded model of our wearable IMU-based gadget that records and measures gait information.

In  Laiba Zahid et al proposed move learning based procedures and AI based procedure is used. In AI based procedure utilized help vector machine (svm) and arbitrary woodland calculation is utilized. The most elevated 99.7 percent exactness on vowel \o\ and read text is noticed utilizing a multifacet discernment. While 99.1 percent exactness saw on vowel \i\ profound elements utilizing irregular backwoods. The proposed procedure beats the current methods on the pc-Gita dataset for Parkinson's infection location.

In Liaqat Ali et al said to dealt with two significant problems, first issue biasedness by imbalanced informational collection for that they utilized irregular under sampling technique using machine learning. What's more, second issue is low pace of arrangement accuracy, to further develop it they utilized flowed learning framework method.it was likewise seen that the proposed flowed framework works on the solidity of traditional has 3.3 percent by Ada-boost model and decreases its intricacy. Moreover, the flowed framework accomplished characterization precision of 76.44%, awareness of 70.94% and explicitness of 81.94%. Restrictions are Not more exact contrasted with different calculations.

In Federica Amato et al separated 60 elements from previous handled vocal signals and involved them as contribution to a few AI models. 92 % exactness in 10-crease cross-approval utilizing SVM. Limitations are Just the SVM calculation works among all the diff calculation utilized in the undertaking. Results show that the 7 elements chose for SVM accomplishes great generally speaking exactness of 83.33%, great location rate for Parkinson`s infection of 75% and low bogus positive after effects of 16.67%. There is a model for recognizing Parkinson's utilizing voice. The diversions in the voice will affirm the side effects of Parkinson's sickness. This task showed 73.8% effectiveness. In our model, a colossal measure of information is gathered from the typical individual and furthermore recently impacted individual by Parkinson's sickness. There is a model for recognizing Parkinson's utilizing voice. The diversions in the voice will affirm the side effects of Parkinson's sickness. This task showed 73.8% effectiveness. In our model, a colossal measure of information is gathered from the typical individual and furthermore recently impacted individual by Parkinson's sickness.

In [16], N. P. Narendra et al proposed two approaches to recognize the disease first approach is standard pipeline approach in that they used three methodology, they are system structure, baseline features and gif strategies and glottal features. Second approach is beginning to end approach called Pipeline approach. From the ordinary pipeline structures, the most raised grouping accuracy (67.93%) was given by blend of benchmark what's more, QCP-based glottal components. From the very outset to complete systems, the most raised accuracy (68.56%) was given by the structure arranged using QCPbased glottal stream signals. Regardless of the way that gathering exact-nesses were honest for all systems, the audit is engaging as the extraction of voice source information was seen as most convincing in the two strategies.

K. Polat proposed the cross variety approach to investigate Parkinson's contamination using talk signals. It combines resounding and time-repeat parts for strong component extraction. Ability of Convolutional Mind Associations for voice signal assessment is researched. The author used two strategies they are SMOTE(Synthetic Minority Over-testing Strategy) and sporadic forest area. SMOTE approach, the amount of tests for minority class in the PD dataset has been falsely extended to change the dataset.

Precision of simply irregular backwoods is 87.037% combination of both Destroyed and irregular timberland precision is 94.89%. Accuracy is the degree of closeness between a measurement and its true value. Precision is the degree to which repeated measurements under the same conditions show the same results. The sporadic forest area technique achieved a higher responsiveness than the decision tree model. Subsequently, it is fitting to cultivate a show to helpfully recognize starting stage PDD considering the PD-MCI gauge model made in this survey, to spread out individualized seeing to follow high-risk social occasions. People with Parkinson's disease may lose their sense of smell and experience sleep disruptions during the rapid eye movement sleep period. It is estimated that 1% of the population over the age of 60 suffers with paralysis against. . Parkinson's disease symptoms vary from person to person. Early warning indicators are subtle and go overlooked. Symptoms normally start on one side of your body and progress to the opposite side before affecting both edges.

