• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Dm-Health App: Diabetes Diagnosis Using Machine Learning with Smartphone

    2022-08-24 12:56:48EliasHossainMohammedAlshehriSultanAlmakdiHananHalawaniMdMizanurRahmanWahidurRahmanSabilaAlJannatNadimKaysarandShishirMia
    Computers Materials&Continua 2022年7期

    Elias Hossain, Mohammed Alshehri, Sultan Almakdi,*, Hanan Halawani, Md.Mizanur Rahman,Wahidur Rahman, Sabila Al Jannat, Nadim Kaysarand Shishir Mia

    1Department of Software Engineering, Daffodil International University, Dhaka, 1207, Bangladesh

    2Department of Computer Science, Najran University, Najran, 55461, Saudi Arabia

    3Department of Computer Science and Engineering, Rajshahi University of Engineering and Technology, Rajshahi, 6204,Bangladesh

    4Department of Computer Science and Engineering, Mawlana Bhashani Science and Technology University, Santosh,Tangail, 1902, Bangladesh

    5Department of Computer Science and Engineering, BRAC University, Dhaka, 1212, Bangladesh

    6Department of Computer Science and Engineering, World University Bangladesh, Dhaka, 1230, Bangladesh

    Abstract: Diabetes Mellitus is one of the most severe diseases, and many studies have been conducted to anticipate diabetes.This research aimed to develop an intelligent mobile application based on machine learning to determine the diabetic, pre-diabetic, or non-diabetic without the assistance of any physician or medical tests.This study’s methodology was classified into two the Diabetes Prediction Approach and the Proposed System Architecture Design.The Diabetes Prediction Approach uses a novel approach, Light Gradient Boosting Machine (LightGBM), to ensure a faster diagnosis.The Proposed System Architecture Design has been combined into sevenmodules;;the Answering Question Module is a natural language processing Chabot that can answer all kinds of questions related to diabetes.The Doctor Consultation Module ensures free treatment related to diabetes.In this research,90% accuracy was obtained by performing K-fold cross-validation on top of the K nearest neighbor’s algorithm (KNN) & LightGBM.To evaluate the model’s performance, Receiver Operating Characteristics (ROC) Curve and Area under the ROC Curve (AUC) were applied with a value of 0.948 and 0.936, respectively.This manuscript presents some exploratory data analysis,including a correlation matrix and survey report.Moreover, the proposed solution can be adjustable in the daily activities of a diabetic patient.

    Keywords: Machine learning; diabetes-prediction; support vector machine(SVM); LightGBM; eHealth; ROC-AUC

    1 Introduction

    Information has become the basis for more substantial business and innovations.The more data we have, the morewe can organize ourselves optimally to produce the best results in diagnosing diseases in the healthcare sector.Various big data outlets in the healthcare sector include hospital reports,patientmedical records, medical test outcomes, and applications of the Internet of Things (IoT).Large databases can be found in the healthcare industry; these databases might contain structured, semistructured,or unstructured data [1].Big data analytics is the method that analyzes enormous sets of data and exposes hidden data, hidden patterns to discover meaningful insight from the data.In recent years, DiabeticMellitus (DM) has been a highly serious disorder.This is a Non-Communicable Disease (NCB), and numerous individuals are experiencing it.According to statistics in [2], from 2017, about 425 million individuals have diabetes.Diabetes causes between 2-5 million people to lose their lives each year.This is said to grow to 629 million by 2045 [2].The shocking fact is almost one in ten adults in developing countries conveys diabetes.

    Around 83 perent of people in developing countries are uninformed about their blood glucose levels, and they never search for diabetes most of the time.According to World Health Organization(WHO), deaths reached almost 5.09 percent of the developing countries, including Bangladesh, in diabetes, published data in 2017.Diabetes Mellitus (DM) refers to Type-1, Type-2, and Type-3 gestational diabetes.This type of diabetes is caused by the body’s inability to create enough insulin,necessitating insulin injections.Non-Insulin-Dependent Diabetes Mellitus (NIDDM), or Type-2 Diabetes Mellitus, is a type of diabetes that is not dependent on insulin.This type of diabetes arises when body cells are unable to utilize insulin effectively.Aspike characterizes type 3 gestational diabetes in blood sugar in pregnant women who do not have diabetes.This type of diabetes develops more quickly.Long-termproblems are linked to diabetesmellitus.Diabetes are at a higher risk of developing a variety of health problems.

