Optimized Cloud Infrastructure for Accelerating Early PD Prediction Through Maximized Mutual Information Feature Selection Using Voice Signal Analysis

Neurodegenerative disorders, most notably Parkinson's disease (PD), pose a significant threat to the elderly population, which includes people aged 50 and beyond. This population is particularly susceptible to the severe effects of these conditions [1]. Because PD affects neurological issues, it affects gradual loss of control over the muscles. The presence of genetic variables has been identified as a key contribution to PD [2]. The characteristic tribulation dementia that is seen in PD patients highlights the vital need to receive therapy in an early manner. Despite the fact that research is still being conducted, there is still a noticeable void in the field. This calls for a detection method that is both more accurate and more easily available, and it would be perfect if it were in the form of a mobile application [3].

In this study, a cloud-based strategy is recommended for early PD detection. The goal of using AI techniques with Parkinson's statistics is to reduce the dimensionality of the data via selecting characteristics while continuously maximising the efficiency of classification. For the purpose of determining which method of PD diagnosis is the most efficient, it is essential to use classification algorithms such as random forest (RF) and bidirectional long short-term memory (Bi-LSTM) along with Maximized Mutual Information (MMI) feature selection. Then, grid-based fuzzy one-class support vector regression is accelerated to preserve the privacy of the medical data. These algorithms are then put through extensive training and testing. Cloud computing, which has the ability to provide unrestricted access to computer resources under favourable conditions [4], is an essential component of this strategy. With the assistance of the cloud, the enormous amount of data is properly saved and managed in an economical manner.

Cloud computing offers characteristics such as scalability, stability, and storage, which make it easier to manage large amounts of data. This is in recognition of the fact that significant security measures are required when dealing with this data in order to prevent corruption and malicious activity. The most significant accomplishment of this study is the creation of an automated PD diagnostic platform that includes a smartphone application. After presenting the results regarding performance assessments, the next sections will provide a detailed review of earlier studies, expound on the suggested method for diagnosing PD and its intricate design, and conclude with some last observations. The main innovation of the study is the incorporation of cutting-edge data mining methods into a cloud-based diagnostic application for speech analysis-based early PD identification. The key contributions include:

•This research combines RF and Bi-LSTM models, offering a synergistic approach that leverages the strengths of both techniques. This fusion enhances the ability to capture subtle patterns indicative of early-stage PD in patient voice data.

•The use of MMI-based feature selection optimizes the selection of relevant features, ensuring that the diagnostic model is built upon the most informative subset. This contributes to the accuracy and reliability of the PD diagnosis in its initial stages.

•The incorporation of a grid-based fuzzy one-class support vector regression at both the physical and abstraction layers of the cloud addresses the challenge of outliers, enhancing the robustness of the diagnostic system in real-world scenarios.

•The thorough evaluation and optimization of six machine learning models pre- and post-hyperparameter tuning contribute to achieving the outstanding accuracy of 93.4%, showcasing the effectiveness of the proposed model.

The subsequent sections are structured in the following manner: Detailed information regarding the background of the work and the proposed work’s highlights in Section 2, which provides a complete assessment of research that has utilised machine learning approaches in the identification and classification of PD. For the purpose of performing PD categorization, Section 3 provides an overview of the suggested feature selection process that incorporates fault tolerance in the cloud context. Section 4 evaluates the suggested technique and compares it to existing machine learning methods for PD classification. Summary of research results and conclusions in Section 5.

In a description of the primary components of the article, which include the feature selection technique and ML algorithms for classification, as well as the specifics of the needs for cloud computing. It involves identifying the optimal subset of features (m) from a pool of (M) features, excluding redundant or irrelevant ones. This process streamlines system complexity and computational time.

The MMI attribute selection technique is used in this work. By choosing a subset (m) that successfully separates target labels determined by data that is shared, MMI seeks to find relevant characteristics. The following outlines the MMI feature selection approach.

To find out how much information is shared among two independent random variables (x and y), you use their individual potential outcomes (P(x), P(y)) and their joint proportion (P(x, y)) [33-35].

$I\left( X,Y \right)=\underset{x\in X}{\overset{n}{\mathop \sum }}\,\underset{y\in Y}{\overset{n}{\mathop \sum }}\,p\left( x,y \right)\log \left( \frac{p\left( x,y \right)}{p\left( x \right)p\left( y \right)} \right)$          (1)

Here, x stands for a particular attribute selected from the collection of features X, and y is the class identifier found in the set of templates Y.