2.2     EXISTING SYSTEM & DRAWBACKS

In PD can only be recognized at secondary stages under the current system, making disease prevention is challenging and potentially fatal. In this system, the doctor must manually assess the patient's symptoms by taking the MRI scans which describes the nigral structure of patients to identify the disease. Each person has a unique set of sickness symptoms. Based on the brain condition and dopamine levels, doctors with special training in brain disorders can identify the disease. Doctors typically use MRIs, SPECT scans, and blood tests. They are also known as image tests. In order to create 3D images of the brain that allows the doctor to understand the status of the brain by the SPECT which uses nuclear radio power. To provide an image of the organs, MRI uses radio frequency bands and magnetic flux. There are various systems for the detection of Parkinson's disease, ranging from clinical assessment tools to computer-based diagnostic methods. Some of the most commonly used methods are:

1)   Clinical Assessment: The most common method for diagnosing Parkinson's disease is through clinical assessment by a neurologist. The neurologist will evaluate the patient's medical history, physical examination, and symptoms to determine if they meet the criteria for a diagnosis of Parkinson's disease.

2)   Rating Scales: There are several rating scales available that can be used to assess the severity of Parkinson's disease symptoms. These scales include the Unified Parkinson's Disease Rating Scale (UPDRS) and the Hoehn and Yahr Scale.

i. Unified Parkinson's Disease Rating Scale (UPDRS): This is a clinical tool used to evaluate the severity of Parkinson's disease symptoms. 

It consists of four parts: 

1. mentation, behavior, and mood, 

2. activities of daily living, 

3. motor examination, and 

4. complications of therapy.

ii. Hoehn and Yahr Scale: This is another clinical tool used to assess the severity of Parkinson's disease based on the degree of disability and motor impairment. It consists of five stages, ranging from mild symptoms on one side of the body to severe disability on both sides.

3)   Montreal Cognitive Assessment (MoCA): This is a cognitive screening tool used to detect mild cognitive impairment in patients with Parkinson's disease. It consists of various tasks such as memory, attention, language, and visuospatial abilities.

4)   Computer-based methods: Various computer-based methods have been developed for the detection of Parkinson's disease, such as machine learning , deep learning models, and neural networks. These methods use data from various sources such as voice recordings, gait analysis, and eye movements to detect Parkinson's disease with high accuracy.

5)   Imaging techniques: Imaging techniques such as positron emission tomography (PET) and magnetic resonance imaging (MRI) have been used to detect changes in the brain that are associated with Parkinson's disease. These techniques can detect changes in the dopamine system and other brain structures affected by the disease.

6)   Electrophysiological Techniques: Electrophysiological techniques such as electroencephalography (EEG) and electromyography (EMG) can be used to measure brain activity and muscle activity, respectively, to help diagnose Parkinson's disease.

7)   Blood Tests: Blood tests can be used to detect specific biomarkers associated with Parkinson's disease, such as alpha-synuclein and DJ-1.

8)   Speech Analysis: Changes in speech patterns can be an early sign of Parkinson's disease, and speech analysis tools can be used to detect changes in speech patterns that may be indicative of the disease.

2.2.1 Drawbacks

There are several limitations of existing systems for the detection of Parkinson's disease, including:

1)   Lack of accuracy: Existing systems for Parkinson's disease detection may not be accurate enough to provide reliable results. This can be due to factors such as variability in symptoms, differences in disease progression among patients, and the sensitivity of the detection method.

2)   Limited availability: Many existing systems for Parkinson's disease detection are only available in specialized clinics or research centers, making them inaccessible to a large number of patients.

3)   Costly: Some systems for Parkinson's disease detection can be costly, making them prohibitive for patients who cannot afford the expense.

4)   Lack of standardization: There is currently no standardization in the way that Parkinson's disease is diagnosed and detected, which can lead to inconsistencies in the results obtained from different systems.

5)   Limited scope: Some existing systems for Parkinson's disease detection focus only on certain symptoms or aspects of the disease, which may not provide a comprehensive picture of the patient's condition.

6)   Invasive methods: Some existing systems for Parkinson's disease detection require invasive procedures such as lumbar punctures or brain imaging, which can be uncomfortable and risky for patients.

7)   Subjectivity of Diagnosis: Currently, the diagnosis of PD relies heavily on clinical assessments and subjective evaluations by neurologists, which can lead to inconsistencies and inaccuracies in diagnosis.

8)   Lack of Early Detection: PD symptoms often do not appear until significant damage has already occurred in the brain, which can limit the effectiveness of treatments and interventions.

9)   Lack of Personalization: Current diagnostic tests do not take into account individual differences in disease progression or response to treatment, which limits the ability to personalize treatment plans for PD patients.