    Recently, the computational approach [3] is being widely used in the healthcare sector.Predictive Analysis is a collection ofmachine learning algorithms, data mining techniques, and statisticalmethodologies for uncovering insights and forecasting future occurrences based on current and historical data.By applying predictive analytics to healthcare data, important decisions and predictions can be made.Predictive Analysis aims to accurately diagnose diseases, improve patient care, optimize resources,and improve clinical outcomes [2].Machine learning is one of the most important aspects of artificial intelligence because it allows computers to learn from past experiences without being programmed[4,5].Machine learning’s realistic implementations drive business outcomes that can significantly impact the bottom line of a corporation.New techniques are increasingly emerging in the field and have extended Machine learning to almost limitless possibilities [6,7].As of now, to detect diabetes,fasting blood glucose and the laboratory tests oral glucose tolerance.However, this procedure takes a long time.

    According to [8], the android application has become increasingly popular in recent years.62.3%of people across the world use mobile data andWi-Fi.Besides, it is noted that when the designmethods used in computers and humans are incorporated during the life cycle into the software development process, the projects developed are more successful, better quality, and more user-friendly.

    In the last few years, there have been many systems that have been proposed to monitor diabetic patients.So, monitoring systems offer several advantages for diabetic patients, such as the quality of life of diabetic patients has improved, and the number of hospitalized diabetic patients has reduced [9].Thus, the main goal of this app is for health-related issues, especially diabetes diseases using modern computation such as machine learning and android application.We have provided multiple packages in this research, and there are three contributions found in this proposed research:

    ?Firstly, this research has shown a novel technique for predicting diabetes through which it is possible to get very fast and promising accuracy, especially for diabetes prediction.While the LightGBM algorithm employs two innovative approaches named Gradient-Based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB), which allow the algorithm to run faster while retaining excellent accuracy, we have classified it as a Novel technique.

    ?Secondly, the machine learning model has been integrated into an android application and shown the pipeline of model deployment separately, which will serve as a benchmark in the research community.This research shows a functional overview of how to connect the Machine Learning Model with Android Application details.

    ?Finally, before developing the mobile application for diabetes prediction, a survey was conducted online.Various suggestions have been taken from the users, especially for people with diabetes.

    The paper is divided into five sections that are all interconnected.The results of the literature search and reviews and related debates are presented in section two.Section three shows the overall research methodology and proposed system analysis.Section four presents the result of the discussions.Finally, section five depicts the manuscript’s conclusion.

    2 Literature Reviews, Search and Discussion

    This section is divided into three interconnected components: Literature Search Strategy, Literature Selection Criteria, and Literature Review.In this research, papers have been collected to review literature from all the databases, and the overall search strategy technique has been shown in the Literature Search Strategy section.The literature Selection Criteria section shows how to select related papers based on the selection criteria after collecting the article from the database.Finally, in the Literature Review part, the Literature Review is written in-depth based on the selected articles.

    2.1 Literature Search Strategy

    We have explored several electronic databases from 2010 to November 2020.IEEE Xplore,PubMed, National Center for Biotechnology Information (NCBI), Springer, Google Scholar, and Elsevier have been applied to recognize similar published research papers.

    Following the steps in‘An organized method to documenting a search strategy for publishing’, an instigated literature search approach was established [10].Moreover, after the preliminary identification,each paper is scanned by following the Exclusion and Excl

    2.2 Literature Selection Criteria

    Based on the following principal criteria, we have initialized our selection criteria for receiving a research article and have retrieved only those papers that meet the following requirements.Some criteria have been set for literature selection because everyone downloads papers randomly while reviewing literature and working there.In this way, the literature review section can be made much more substantial, making the quality of the paper much effective.

