8.2.1 Machine Learning in PD Diagnosis

This section elaborates the primary techniques of categories of learning: Machine Learning, Autonomous learning, Neural Network, and self organizing networks used previously for PD findings. Additional information in the form of datasets, and set of images are also quoted with corresponding references.

The literature stated about PD patients’ handwriting and speech related noteworthy work. The assessment of tremor in related body parts and scale of severity of the disease is measured partially; the most conventional techniques of Neural Network and regression in recent times are implemented for the same. To improvise the accuracy of the disease finding, this chapter suggests the application of scientific finding tool of PD i.e. Spiral drawing up to some extent. There are evidences that some of the Machine Learning models and enhanced ensemble models work better in likelihoods of disease. Extensive feature engineering supports for exposing more details of symptoms of neurodegenerative disorder.

Authors in [16] stated the scientific method of three types of drawing the spiral and which are then known as test tool namely, Stability Test on a certain fixed coordinate, tracing of spiral in static manner, and tracing on spiral in dynamic manner.

The author [15] recommended the investigation of PD patient handwriting by applying feature selection algorithm and Machine Learning SVM method. This is one of the early works in identifying the magnitude of various in-air/on-surface hand movements while tracing the spiral, and these can be used in finding of motor disorder of neuro degenerative diseases. The outcomes of the paper confirmed the major impact of these kinds of movements during assessment of handwriting with 85.61% prediction accuracy in [14]. PD patients have performed eight different kinds of handwriting tasks and thus stored these handwritten samples into PaHaW PD database presented by the works of 2014. Later on, on the same dataset, the work [15] investigated some additional features. Extensively unique pressure features and some innovative stroke related features are extracted using dynamics of handwriting. The authors implemented three different types of classifiers: KNN, AdaBoost ensemble and SVM to reveal 81% classification accuracy. However, a limitation of the work is lack of consideration of related demographic features in support of handwriting test, and this leads somewhat less practical and effective diagnosis of PD patient. In lieu of this handwriting experiment, Static and dynamic manner of Archimedean spiral based scientific methods of drawing the spiral can be assessed as more pertinent and can be associated with the power of ML methods.

The handwriting as a powerful automatic marker to develop PD finding tool is one of the significant offerings of the authors of [17]. The authors have assumed the same PaHaw dataset and implemented ML classification framework. They obtained better specificity performance measures. The work not only illustrated binary discrimination of the patients into healthy and patients, but above this, the work in the said resource extended to group the patients, on the basis being the degree of severity of the illness, to further support early cure and medication to diseased persons. Though the work has provided extensive dataset, and clustering of patients based on illness, for proper classification, scientific tools of spiral can be extended for simple performance.

The authors [18] also analyzed the kinematic features of micrographia related with the disease when patients are performing the repetitive writing task. The experimental setup illustrated in the work was for samples of letter “e” and this is chosen due to its similarity with spiral pattern. To interpret the experiment, common size of initial five and final five letters were compared. Standard configured test of Wilcoxon signed-rank experimental test evaluated these handwriting samples. Later on they are also diagnosed with Kruskal–Wallis test.

Authors in [19] described the applications of well known machine learning frameworks using support vectors machine, multi-layer perceptron neurons for remotely tracking progress of Parkinson’s. It is difficult to take the patient every time to the clinic and so voice recordings are done periodically and the processed speech signals can be input to the general regression using most common neural network which outputs the Parkinson’s progress. This work is implemented on telemonitoring dataset that consists of the early stage PD patients’ voice recordings and twenty six attributes of voice are considered. This data set is computerized for remote assessment. Customary Performance measures of absolute error, MAE, MSE and correlation coefficients are identified to evaluate the prediction model. The conclusion findings are UPDRS can be mapped with vocal features in superior manner with LS-SVM as compared to other traditional ways. Vocal features are classical and non classical, and non linear in nature. Normalization of voice features after extraction is complex process, that needs wider range for normalization and accuracy level and the signal processing algorithms are needed. This is traditional mechanism to verdict of the neuro degenerative disease like Parkinson’s and in premature stage the symptoms may not be seen keenly in the patient. So UPDRS mapping with only voice features may not lead to accurate diagnosis. Greater computational burden and proneness to over fitting are the two common drawbacks of machine learning model for PD dataset, presented in this literature. To overcome the limitation of diagnosis by only few features either based on voice or drawings, it can be integrated and further instead of classifying them, they can be extended to check the severity of the disease, so that diagnosis and treatment can be done.