2.3     PROPOSED SYSTEM

The proposed system for the detection of Parkinson's disease using Random Forest and XGBoost would involve the use of spiral drawings and voice deflections as input data for the model. Firstly, the patient would be asked to draw a spiral on a piece of paper using a pen or pencil. The drawing would then be scanned and processed to extract features such as the number of turns, the length of each turn, and the amount of tremors during the drawing. These features would then be fed into the Random Forest and XGBoost models to predict the likelihood of the patient having Parkinson's disease.

In addition to spiral drawings, the patient's voice would also be recorded and analyzed for signs of Parkinson's disease. Features such as tremors, pitch, and volume would be extracted from the voice recording and used as additional input for the models.

The XGBoost model would be trained on a large dataset of spiral drawings and voice recordings from both Parkinson's disease patients and healthy individuals. The models would then be used to predict whether a new patient is likely to have Parkinson's disease based on their spiral drawing and voice recordings.

To develop a machine learning model for PD detection using voice recordings, a dataset of voice recordings from both healthy individuals and those with PD would be needed. The dataset would need to be labelled to indicate which recordings were from individuals with PD and which were from healthy individuals.

The next step would be to pre-process the data, which might involve feature extraction to identify relevant characteristics of the voice recordings that could be used as inputs to the

machine learning model. These features could include measures of pitch, loudness, and other acoustic properties of speech [18]. Once the data has been pre-processed, it could be used to train a machine learning model using techniques such as logistic regression, decision trees, or neural networks. The model would be trained to recognize patterns in the voice recordings that are indicative of PD.

Finally, the trained model could be tested on a separate dataset of voice recordings to evaluate its performance. This evaluation might involve metrics such as accuracy, precision, recall, and F1 score to assess how well the model is able to correctly identify individuals with PD based on their speech patterns.

The proposed system overcomes the limitations of the previous models with the high accuracy Spiral data and speech datasets are both inputs for this system. To measure the deflections in the voice, linear regression and XGBoost is applied to the voice dataset. Due to a variety of factors, like the inability to hold a pen properly or fear, not everyone is able to draw the spiral shape. The spiral model's issues are resolved by the voice model. The system is hence highly accurate. The massive amount of spiral and audio data from healthy and ill patients are collected. Training and testing sets were derived from these datasets. When used separately, spiral and voice models are unable to produce accurate outcome. So, each model has its own drawbacks. This combines the results of the spiral and voice models to identify the illness.

2.1     FEASIBILITY STUDY

The machine learning feasibility study phase is used to dig into the data and quickly conduct experiments to establish baseline performance on a task. Some key considerations when conducting a feasibility study for detecting PD using machine learning are given below :

1. 1. 1. Data collection: Voice and spiral datasets were collected kaggle repository.

      2. Data preprocessing: Once the data has been collected, it should be preprocessed. Preprocessing involves cleaning and transforming the data into a suitable format for machine learning algorithms. It involves removing missing values, outlier detection, normalization, and feature selection.

      3. Machine learning algorithms: Machine learning algorithms such as Random Forest and XGBoost to detect PD. Random forest is applied on spiral dataset and XGboost was applied on voice dataset. These algorithms are powerful and have been proven effective in detecting complex patterns in large datasets.

      4. Model evaluation: We can evaluate the performance of the machine learning model to determine its accuracy in detecting PD. The model should be evaluated using various metrics such as accuracy, confusion matrix.

      5. Clinical validation: The final step in conducting a feasibility study is to clinically validate the machine learning model. Clinical validation involves testing the model on a separate dataset to determine its generalizability and accuracy in real-world settings.

2.4.1      Economical Feasibility

The economic feasibility of using machine learning algorithms like Random Forest and XGBoost for detecting PD using spiral drawings and voice deflections can be analyzed from various perspectives [17]:

• Cost of data collection: Collecting data for training and testing the machine learning models can be expensive. In the case of PD detection, the cost of collecting spiral drawings and voice deflections from patients can be high. However, this cost can be reduced by using digital platforms that allow remote data collection and sharing.

• Cost of model development: Developing machine learning models requires significant expertise and resources. Hiring data scientists, machine learning engineers, and software developers can be expensive. However, there are open-source tools and libraries available that can be used to build models without incurring significant costs.

•  Accuracy of the model: The accuracy of the machine learning models is crucial in determining their economic feasibility. If the models are not accurate enough, then they will not be useful in clinical settings. The Random Forest and XGBoost algorithms have shown promising results in detecting PD using spiral drawings and voice deflections, with accuracies of up to 90%.