    Figure 1: Block diagram of literature search & selection strategy of the proposed research

    Inclusion Criteria

    1.The research paper must be required to be a conference paper or journal article.

    2.Machine learning-based model or program intended solely for diabetes purposes must be included in the research paper.

    3.Each research study’s goal must concentrate on diagnosing or screening patients with diabetes through Machine learning methods or a solution based on mobile applications.

    4.The time frame for each scientific manuscript is 2010 to 30 November 2020.

    Exclusion Criteria

    1.Any research work that carried out as a preprint, early works, or not peer-reviewed.

    2.Any comparative studies.

    3.Editorials, Review Papers or Research Letters.

    4.In the domain of diabetes prediction, no method based onmachine learning has been discussed.

    2.3 Literature Review

    The paper “Usability Pitfalls of Diabetes mHealthApps for the Elderly” [11] shows that increasing the usability of an app in a number of categories changes a small percentage of the total number.The paper [12] entitled is“Mobile Applications for Diabetes Self-Management: Status and Potential”provides an overview of smartphone technology and health apps.For the prediction of diabetes, the article [13] offered ANN, Random Forest (RF), and K-means approaches.The ANN technique gave the best accuracy of 75.7 percent and can help medical professionals with care decisions.

    [14] Shows a model for predicting diabetes using classification algorithms.The algorithms Naive Bayes, Support Vector Machine (SVM), and Decision Tree (DT) have been proposed to complete the research.Performance evaluations on various measures as well as comparative analysis based on the accuracy of different models have been illustrated and finally achieved 76.30% accuracy through the Na?ve Bayes algorithm.Another study [15] has been Ensemble approaches used to conduct towards Diabetes Prediction.The Bagging approach accurately categorized 95.59 percent of diabetes forms,while Decorate accurately categorized 98.53 percent.

    The authors of this study [16] concentrate on developing a predictivemodel for diabetes prediction by applying Machine Learning (ML) algorithms and techniques of data quarrying.They have demonstrated the pipeline of diabetic prediction through the use of Boosting approach with various traditionalML algorithms and have increased the accuracy of classification algorithms.The paper [17]focused on developing risk assessment of diabetes among some individuals about their daily life and family history.A machine learning mechanism is adopted to train and validate the proposed model.In the paper [18], the authors utilized the Logistic Regression model to assess the risk of diabetes upon p-value and odds ratio.Four classifiers had been taken to predict the risk of diabetes.

    The paper [19] aimed to implement an intelligent mHealth application based on anMLmechanism to predict diabetes and classify it into three categories: diabetes, prediabetes, and non-diabetes.The paper [20] had reflected on the execution of a scheme to analyze diabetes infection utilizing different types of ML algorithms.

    These reviewed articles had only focused on the established model to predict diabetes disease,but no mechanisms to build a mobile application for diabetes patients had been proposed.This proposed study’s key idea and motivation are to design a system that ensures the computer system and general people interaction by deploying the machine learningmodel into the Android Application and quickly predicted diabetes.Additionally, a table has been interpreted in this section to demonstrate the literature review summary that is illustrated in Tab.1.

    Table 1: Summary of the Literature Review

    Table 1: Continued

    Table 1: Continued

    3 Methodology of the Proposed System

    This section’smethodology is divided into the Diabetes Prediction Approach and System ArchitectureDesign.The Diabetes Prediction Approach part shows how the algorithm accomplishes diabetes prediction, and the overall system architecture design is described in the System Architecture Design section.Fig.2 provides the number of algorithms used in this manuscript and the corresponding model architecture.In this proposed study, a mobile application has been developed for diabetes users to identify the possibility of diabetes through this application.This mobile application has a Health Module component, where the machine learning mechanism is integrated and analyzed to predict the probability of diabetes based on some interconnected parameters.