When two traits are highly dependent, removing one does not appreciably affect their discriminative power. Finding the smallest collection of features (S) with characteristics {x_i} (n total) that show the highest dependence on the target class y while minimising duplication among them is the aim of the MMI approach. Method 1's Python-implemented MMI feature selection method seeks to accomplish this. The maximisation of dependence among the variable x_iwith label y is shown in Eq. (2). Meanwhile, Eq. (3) shows how to minimise dependence among features (x_i and x_j), acting as an amplifier for incompatible characteristics.

$MaxRv\left( {{x}_{i}},y \right)=\frac{1}{m}\underset{{{x}_{i}}s\in }{\overset{n}{\mathop \sum }}\,l\left( {{x}_{i}},y \right)$    (2)

$MinRd\left( {{x}_{i}},{{x}_{j}} \right)=\frac{1}{m}\underset{{{x}_{i~xj}}\in s}{\overset{n}{\mathop \sum }}\,l\left( {{x}_{i}},{{x}_{j}} \right)$         (3)

It uses a progressive searching procedure to find the probable optimum set of characteristics S, as shown in Eq. (5).

$max\varphi \left( Rv,Rd \right),\varphi =Rv-R$            (4)

$\max _{x j \epsilon-s m-1}\left[l\left(x_i, y\right)-\frac{1}{m-1} \sum_{x i \epsilon s m-1} l\left(x_i, x_j\right)\right]$            (5)

This research uses MMI to reduce data attributes while keeping good accuracy after learning the classification model with the chosen features. This process involves eliminating irrelevant features that could otherwise confuse the classification model's decision-making process, ultimately affecting accuracy. Furthermore, removing redundant features helps decrease the overall complexity.

Clinical Significance of Selected Features: The MMI feature selection process prioritized features exhibiting the strongest dependency on PD status while minimizing redundancy among features. Several selected attributes have established clinical relevance. Jitter-related parameters reflect cycle-to-cycle frequency instability caused by impaired vocal fold control. Shimmer measures amplitude perturbations associated with neuromuscular dysfunction. Harmonic-to-Noise Ratio (HNR) and Noise-to-Harmonic Ratio (NHR) capture breathiness and vocal roughness commonly observed in PD patients. Nonlinear measures such as RPDE, DFA, Spread1, Spread2, D2, and PPE characterize abnormalities in vocal dynamics and phonatory control. These features collectively represent known manifestations of Parkinsonian speech impairment, supporting the clinical relevance of the selected feature subset.

3.1 Utilizing machine learning models for classification

Working as an Artificial Intelligence (AI) tool, the classifier divides input data into a number of predetermined classes (k). Its main goal is to correctly identify test samples with the least amount of information and characteristics. The hybrid RF-BiLSTM architecture was selected to exploit the complementary strengths of both learning paradigms. RF effectively captures nonlinear relationships among engineered voice features and provides robustness against noise and overfitting. In contrast, Bi-LSTM models’ temporal dependencies and sequential patterns present in voice signals that may not be captured by conventional machine learning algorithms. The integration of both models enables simultaneous learning of statistical feature interactions and temporal speech dynamics. Final predictions are generated through ensemble decision fusion, where outputs from both models are combined to improve classification stability and diagnostic accuracy.

This research uses two different AI methodologies, which are as follows:

3.1.1 Bidirectional long short-term memory integration

The long short-term memory (LSTM) belongs to the category of artificial recurrent neural networks, commonly applied in deep learning. Known for its ability to retain data over extended periods, LSTM finds utility in data prediction and classification tasks. LSTM have four layers: an input layer, two concealed, and a final outlet. The first figure shows their fundamental components.

LSTM architecture incorporates specialized memory blocks known as cells, where data is stored and managed through elements termed gates. The forget gate determines what information to retain or discard within the cell. It calculates the relevance of the previous states in relation to the current state. Meanwhile, the input gate controls the addition of pertinent information to the cell state, which is eventually extracted for output or passed on to subsequent cells through the output gate.

1) Forget gate: In initial stage within LSTM involves the forget gate determining whether to retain the information from the cell. The outcome of this forget gate, denoted as b_f  in Eq. (6).