• Cost of deployment: Deploying the machine learning models in clinical settings can be expensive. The cost of hardware, software, and maintenance needs to be considered. However, the cost can be reduced by using cloud-based solutions that allow remote deployment and maintenance.

•  Economic benefits: The economic benefits of using machine learning models for detecting PD are significant. Early detection and treatment can reduce the progression of the disease, improve the quality of life of patients, and reduce healthcare costs. In addition, machine learning models can help in identifying patients who are at high risk of developing PD, enabling early intervention and prevention.

The economic feasibility of using machine learning algorithms like Random Forest and XGBoost for detecting PD using spiral drawings and voice deflections depends on several factors, including the cost of data collection, model development, accuracy of the model, cost of deployment, and economic benefits. However, the potential benefits of early detection and treatment of PD outweigh the costs, making it a promising area for research and development.

2.4.2 Technical Feasibility

The Technical Feasibility behind the detection of PD using machine learning techniques such as XGBoost using spiral drawings and voice deflections is well established. PD affects the central nervous system, which controls the movement of the body. One of the key symptoms of PD is tremors, which can be observed in the patient's handwriting and voice[18].

Spiral drawings can be used to analyze the patient's handwriting, and voice recordings can be analyzed to detect changes in speech patterns. Machine learning algorithms such as Random Forest and XGBoost can be trained on large datasets of spiral drawings and voice recordings to identify patterns that are indicative of PD. These algorithms can then be used to classify new data as either Parkinson's positive or negative. The accuracy of the classification model can be improved by including additional features such as age, gender, and other medical history [21].

Overall, the technical feasibility of using machine learning algorithms to detect PD using spiral drawings and voice deflections is well established, and this approach has the potential to improve the accuracy and speed of diagnosis for PD, which will help in the getting more correct predictions of the disease. The technical feasibility is going to give the better results for the diagnosis of the disease [22].

2.4.3   Social Feasibility

The Social Feasibility of using machine learning models such as XGBoost for the detection of PD using spiral drawings and voice deflections depends on several factors includes the following [23]:

• • • Accessibility: The availability of technology and resources required for the detection of PD using machine learning models may vary across different regions and populations. Therefore, it is important to ensure that the technology is accessible to all communities, regardless of their socioeconomic status or geographic location.

    • Acceptance: The use of machine learning models for medical diagnosis may not be readily accepted by some individuals, particularly those who may not trust technology or may prefer traditional methods of diagnosis. Therefore, it is important to educate and inform the public about the benefits and limitations of using machine learning models for medical diagnosis.

    • Privacy: The use of machine learning models for medical diagnosis raises concerns about privacy and data security. It is important to ensure that patient data is protected and used only for the intended purpose of medical diagnosis, and that patients are informed about how their data is being used.

    • Accuracy: The accuracy of machine learning models for the detection of PD using spiral drawings and voice deflections is critical to their social feasibility. It is important to validate the accuracy of these models across different populations to ensure that they are effective and reliable.

Proposed projects are beneficial only if they can be turned out into information system. That will meet the organization’s operating requirements. Operational feasibility aspects of the project are to be taken as a crucial a part of the project implementation. The important issues raised are to check the operational feasibility of a project includes support, application specific and management issues [24]. Overall, the social feasibility of using machine learning models for the detection of PD using spiral drawings and voice deflections depends on their accessibility, acceptance, privacy, and accuracy. By addressing these factors, we can ensure that these models are effective, reliable, and socially acceptable.

2.2     DATASET

2.5.1 Voice Dataset

Voice dataset is collected from Kaggle [25]. This contains 195 persons containing 147 `Parkinson, 48 healthy people data. This dataset contains 24 features. They are

• • • • MDVP:Fo(Hz) : Average fundamental frequency of the voice. The average of all the fundamental frequency values that were derived from different periods. Areas where voices break apart are ignored.

    • MDVP:Fhi(Hz) : Maximum fundamental vocal frequency. the highest value for fundamental frequency among all retrieved period-to-period. Areas with voice breaks are not included. The pitch extraction range is set to either search for periods from 200 to 1000 Hz or from 70 to 625 Hz. As a result, determining a fundamental above 625 Hz falls outside the "normal" range.

    • MDVP:Flo(Hz) : Minimum fundamental frequency for voice. the lowest of all extracted fundamental frequency values for period-to-period. Areas where voices break up are excluded. The lowest fundamental for the specified time period is taken out and presented in the form of Flo. The pitch extraction range, however, is set to either search for periods between 200 and 1000 Hz or between 70 and 625 Hz.