    3.1 Diabetes Prediction Approach

    The Diabetes Prediction Approach has been classified into four segments, to illustrate, Experimental Setup, Data Preprocessing, Algorithm Selection, and Model Deployment.The dataset has been used in this proposed research, which is explained in the Research Dataset Section.The data preprocessing pipeline is shown in detail in the Data Preprocessing section.All the machine learning algorithms used in this research are described in the Algorithm Selection.Finally, a model is deployed and communicates with a mobile application, explained in the model deployment part.There are many reasons to use various machine learning algorithms in this research, and there are specific algorithms that are frequently used for this particular work beforehand.Those who conduct research will quickly understand all these soft computing’s performance in this type of work; then, there will be no problem during the model selection.Using this soft computing, we received an idea about algorithms’computational cost while creating the model.Model selection is an essential issue during Disease Detection in the research community.It will be straightforward to select the model by looking at the interpretations of different types of algorithms.

    Figure 2: Algorithms and the corresponding model deployment

    3.1.1 Research Dataset

    The National Institute of Diabetes and Digestive and Kidney Diseases provided this data.The goal of the dataset is to use particular diagnostic metrics in the dataset to diagnose whether a patient has diabetes.Several other online portals fromwhich a wide variety of datasets are commonly available such as Kaggle and UCL Machine Learning Repository are very popular.The goal of the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK)’s is to enhance people’s wellbeing and excellence of life by undertaking and promoting medical studies and research training, as well as disseminating science-based knowledge on diabetes and other endocrine and metabolic diseases; digestive diseases, dietary disorders, obesity, kidney, urologic, and hematologic diseases.Since the information from NIDDK is very authentic, we have taken the dataset from this particular place because since this research has been carried out on essential issues like diabetes, accurate and authentic information is essential.The more authentic the information, the better the research quality and all the researchers worldwide will be able to believe our research’s messages.

    The selection of these examples from a larger database was subjected to a number of restrictions.All of the patients at this clinic are Pima Indian women who are at least 21 years old.Pima Indians in Arizona had the highest prevalence of type 2 diabetes (NIDDM) in the world in 1993.Scientists from the National Institutes of Health (NIH) studied this phenomenon for over 30 years.The majority of findings of the NIH suggest that both acquired (environmental) and genetic factors play a critical role in the development of type 2 diabetes (NIDDM) in this population.Disease complications in this group are disruptive or fatal, and further research into the causes and prevention of disease will be beneficial.There are various medical predictor factors in the dataset, as well as a target variable,Outcome.The predictor variables combine the patient’s number of pregnancies, BMI, insulin levels,age, etc.We have collected the dataset from the Kaggle [21].Tab.2 shows the attribute of the proposed research dataset.

    Table 2: Attribute of the proposed research dataset

    3.1.2 Data Preprocessing

    Data Preprocessing is divided into three sections, to illustrate, Data Cleaning, Data Transformation, and Data Reduction.Data preprocessing is essential in any data mining process because it directly affects the project’s success rate.If there are attributes, attribute values, noise or outliers, and redundant or missing data, it is called data impurity.The presence of any of these will degrade the quality of the results.There can be several irrelevant and missing components of the results.Data cleaning is done to handle this portion.It includes taking missing data, noisy information, etc.The Data Transformation phase is kept in place to transform the data into formats that are acceptable for mining.This process combines Normalization, Attribute Selection, Discretization, and Concept Hierarchy Generation.When dealing with a large volume of data, analysis becomes more complicated when the data dimension is high.To get rid of this, the method of data reduction is used.It aims to increase the efficiency of storage and decrease the cost of data storage and analysis.Fig.3 shows the missing value in our dataset when retrieved online and the disappeared value stage after the clean dataset was created at the end of preprocessing.So, data processing criteria have been accomplished through the discussed approaches.

    3.1.3 Algorithm Selection

    Several classification algorithms of machine learning have been applied in the Diabetes Prediction Approach part.This section has highlighted the SVM, KNN, and Light GBM algorithm because they workwell for this dataset.Thus,we have illustrated the mathematical interpretation of these algorithms in the following subsections.