${{f}_{t}}=\sigma ({{w}_{f}}\left[ {{h}_{t-1}},{{x}_{t}} \right]+{{b}_{f}}$            (6)

In this case, $\sigma $ stands for the sigmoid function, whereas w_f and h_t-1 are the weighted matrices connected to the forget gate. Furthermore, the bias is denoted by b_f, and the input characteristics used in the calculation are indicated by x_t.

2) Input gate: The next step is to save fresh data in the cell state using the output of the input gate, as computed in Eq. (7). This procedure begins by upgrading the cell memories with alternative principles, as shown in Eq. (8).

$i_t=\sigma\left(w_i\left[h_{t-1}, x_t\right]+b_i\right)$         (7)

$C=\tanh \left(w_c\left[h_{t-1}, x_t\right]+b_c\right)$         (8)

Here, w_i and w_c denote the weights associated with the input level while ${{h}_{t-1}}$ and ${{x}_{t}}$ represent the bias to their corresponding incoming gate. Additionally, tanh represents the activation function employed in the process.

3) Cell state: At the current time step, the cell state is identified by multiplying the forget gate by the current state, then adding the resultant of the input gate elements, as shown in Eq. (9).

${{c}_{t}}={{f}_{t}}*{{c}_{t-1}}+{{i}_{t}}*c$             (9)

'i_t' here indicates the input gate's output, 'c' the cell state, 'c_t−1' the condition of the cell prior to now, and 'c_t' the current cell state.

Completing the procedure for making choices within the resultant gate, as shown in Eq. (12), is the last stage. To do this, run the initial output—which was previously determined in Eq. (10), over the sigmoid activating function, as shown in Eq. (11).

$o_t=\sigma\left(w_o\left[h_{t-1}, x_t\right]+b_o\right)$           (10)

${{h}_{t}}={{o}_{t}}\tanh ({{c}_{t}})$            (11)

$outpu{{t}_{class}}=\sigma \left( {{h}_{t}}*{{W}_{out~parameter}} \right)$           (12)

Here, ${{W}_{o}}$ and ${{W}_{out~parameter }}$represent the weights associated with the output gate, while b_o signifies the bias term. Additionally, $Outpu{{t}_{class}}$ denotes the classification outcome and stands for the sigmoid activated function used in this process.

3.1.2 Random forests integration

Among the data extraction methods used mostly for supervised training is RF. It constructs a 'forest' by generating multiple decision trees through a technique called 'bagging,' where each tree is trained independently to enhance classification accuracy, as shown in Figure 1.

By building numerous categorical decision chains throughout learning, RF is used in regression, classification, and other applications. The essential characteristics of RF are:

•Handling multiclass classification and regression naturally.

•Swift training and prediction.

•Dependency on one or two factors for tuning.

•Built-in estimation of generalized fault.

•Applicability to high-dimensional problems.

•Feasibility for parallel implementation.

•Capability to measure variable importance.

•Support for differential class weighting.

•Unsupervised learning with potential for visualization and outcome received.

tu_pian_1.png

Figure 1. Framework for bidirectional long short-term memory (Bi-LSTM) network

The operational steps of RF can be summarized as below:

Examine a collection of real-value inputs for their variables X = X_{1 …..} X_p that have 'p' values.

$E_{x y}=(L,(Y, f(x)))$            (13)

Here, $\left( Y,f\left( X \right) \right)$ denotes the loss function utilized for computing the predictive function $f\left( X \right)$, where 'Y' stands for the actual outcome. The common representation of the loss function is through the root mean square error. However, in RF classifiers, the entropy information gain often serves as the default loss function.

$L(Y, F(X))=I(Y \neq F(X))=\left\{\begin{array}{c}0 \text { if } Y=F(X) \\ 1 \text { otherwise }\end{array}\right.$          (14)

The potential range of 'Y' are encompassed within the real number space (ℝ). EXY (L(Y, f(X))), particularly for 0 – 1 loss, yields the Bayes rule as depicted in Eq. (15).

$f\left( x \right)=argma{{x}_{y\in R}}P(Y=y|X=x)$          (15)

Eq. (16) shows how the collective prediction 'f,' which includes h_j(x) base learners, votes to decide the projected class.