The authors [20] poised the thirty seven numbers of scripting samples of the patients which are taking regular medication for the disease. This group is of approximately thirty eight in age and sex-matched persons that did the seven scripting handwriting activities and tasks. Customary kinematic, spatial–temporal handwriting measures and some self defined novel handwriting actions based on entropy are the outcomes. Further signals are breakdown of the handwriting signals. Selected features are categorized with SVM classifier to demonstrate the 88% accuracy with highest sensitivity and specificity. However, this method only contributed to complexity of model without significantly improving the performance. The authors [21], stated that Parkinson’s disease is a progressive and chronic neurodegenerative disorder. The study revealed that as the dopamine-generating neurons degrade over the time or expire, the patients start experiencing difficulty in regular activities of speaking, writing, walking, or other simple tasks. Based on this research and other clinical research, it can be assured that by combining measures of all such activities, extensive data set can be prepared to analyze these symptoms. Patients’ symptoms of regular activities start degrading over time, thus severity in patients start increasing. Most of the work is based on identification of the disease. But this paper projected a methodology for the extrapolation of Parkinson’s disease severity using deep neural networks. Customary dataset of telemonitoring voice signals of patients widely available with UCI was used. It predicted severity by application of ‘TensorFlow’ deep learning library of python. The accuracy values obtained by this method are comparable to the previous research work. However, this method is based on single factor and based on some assumptions of clinical values.

Authors in [22] presented the technique to predict UPDRS score using machine learning algorithms which is relatively a rare approach and significant contribution to the field. Tremor signals collected by wrist wearable device is transformed to dataset and with dimensionality reduction, selected features were trained to automatic scoring of UPDRS. Four ML classifiers are implemented and comparative analysis is done. The authors suggested score on a scale of 0 to 4, describing absence of disease (0), slight tremor (1), mild amplitude changes in writing (2), moderate in amplitude (3) and severe (4). The authors have generated kinematic drawing features by recording hand movement of a person with wrist wearable device. Eighty five patients with PD have participated in the study. A video camera attached to the hand sensor has recorded the patients’ hand movements and two neurologists have evaluated the same. After mutual discussion of two neurologists, UPDRS score is mapped to the corresponding patient recording and stored in the dataset. This is one of the unique efforts for predicting disease severity.

8.2.2 Challenges of PD Detection

Parkinson’s disease is a neurodegenerative disease affecting slowly all daily activities of the person. Clinical tests available are fully based on expertise and may mislead to diagnosis. There is no confirmed medication available for the disease that can partially or completely cure the patient. At an early age, if it is detected there are costly medication methods and therapies available.

Machine learning and deep learning methods are significantly utilized in the Healthcare sector, where the symptoms are available in the form of several facts.

The possibilities of diagnosis of PD using soft computing can be explored for following means

• Brain scans, images
• Handwriting and drawing patterns
• Behavioral instances decoded in some facts
• Speech and Voice signal processing
• Dynamics and kinematics of movement of arms.

The challenges are variety of format representation, for images, signals, datasets and databases [8, 9]. Normalization generally cannot be standardized using clinical methods. Mixed symptoms cannot be categorized into actual causes.

To tackle this problem, this chapter has proposed a prototype work with dummy or prototype dataset, partially contributed by various sources and integrated for data science experiments. The assumptions are stated based on scientific mechanism.