    (a) Support-vector Machines (SVM):In machine learning, SVMs are supervised learning models with related learning algorithms that examine classification and regression analysis data.A support vector machine generates a group of hyperplanes in an infinite or highdimensional space to build the classifier.The hyperplanes also can be used for regression and other tasks, such as spotting outliers.Support vectors are the vectors or cases that contract the hyperplanes.Fig.4 shows the linear SVM and its mathematical explanations.From this figure,we can illustrate that any hyperplane can be written as:

    Figure 3: (a) The combination of missing values before data preprocessing (b) The combination of clean data after preprocessing

    Figure 4: Hyperplanes with a maximum margin and SVM margins for samples of two classes trained

    Here, w is the (not necessarily normalized) average vector to the hyperplane.The“margin”is the area or region bordered by these two hyperplanes, and the maximum margin hyperplane is the hyperplane that lies halfway between them.These hyperplanes can be defined by equations using a normalized or standardized dataset.

    So, we can write the width or the margin of the two hyperplanes for data classification can be written as the following equations:

    Radial Basis Function (RBF) Kernel Support Vector Machine (SVM):

    The SVM has been shown to be effective on both linear and nonlinear data.The radial base function was introduced with this approach to categorizing nonlinear data.The kernel function is crucial for putting information into the function space [22].For example, if we plot more than one variable in a typical scatter plot, in many cases, that plot cannot separate two or more data classes.The kernel of SVM is a unique sort of approach for converting lower-dimensional input into higherdimensional space and distinguishing between classes.Linear kernels, polynomial kernels, and radial basis function kernels are some of the types of SVM kernels accessible.The radial basis function is also a kind of non-linear function.This function is the most popular function of the support vector machine.This kernel can map any input to infinite-dimensional space.

    The Radial basis function (RBF) kernel is another name for a Gaussian function.In Fig.4, the input space divided by feature map (Φ).By applying Eq.(1), we get:

    By applying Eq.(3) in (4), we get a new function, where N represents the trained data.

    (b) Light Gradient Boosting Machine (LightGBM) Algorithm

    We employ the concept of [23] verdict trees to transfer a function, for example, from the input space X to the gradient space G.It is believed that a training set with instances like x1, x2, and up to xn is used, with each element being a vector of s dimensions in the space X.All negative gradients of a loss function about the output model are denoted as g1, g2, and up to gn in each repetition of a gradient boosting.The decision tree separates each node at the most revealing feature, giving rise to the greatest evidence gain.

    Y= Base_tree (X)--Ir*Tree1 (X)--rr*Tree 2(X) - lr*Tree (X)

    Explanation, Let O represent a training dataset on a fixed node of a decision tree, and the variance gain of splitting measure j at a point d for a node is defined as:

    Gradient One-Sided Sample, or GOSS, uses every case with a more significant gradient and performs random sampling on the many instances with small gradients.For each node of the Decision tree, the training dataset is represented by the notation O.The variance gain of j, or the dividing measure at node position d, is given by:

    Where,

    Al= {xi∈A:xij≤d},Ar= {xi∈A:xij>d}

    Bl= {xi∈B:xij≤d},Br= {xi∈B:xij>d}, and the coefficientis used to normalise the sum of the gradients over B back to the size ofAC

    (c) K-nearest Neighbors Algorithm (KNN):The K Nearest Neighbor (KNN) said that the Knearest neighbor of the unseen data point would find the K-nearest neighbor for a given K algorithm value and then allocate the class to the unspecified data point by making the class with the maximum number of individual points out of all K neighbor classes [24].

    Finally, with the uppermost possibility, the input x is allocated to the class.

    (d)Hyperparameters Tuning:Random search (RS) is a strategy for finding the optimal solution for a created model by using random combinations of hyperparameters.GridSearchCV, which calculates all potential combinations, is generally faster and more accurate than RS.We indicate the number of combinations we want with Random Grid.The following parameters have been tuned for the case of LightGBM:

    Learning rate: The learning rate specifies how much each tree affects the final result.GBM works by starting with an initial estimate and updating it with each tree’s output.The learning parameter controls the magnitude of this change in estimates.