$f\left( x \right)=argma{{x}_{y\epsilon R}}\underset{j=1}{\overset{J}{\mathop \sum }}\,I\left( y={{h}_{j~}}\left( x \right) \right)$            (16)

Here, 'Y' stands as the RF denoting the output, while $X$ represents the predictive function used for estimating Y. $EXY$ signifies the loss of outcome, with ℝ denoting the possible values of Y. Additionally, $h_j~\left( x \right)$ corresponds to the jth base within the ensemble.

3.2 GAE: Google application engine

In 2008, the introduction marked Google's entry into the demand-driven computing market by giving users access to a framework that allows them to create and host a variety of web-based apps within its data facilities. This service became publicly available in November 2011 alongside other cloud services provided by Google. GAE essentially functions as a comprehensive file system, providing users with more extensive capabilities than a standard engine, including read/write permissions. GAE standard users have a fixed set of libraries in serverless technology settings, restricting third-party library distribution. Representational State Transfer (REST) serves as a web service design architecture and stands as a popular alternative to SOAP. REST is favoured for its lower bandwidth consumption, making it a preferred choice for internet usage. SOAP operates on a client/server logic protocol using the RPCs model, resulting in services becoming XML documents that define the methods for calling. Because XML is used in SOAP exchanges often, greater storage capacity is used. REST sends information in XML or JSONs forms to the website's server via requests made over HTTP. Cloud-based APIs from Google, Amazon, Twitter, and Ms choose REST owing to its flexibility and inexpensive construction. A web service that uses the REST framework is known as a RESTful API in Figure 2.

tu_pian_2.png

Figure 2. Framework for random forest (RF)

3.3 Parkinson's disease diagnosis process in cloud environment

Utilising a cloud-hosted environment is how the PD screening system operates. A visual depiction of the elements and setup of the cloud-based PD detection platform is shown in Figure 3. The cloud-powered ML-based PD diagnostic technique, the RESTful API, and the Google Play mobile app have been recognised as the method's three key elements.

tu_pian_3.png

tu_pian_4.png

Figure 3. Overall flow of the proposed taxonomy

3.3.1 Cloud deployment phase

The machine learning methodology that has been proposed for the development of the PD detection system combines many distinct elements. These consist of the PD evaluation database, the initial processing stage, which is in charge of getting the information set ready for training the machine developing categorization approach, the feature identification stage, which lowers dimensions by selecting the best characteristics that show exceptional performance in task categorising, and the algorithms used for machine learning phase. In the last stage, a number of different ML algorithms get trained by making use of the data that was collected in the stage before it. Subsequently, the model's capacity for classification is calculated to determine the most efficient approach for creating a model for categorization appropriate for PD identification.

The data used in this study regarding PD were gathered by Max Minor from the University of Cambridge and the National Centre of Voice [14]. This dataset contains information derived from voice recordings of 32 individuals. Among them, 24 have a diagnosis of PD, while the other 8 are considered to be in good health. There are an overall of 195 recordings of voices included in the dataset, and each recording produces 22 retrieved features, as summarised in Table 2. The extraction of these traits is accomplished either through prolonged vowel phonation or through a running voice test [20]. Individuals are asked to articulate the vowel "a" for a period of time that is shorter than ten seconds while using lab microphones and recording tools. This technique is known as sustained vowel phonation. Each recording is linked to a binary PD score for the purpose of prediction. The benchmark data that was accessed from the ML data collection in reference [21] was the database that was utilised for this researcher's investigation. For practical deployment, the proposed system was implemented using Google Application Engine, which provides automatic resource scaling and high availability. The average audio recording duration is limited to 10 seconds, resulting in relatively low bandwidth requirements during transmission. Since only extracted acoustic features are transmitted to the cloud rather than raw continuous audio streams, communication overhead is minimized. Preliminary testing indicated response times suitable for near real-time diagnosis. Future studies will include detailed latency analysis, cloud resource utilization measurements, and cost evaluation under large-scale clinical deployment scenarios.

Table 2. 22 feature sets

Note: MDVP = Multi-Dimensional Voice Program; NHR = noise-to-harmonics ratio; HNR = harmonics-to-noise ratio.

Within the scope of this investigation, each of the 22 characteristics is addressed. It is possible to aggregate the values that correspond to these characteristics for an identical person by employing functions such as the minimum, the range, the standard deviation, the mean, and the maximum. The combining process is carried out in an iterative manner for each participant, which ultimately results in an information set size of 32 by 110 measurements. It has been decided that the set number of features for the MMI algorithm will be (5%, 10%, 25%, 50%, 75%, and 100%), correspondingly.