8.2.3 Structuring of UPDRS Score

The researchers have been working for a long time to find out a method to quantify the ill effects. As provided in the literature [23], a well known score of PD, UPDRS is calculated with subscores of the activities: 1. mentation, behavior and mood, 2. activities of daily living 3. motor examination, and 4. complications of therapy such as dyskinesias, clinical fluctuations. These sections comprise a total of 199 features and every feature is quantified on the scale of 0 to 4. By summation of scores and scoring of patient based on questionnaire and symptomatic analysis, automatic score can be calculated and suggested. This is further termed as daily living stage as modification of further research and according to this, disease is classified into one to five stages. Complete follow-up of this neuro toolkit may give extensive exploration for disease severity. However, motor examination subscores are more quantifiable and a distinct significant feature as per our analysis and study of recent research work for the same healthcare sector.

This section elaborates the features on which automatic scoring is done for the given patients and healthy persons. A set of 40 spiral drawings [18] (drawn with dynamic spiral test and static spiral test) is considered. The dataset is comprising of csv files stating x,y,z coordinates, pen pressure, grip angle and timestamp. So in a drawing, csv file contains 1,000 recordings of drawing coordinates. These recordings are processed using kinematics formulae [14, 15] and a csv file of 40 records is prepared. Correlation heatmap is plotted with target as patient or healthy (dependent variable is consisting of 0 or 1) and 38 kinematic features are calculated.

Similarly, 16 biomedical voice features are recorded for these 40 persons, where the values are approximated from the UCI Machine voice dataset. This voice dataset consists of Total UPDRS and Motor UPDRS scores. Table 8.1 tabulated all the features, independent and dependent variables for the proposed work.

The dataset of 40 records with large number of features, correlation matrix, combined correlation matrix and standard deviation, mean absolute deviation, coefficient of variance are implemented by hyperparameter tuning of classification algorithms and the assumptions are also supported by the spiral drawings and voice recordings. Class labels for classification, are tabulated in Table 8.2.

8.3.1 Overview of Data Driven Intelligence

The Progression of Data-Driven Artificial Intelligence

The birthplace of Artificial Neural Network began back in 1940s, when the researchers used to talk about how neurons functioned in the human mind. At the popular workshops, the noteworthy man-made reasoning abilities like playing chess games and taking care of basic rationale issues were built up to understand the contribution of AI to the society and man-made problems.

In 1956, Perceptron a scientific model was proposed to mimic the sensory system of social learning with straight improvement. Next, a system model called Adaptive Linear Unit was created in 1959 and had been effectively utilized in useful applications, for example, correspondence and climate determining. The restriction of early computerized reasoning was likewise censured because of the trouble in taking care of non-straight issues, for example, exclusive OR (or exclusive NOR) order. With the advancement of Hopfield arrange circuit man-made reasoning ventured forward to the subsequent upsurge (the 1980s). The calculation for Back Propagation (BP) was suggested in 1974, in order to address non-straight problems in a complex neural system. Arbitrary mechanisms were introduced in the Hopfield system and Boltzmann (BM) was established in 1985 with the development and advancement of factual learning.