    N estimators: N estimators are the number of trees that have been estimated (or rounds), num leaves: the total number of leaves in the tree; the default is 31.Minimum child samples: the smallest number of data samples in a single leaf.Minimum child weight: smallest sum hessian in one leaf that has been used to deal with over-fitting.Subsample: choose a portion of data at random without resampling, maximum depth: It specifies the tree’s maximum depth.This parameter is used to deal with overfitting in models.Colsample_bytree: if colsample bytree is less than 1.0, LightGBM will randomly select parts of features on each iteration.LightGBM, for example, will select 80 percent of features before training each tree if we set it to 0.8.Regularization (reg alpha) and regularization (reg lambda) were tuned properly.

    Early stopping rounds: This parameter made our analysis go faster.If one metric of one validation data does not improve in the last early stopping round rounds, the model will cease training.Excessive iterations will be reduced as a result of this.On the other hand, for the case of knn, the best parameters obtained: {‘knn__n_neighbors’: 18}.

    3.2 Model Deployment

    Machine Learning is a branch of AI that allows machines to learn, discover, and predict outcomes without the need for human interaction [25].In numerous fields, machine learning has been applied and currently actively serves to make mobile apps.Tensorflow lite, which is continually changing,has produced new and exciting features for their mobile apps, simple for mobile developers.Robust machine learning mobile applications can exploit great business models and perform complex tasks such as face recognition or automated image captioning, all in actual time and without Internet access[26].This research has come across a REST API called Retrofit.Java API, i.e., Retrofit, will send and receive a response from Python API.In this research, Retrofit is adopted as a Java API because it is easy enough to use.The most significant benefit is that it allows API calls as quickly as the Java method calls for everyone.As a result, developers have more flexibility in defining URLs to hit and determining request/response type parameters as Java classes.Retrofit made networking in Android applications a lot simpler.It includes various capabilities that help us to reduce boilerplate code in our application and easily consume web services, such as adding custom headers and request types,file uploads, simulating replies.However, details on how to integrate with Android apps by deploying machine learning models are shown in the Model Deployment section and shown in Fig.5, along with diagrams.

    Figure 5: Architectural diagram for deploying machine learning model into mobile apps

    Model Creation for Machine Learning includes a pipeline that begins with data collection,exploratory data analysis (EDA) and goes to the real world for model implementation.The model has been deployed in Heroku using the services of Flask Restplus.We used Heroku Server because it provides a free plan to learn and get started [27].Heroku is a platformthat assists developers in honing their skills to build feature-rich applications.Developers will benefit from the experience because they will have access to valuable resources for speeding up key development processes.Heroku’s free version is suitable for smaller software projects.Developers may also select from a range of tier packages that are ideally tailored to large companies’diverse needs.The user-friendly Heroku platform dashboard allows scaling, management, and application monitoring.Heroku has many more platforms as an alternative, but since it can be used primarily for free using various modules, we have selected this server for this research.The following approaches need to be taken after the Flask API’s deployment to allow our Flask Application Programming Interface (API) to communicate with the Java Client.Flask is a Python web framework that includes features for creating web applications such as handling Hypertext Transfer Protocol (HTTP) requests and rendering templates.Flask has several alternatives,the most common of Django, Tornado, Express JS, Node.js, and React.As far as we are concerned,the flask’s most significant benefit is its design, which is both lightweight and modular.It also has excellent community support and good documentation for developers to get started.Flask appears 100% compliant with Web Server Gateway Interface (WSGI) makes it easy to deploy for production.The details sequence and consequences are shown below:

    ?Create an object for Retrofit Class-This phase can be described as a‘getApiClient’generic method that returns a retrofit object.Besides, we have to mention BASE Uniform Resource Locator (URL) Endpoints in this process, pointing to our Flask URL.