    • MDVP:Jitter(%) : Relative evaluation of the period-to-period (very short-term) variability of the pitch within the analyzed voice sample. Voice break areas are excluded. The pitch of the voice can vary for a number of reasons. Cycle-to-cycle irregularity can be associated with the inability of the vocal cords to support a periodic vibration for a defined period. Usually these types of variations are random. They are typically associated with hoarse voices.

    • MDVP:Jitter(Abs) : Absolute jitter (/usec/) is a measurement of the pitch period's period-to period variability within the analyzed speech sample. Areas where voices break up are excluded. The voice's pitch can change for a variety of causes. Cycle-to- cycle variability may be related to the vocal cords' failure to sustain a periodic vibration for a predetermined amount of time. This kind of fluctuation typically occurs at random. Hoarse voices are frequently connected to them.

    • MDVP:RAP : The relative average perturbation is %. a three-period smoothing factor was used to evaluate the period-to-period fluctuation of the pitch within the studied voice sample. Areas where voices break up are excluded. The inability of the vocal cords to support a periodic vibration with a known period can be linked to cycle-to- cycle inconsistency. RAP may be higher in voices that are hoarse or breathy.

    • MDVP:PPQ : Pitch Period Perturbation Quotient /%/ - relative assessment of the pitch's period-toperiod fluctuation within the studied voice sample using a 5 period smoothing factor. The inability of the vocal folds to support a periodic vibration with a known period can be linked to cycle-to-cycle variability. Voices that are hoarse or breathy may have a higher PPQ.

    • Jitter:DDP : Several measures of variation in fundamental frequency.

    • MDVP:Shimmer : Shimmer Percent /%/ - /%/ Shimmer Percent - relative assessment of the peak-to-peak amplitude variability within the studied voice sample from period to period (very short term). The inability of the cords to support a periodic vibration for a specific amount of time and the existence of turbulence noise in the speech signal can both contribute to cycle-to-cycle variability of loudness.

    • MDVP:Shimmer(dB) : Shimmer in dB /dB/ - Evaluation in dB of the period-to-period (very short-term) variability of the peak-to-peak amplitude within the analyzed voice sampleAreas where voices break up are excluded. The inability of the vocal folds to sustain a periodic vibration for a specific amount of time and the existence of turbulent noise in the voice signal can both be factors in cycle-to-cycle irregularity of loudness.

This kind of fluctuation frequently occurs at random. Voices that are breathy and harsh are frequently associated with it.

• • Shimmer:APQ3 : The Smoothed Amplitude Perturbation Quotient is %. relative assessment of the peak-to-peak amplitude variability within the studied speech sample over the short or long term at a smoothing factor specified by the user. The smoothing factor's factory configuration is 55 periods (providing relatively long- term variability; the user can change this value as desired). Areas where voices break up are excluded. They frequently have harsh and breathy voices.

  • MDVP:APQ : Amplitude Perturbation Quotient /%/ - relative assessment of the peak- to-peak amplitude period-to-period variability within the examined voice sample at a smoothing of 11 periods. Areas where voices break up are excluded. There are many different factors that can affect the voice's amplitude. The incapacity of the cords to maintain a periodic vibration with a known period and the existence of turbulent noise in the speech signal can be linked to cycle-to-cycle irregularity of amplitude. Hoarse and breathy voices typically have a higher APQ.

  • Shimmer:DDA : Several measures of variation in amplitude

  • NHR : Two Noise to Harmonic Ratio (NHR) is another useful measure. This can be routinely measured using MDVP. For a signal that can be assumed to be periodic, the signal-to-noise ratio will be equal to the harmonics-to-noise ratio.

  • HNR : and it is this that I prefer to calculate when using Praat. Praat declares that a healthy voice phonating /a/ or /i/ should have an HNR of 20, and 40 for the phonation of the vowel /u/. Consequently, an HNR below 20 is considered to be a measure of noticeable hoarseness.

  • Status : Health status of the subject (one) - Parkinson's, (zero) - healthy

  • RPDE : Recurrence Period Density Entropy.

  •  D2 : It is a nonlinear dynamical complexity measures.

  •  DFA : Signal fractal scaling exponent

  •  spread1, spread2, PPE : Three nonlinear measures of fundamental frequency variation

The spiral dataset also collected from Kaggle repository. In training dataset contains 72 persons data (36 parkinson,36 healthy). In testing dataset contains 30 persons data (15 parkinson,15 healthy)