In any case, these conventional AI strategies require human aptitude to include extraction to decrease the element of information, and accordingly their presentation profoundly depends on the designed highlights. The introduction of profound gaining benefits not just from the rich gathering of customary AI strategies, yet additionally the motivation of factual learning. Profound learning utilizes information portrayal adapting instead of unequivocal designed highlights to perform undertakings. It changes information into unique portrayals that empower the highlights to be educated. Restricted Boltzmann Machine (RBM) was created in the year 1986 by getting the likelihood dispersion of Boltzmann Machine and the concealed layers were utilized as highlight vectors to describe the information. In the interim, Auto Encoder (AE) was proposed utilizing the Greedy learning calculation in layered manner to limit the misfortune work. In 1995, a neural system with coordinated topology associations between neurons, called Recurrent Neural Network (RNN), was proposed for include gaining from succession information. In 1997, an improved variant of repetitive Neural system, named Long transient Memory (LSTM), was proposed to handle the evaporating angle issue and manage complex time grouping information. In 1998, Convolution Neural Network (CNN) was advanced to deal with two dimensional information sources (for example picture), in which highlights learning were accomplished by stacking convolution layers and pooling layers As the various leveled structures of profound learning models getting further, model preparing and parameter enhancement become increasingly troublesome and tedious, in any event, prompting over fitting or nearby advancement issues. Numerous endeavors were made to grow profound learning models, yet no agreeable execution was accounted for before 2006. Profound Belief Network (DBN) was created and made progress in 2006 It permitted bidirectional associations in top layer just as opposed to stacking RBMs straightforwardly to lessen computational unpredictability, and the factors were successfully trained through layer-wise pre-preparing and adjusting. In the meantime, it was proposed to add increasingly shrouded layers and it was named deep auto encoder. It is useful to manage high nonlinear information of the model factors. Some parameters were right off the bat pre-prepared utilizing an eager layer-by-layer solo learning calculation and afterward calibrated utilizing BP calculation. After one year, Sparse Auto Encoder (SAE) was advanced to lessen dimensionality and learn inadequate portrayals. Deep learning increased expanding ubiquity. In 2010, Demising Auto Encoder was exhibited to recreate the stochastically tainted info information, and power the shrouded layer to find increasingly vigorous highlights. In 2012, profound convolution neural network demonstrated prevalent execution in picture finding. In 2014 it was introduced and featured two separate models that were operating as rivals. Generative Adversarial Network (GAN) The generative model was intended to provide irregular example, while the discriminatory model was used to prepare and group genuine and arbitrary examples. In 2016, a consideration based LSTM model was proposed by incorporating consideration component with LSTM. These days, an ever increasing number of new models are being grown even every week.

8.3.2 Comparison Between Deep Learning and Traditional Machine

Learning both profound learning and customary AI are information driven man-made consciousness methods to demonstrate the perplexing connection among information and yield not withstanding the high progressive structure, profound adapting likewise has particular traits over conventional AI as far as highlight learning, model development, and model preparing. Profound learning coordinates include learning and model development in one model by choosing various pieces or tuning the parameters by means of start to finish improvement. Its profound engineering of neural nets with many shrouded layers is basically staggered non-direct activities. It moves each layer’s portrayal (or highlights) from unique contribution to increasingly disconnected portrayal in the higher layers to locate the confounded intrinsic structures. For instance, the highlights, for example, edge, corner, form, and item parts, are disconnected layer-by-layer from a picture. These disconnected element portrayals are then contribution to the classifier layer to perform order and relapse errands. In general, profound learning is a start to finish learning structure with the base human induction, and the parameters of profound learning model are prepared together. Despite what might be expected, conventional AI performs include extraction and model development in an isolated way, and every module is built bit by bit. The carefully assembled highlights are right off the bat separated by changing crude information into an alternate area (e.g., measurable, recurrence, and time-recurrence space) to take the delegate data requiring master area information. Next, include choice is performed to improve the importance and diminish the fake repetition among highlights before nourishing into the AI model. Conventional AI procedures as a rule have shallow structures with all things considered three layers (for example info, yield, and one shrouded layer). Subsequently, the presentation of the built model not just depends on the advancement of received calculations [12, 13] (for example BP Neural Network, Support Vector Machine, and strategic relapse), yet additionally is vigorously influenced by the carefully assembled highlights. For the most part, the element extraction and choice are tedious, and profoundly rely upon space information.

In this manner, profound learning has unmistakable distinction with conventional AI procedures as showed. The significant level dynamic portrayal in include learning makes profound adapting progressively adaptable and versatile to information assortment. Since the information are disconnected, the assorted information types and sources don’t have.

8.3.3 Deep Learning for PD Diagnosis

Deep learning has become an accepted technique for successfully analyzing unstructured data like speech and audio signals, handwriting, and drawings. Deep neural networks use multiple layers of neurons and are used as multi classification and feature selection models. Classes are learned by the network with gradient mechanism by correlating number of input data and epochs for improvising learning. Activation functions also play a significant role. Generic comparison is depicted in Figure 8.1.

8.3.4 Convolution Neural Network for PD Diagnosis

Convolution Neural Network (CNN) is a multi-layer feed forward fake neural system which is initially proposed for two-dimensional picture preparing. It has likewise been examined for one-dimensional successive information investigation including normal language handling and discourse acknowledgment as of late. In CNN, the component learning is accomplished by rotating and stacking convolution layers and pooling tasks. The convolution layers convolve with crude information utilizing various neighborhood piece channels and create invariant nearby highlights. The consequent pooling layers remove the most critical highlights with a fixed-length. Generic architecture is depicted in Figure 8.2.