    ?Create Interface with HTTP Operation-This interface will serve as a bridge to communicate with RESTendpoints.Aresponsefromthe Flask Endpointtothe Plain Old Java Object(POJO)class will be returned.

    ?Defining POJO Class to handle the response-This Class captures the return variables from Flask Endpoint.We have to define a key-value pair in a flask; those keysmust be declared in the POJO class.

    ?Encapsulating the Steps in Main Activity.java-We have to construct an object for the interface at MainAcitivity.java; i.e., however, by applying this object, ApiInterface must call the procedure declared in the interface.Furthermore, it is essential to establish a callback interface.It has two strategies that are called on Response and On Failure.Moreover, if the Java API request is flourishing, the on Response method will be contacted.Otherwise, the on Failure process will be called.Finally, the Response method will allocate the POJO class to the Response object.We will be able to extract the response from POJO using the getter method.

    Fig.5 illustrates the Architectural diagram for deploying a machine learning model into the mobile application.The figure has been divided into two interconnected parts, for instance, Backend and Frontend.When a request comes from the Frontend, it reaches the Backend through HTTP API Service.After coming to the Backend, based on the input, it hits the model, and the model starts to predict and sends it to the User Interface(UI) component via HTTP (REST) and then the output is seen from the UI Component.Through this procedure, the machine learning model completes the prediction through integration with the mobile application shown in Fig.5.

    3.3 Proposed System Architecture Design

    The Proposed System Architecture Design has been categorized into six modules, for instance,Food Recommendation Module, Health Module, Answering Question Module, Daily Movement Activities Module, Daily Awareness Module, and Medicine Reminder Module.This system has been developed by deciding all the possible problems and its solution regarding diabetes.This system would be a compact solution for diabetic patients.The Patient and Doctor will have the authorization to access this developed system.The modules of this system are connected through the Channel, which is the way of data communication.In the Data element portion, we have recorded the data coming through the proposed system.A detailed explanation of the proposed system’s module and diagram are shown in Fig.6.

    Figure 6: Proposed system architecture design

    While developing this mobile application, we have considered everything in terms of health monitoring and diabetes prediction.To put it more simply, for the case of diabetes patients, it is important to avoid having a variety of foods.Also, it is not very difficult for people in first-world countries to get treatment from a doctor, Still, in third-world countries, it is a slightly difficult matter because low-income people find it problematic to see a doctor due to financial constraints.Therefore,to reduce hassle or financial cost,we have put all the following modules in themobile applicationwhere all the problems related to diabetes can be solved.Nonetheless, we have provided multiple features for the case of diabetes and health monitoring.

    3.3.1 Food Recommendation Module

    A healthy body and mind truly depend on a healthy eating plan.This module aims to recommend to users about preferable food.We have collected some food lists for diabetes users from the National Institute of Diabetes and Digestive and Kidney Diseases [28].After collecting the food list of diabetic patients,we have created a database, and it has been integrated into themobile application.When users enter this module, they will be shown a list of foods through which they can be aware of what kind of food they should eat if they have diabetes.For specific individuals, awareness of diabetes can be about avoiding the onset of the disease.Healthy eating and more active lifestyles will avoid type 2 diabetes caused by obesity.Since this system has been developed in a user-friendly way with everyone in mind,through the Food Recommendation Module, anyone can get an idea of what kind of food they can eat if they have diabetes, which will be considered significant.The details sequence and consequence are shown in Fig.7.

    Figure 7: Flowchart of the food recommendation module

    3.3.2 Medicine Reminder Module

    People with diabetes have to remember a lot of things like taking medicine every day, diet control.However, people can’t remember many things, so it will be convenient if there is a system that will automatically remind you to complete a task.A reminder module has been developed focusing on these issues.The system will remind diabetic patients through an alarm notification according to the particular time they have to take medicines, take food, etc.SQLite database has been used to do this because it is very popular as an offline database.When users provide input for their daily records, the data will be stored in the local database, and the application will notify the user at a particular time.The functionality of the Medicine Reminder Module is shown in Fig.8.