It is important to divide the classification data into two or more sets in order to get it ready for use. When the classifier is being trained, one set is utilised for training, and this set will be referred to as the set for training. Another set, which is additionally referred to as the test set, is used to evaluate whether or not the classifier has successfully learnt the rules.

3.3.2 Utilization of the application programming: RESTful

The smartphone or tablet and the PD assessment that is located in the cloud are connected through the RESTful Interface for Application Programming (API), which acts as a connection between the two. The Flask Python toolkit is utilised in order to develop and put in the RESTful API that is hosted in the cloud. POST is a mechanism that is utilised within the mobile application to submit patient information to the Bi-LSTM PD detection model that is located in the cloud, receive the forecast outcome & then send it again to the app. In Algorithm 2, the process of calling the RESTful API is described in some detail. The API is put through testing by sending an instance of patient information to CCs. This is done to confirm that the RESTful API that has been created and installed in Google Cloud is running correctly and that it is able to facilitate the required interaction with the mobile application.

3.3.3 Smartphone app development phase

This Android mobile software was created to diagnose Parkinson's illness, and it has a simple GUI and is built in Java. The individual can capture audio recordings using the attached recording button in order to aid in the diagnosis of Parkinson's illness. The final test result is then displayed on the screen by the application when you have completed the process. The diagnosis model utilizing machine learning, deployed on Google's cloud infrastructure, is responsible for generating the decision. For more specifics about this cloud-based model for diagnosing PD, refer to Algorithm 2. Additionally, Algorithm 3 outlines the pseudocode for the mobile application related to this diagnosis system.

The mobile app prompts patients to record their sound by articulating the note /a/ three times within a ten-second interval. Praat, derived from the Dutch word for "talk," refers to a sound analysis program. It's employed to recognize sound characteristics akin to those utilized in developing the LSTM diagnosis model.

After the recording of the audio sample, the smartphone app communicates with the Praat software service. After the features have been retrieved, they are translated into the JSON format in order to guarantee compliance with the RESTful API. The cloud-based PD assessment model receives this information after it has been submitted to it.

3.3.4 Outliers control using grid-based fuzzy one-class SVR

A cloud-based web application with flaws has been developed by creating a simulated environment with many nodes using a hypervisor to construct virtual machines. Finding the broken node is essential to getting reliable defective data. The cloud data system archives the tasks allocated to these virtual nodes, while the cloud master maintains control of the virtual servers that are produced. Finding malfunctioning nodes is made easier by analysing data from both sources. In order to efficiently obtain false data, procedures such as pulse tactics, contact plans, keep-alive approaches, and intrusion methods are used. Training data is obtained by using a grid-based fuzzy one-class support vector regression for outlier detection.

Cloud systems can have a variety of issues that originate from software, hardware, and other barriers. Network, process, and data failures are additional categories for software errors. The data must be categorised and trained appropriately in order to create an efficient detection and diagnosis model. Training data is classified as a consequence of the fault classification process using the Bi-LSTM technique. A resource-efficient statistical methodology called logistic regression is used to identify faults. This approach is appropriate for the prediction task because it doesn't require a lot of processing power (Algorithm 4).

${{Y}_{i}}={{\alpha }_{0}}+~{{\alpha }_{1}}{{x}_{i}}+~{{\pi }_{i}}$          (17)

where,

i represents the dependent variable.

${{\alpha }_{0}}$ denotes the population y-intercept.

${{\alpha }_{1}}$ signifies the population slope coefficient.

${{x}_{i}}~$stands for the independent variable.

${{\pi }_{i}}$ represents the random error term.

The assessment procedure starts with entering the data to be evaluated, reviewing the file, and then providing an explanation of the test data using the SVR (Eq. (17)). By utilising an encoding tool, the test data is converted. The system next performs a critical phase in which it determines if the given data is considered anomalous or typical. The next step is feature extraction, which selects pertinent features. Next, the precision and accuracy of the provided data are ascertained by utilising the logistic regression method. Notably, SVR is applied only to the significant features that are extracted with the help of Bi-LSTM. Calculated results indicate that the accuracy obtained by using logistic regression is 0.981.