8.4.1 Classification of Patient and Healthy Controls

Logistic regression is the probabilistic model and may be used in parallel with subordinated variable. The calculated reoccurrence is a thorough evaluation, similar to any other retroactive exam. Connection b/w at least one ostensible/interim/ordinal/proportion level autonomous factors and the paired ward variable can be translated well with the assistance of strategic relapse.

The K-nearest algorithm (k(10)-NN) is a non-parametrical technique and can be used as a regress or a classifier. It is an example of appreciation model and is also referred to as a humble student, because he does not create a model at the time of learning. The preparation is processed only. The return value in KNN is a class. When another article is added, its nearest K neighbors are categorized into the classes they have a position, and the new item is placed in the class that has earned the greatest votes. In the recession k-NN the new item has the value the normal value of its nearest neighbors’ K estimates.

Support vector machines (SVMs) [14] come under directed learning model and examinations the information utilized for relapse/order investigations. The SVM calculations use preparatory tests that have a place each with two classes and make the new component master into a class or a class, making it an unlikely two-way direct classifier. The SVM model sees events as spatial centres, so that the models with various groups are separated via a sensitive broad aperture. It is mapped to that corresponding space when another article is to be organized. In light of the hole to which it has a place, it is anticipated to have a place with that classification.

Kernel methods (SVM) goes under class of example investigation & Support Vector Machine (SVM) is its most broadly utilized strategy. The primary aim of model evaluation approaches is to know and evaluate in datasets the general types of relationships (such as head parts, ranks, packages, orders, associations). The rough data in the dataset has to be converted unmistakably to vector depictions which are a partial map defined by the consumer. A customer-defined part can be imported from the R libraries regularly for the SVM bit.

Naive Bayes classifiers use of Bayes hypothesis. It relies on Bayes inference with strong and free assumptions between the characteristics. In view of the probabilistic relation between the class and the features it is generally called probabilistic classificatory. He is unwittingly versatile and does not look for a deterministic relationship. The arrangement is made directly by checking a shut structure statement, not least by using a pathway measurement used by different categorizers.

Decision Tree is a reflection of hubs and bends. The decisions are made in the hubs and one of the children is chosen depending on the result from the hub. It occurs until the leaf hub reaches the final category of the event. It is used in knowledge mining and AI models in general. The method starts from the root and finishes in a single leaf. Branches are used to connect highlights which ultimately induce class marks arranged on the leaves. Relapse trees have unceasing values in objective conditions.

Random forests are expansion of choice trees. They build various choice trees which are produced using arbitrary covering sets of information and traits. The consolidated element of covering sets and arbitrary choice disposes of overfitting. It uses a recurrence/arrangement learning strategy and yields the results dependent on the average trees’ expectations. By consolidating the trees in their vast number, the individual mistakes and limitations of choice trees can be overwhelmed. Figures 8.3 to 8.8 depicts the significant evaluation factor confusion matrix for the deployed machine learning models.

8.4.2 Severity Score Classification

Classification Using Artificial Neural Networks (ANNs)

Severity Class Prediction Using ANN

Neural networks are often compared to decision trees because both methods can model data that has nonlinear relationships between variables, and both can handle interactions between variables. The neural network needs to start with some weights and then iteratively update them to better values. The term kernel initializer is a fancy term for which statistical distribution or function to use for initializing the weights. In case of statistical distribution, the library will generate numbers from that statistical distribution and use as starting weights. Default learning rate 0.001 and Exponential decay rate 0.9 is with optimizers. By changing and training it rigorously, we can get new parameter setting.

Optimizers are algorithms or methods used to change the attributes of your neural network such as weights and learning rate in order to reduce the losses. Optimization algorithms or strategies are responsible for reducing the losses and to provide the most accurate results possible. Adam (Adaptive Moment Estimation) works with momentums of first and second order. The intuition behind the Adam is that we don’t want to roll so fast just because we can jump over the minimum, we want to decrease the velocity a little bit for a careful search. In addition to storing an exponentially decaying average of past squared gradients like AdaDelta, Adam also keeps an exponentially decaying average of past gradients M(t).

A neural network is a computational subdomain under domain of machine learning that can help in predictions based on existing independent data. Neural network model is fruitful for non-linear, complex data relationship. The severity class of PD non-linearly varies for number of voice features and kinematic features. For this proposed work, neural network is built in Keras to solve regression problem with as much accuracy as possible. Keras is an Google API developed to run high level neural network. Figures 8.9 and 8.10 depicted confusion matrix for ensemble models for the given dataset.

We are using the six (6) significant input variables (kinematic dataset for spiral drawings), along with seven hidden layers of 22 neurons correspondingly, and finally using the linear activation function to process the output. We are training our model over 1,000 forward and backward passes, with the expectation that our loss will decrease with each epoch, meaning that our model is predicting the value of y more accurately as we continue to train the model. Adaptive optimization methods Adam, RMSprop and NAdam are also implemented in upper layer of gradient descent to tune hyper parameters of the neural network and obtained superior r2_score that closes the generalization gap.

Major Steps of Proposed Multi Classification Model

• Weights were initialized using the hyper parameter tuning and 10 different kernel initializes

• Forty samples, in a single batch are used in an advance pass through the neural network and generated Z-values and activations for all the layers
• The loss was back propagated through the neural network layers for generating the gradients
• Each step involves using the model with the current set of internal parameters to make predictions on some samples, comparing the predictions to the real expected outcomes, calculating the error, and using the error to update the internal model parameters.
• This update procedure is different for different algorithms, but in the case of artificial neural networks, the back propagation update algorithm is used.
• The neural network starts its iteration with some random weights and then iteratively update these weight values with better modifications. The underlying concept is transforming non linear input data from input data space to output data space in a layered manner and this is done by mathematical concept called as kernel function. This neural network is designed with eight layers for this non-linear transformation using kernel. Initial step is to do kernel initialization, the process in which neural network weights are initialized with some values with definite logic. The library functions will generate numbers based on some statistical distribution and these numbers play the role of weights. Initialization of layer weights is a very important aspect of deep neural network as it is directly related to the performance of the NN model.
• There are default parameters available for neural network, the proposed method suggests hyperparameter tuning with kernel initialize and generated 10–12 models of deep neural network with weight initializer. The experiment is done to find suitable hybrid model and its influence on the nonlinear kinematic and voice data set training process for the solar irradiance prediction. Following weight initializers (kernel initializers) are combined with Relu activation function. Relu is effective and less complex as it spins half of the Z-values (the negative ones) into zeros, effectively removes about half of the variance and it simply doubles the variance of weights to compensate it.

8.5.1 Performance Measures

Sensitivity, specificity and precision of algorithms given in Eqs. (8.1), (8.2), and (8.3) are calculated using:

(8.1)

(8.2)

(8.3)

Where, TP is True_+v e and FP is False_+

TN is True+-v e and FN is False_-v e

P is Summation of all Positives

N is Summation of all Negatives

• R2_score: R-squared is a numerical measure of how close the data are to the fitted regression line. It is also known as the coefficient of determination. It describes the extent to what the variance of one variable explains the variance of the second variable. In short, it comments how much variation of a dependent variable is explained by the independent variable(s) in a regression model. Refer to Table 8.3 for tabular analysis of the three performance measures for the implemented models and relatively better performing model of ANN is also stated.

8.5.2 Graphical Results

Graphical analysis of the three performance measures for the implemented models is demonstrated in Figure 8.11 to present comparative matter at a glance. The findings were resolved with a method of uncertainty. The test sample consisted of 20% of the first set of data and the accuracy of the chaos cross section was calculated. Best results were achieved with 94.87% accurate artificial neural frameworks and least 71.79% was produced by decision tree. Each of the models is performed with R libraries and the precise tuning of the parameter has been achieved with ANN.