    3.3.3 Answering Question Module

    The answering Question Module has been integrated with a chatbot to answer relevant diabetes and health questions.Google Dialogflow API has been combined for accomplishing the task.Dialogflow is a platform for natural language understanding that makes it simple for a mobile app, web application, bot, interactive voice response system, etc., to design and incorporate a conversational user interface.Dialogflow can evaluate various user input forms, including text or audio (such as phone or voice recording) inputs.A few ways, either by text or through synthetic expression, can also respond.We have predefined some of the frequently asked questions to answer when users ask questions automatically.The working procedure of this proposed module is shown in Fig.9.

    Figure 8: Flow chart of the medicine reminder module

    3.3.4 Daily Movement Activities Module

    The Daily Movement Activities Module has been developed to visualize users’daily movements and activities.Simply put, this module is designed to generate a report on how long someone is cycling or walking every day.Google Fit API has been integrated for this particular work & real-time tracking of the movement activities.To create smarter health and wellness apps, Google Fit provides Android and REST APIs.Google Fit APIs will help users find new ideas that they want to share with others.To help people exercise harder, eat healthy, stay calmer, and sleep better, use these experiences to create useful new features [29].In addition to these, this module incorporates a shaking feature by integrating the accelerometer sensor, which will detect a user’s current situation.Almost all smartphones have built-in accelerometer sensors, so combining them and using proper existing technology is good.To use the shaking feature, the user has to register some known numbers in the application.If there is any problem, double shaking the phone will send the location in the form of a message to the known person through current location tracking with Google map link.When people get older, they face various problems, especially diabetics who have to do regular activities such as morning walks, so it is crucial to track their location so that family members can be aware of their whereabouts.Fig.10 shows the functional explanation of the Daily Movement Activities Module.

    Figure 9: Flowchart of the answering question module

    3.3.5 Daily Awareness Module

    Diabetes is spreading more and more, especially in children and adolescents, and it is present in both type 1 and type 2.This means that the public does not really know or care enough about the disease or is unaware of its harmful aspects.This module informs the user through daily push notifications and makes them aware of what kind of food should be avoided.This module integrated the push notification API for accomplishing this task.The purpose of the daily notification is to enhance awareness regarding diabetes.Users need to keep the Internet open to receive notifications because the system will communicate with the cloud via an Internet connection.This module’s main objective is to inform people about diabetes, especially in rural areas, because rural people are less aware of significant diseases.The working procedure of this module is illustrated in Fig.11.

    Figure 10: Flowchart of the daily movement activities module

    3.3.6 Health Module

    Fig.12 illustrates the block diagram of diagnosis using Machine Learning.The figure is classified into three interconnected phases.Phase-1 provides a diagnosis of diabetes using experimental parameters.Phase-2 gives the data preprocessing methodology with data process, data resizing and data leveling.Eventually, Phase-3 focuses on the artificial diagnosis of diabetes using the traditional approaches.To predict diabetes, ML algorithms have been integrated into the Health Module.To put it more simply, this module has been trained through various machine learning algorithms.Finally, we have explored that the LightGBM & KNN Algorithm performs well with 90% accuracy for predicting diabetes of the individuals.The user has to enter their Age, Body mass index (BMI), Heart Rate, and Glucose.Our module will automatically notify whether anyone has diabetes.The prediction approach and how cooperates with mobile apps are shown in Fig.13.

    Figure 11: Flowchart of the daily awareness module

    4 Results and Analysis

    This section has been categorized into two parts, to illustrate, Observation of Machine Learning Approach(OMLA)and Observationofthe Proposed System(OPS).TheOMLAis further subdivided intofiveparts:Experimental Result,Model Performance,Comparative Analysis,Feature Importance,and Exploratory Data Analysis (EDA).The OPS is also sub-divided into three sections: Survey Data Analysis (SDA) and Developed System Interface (DSI).

    4.1 Observation of Machine Learning Approach (OMLA)

    The results obtained using the machine learning algorithm are discussed in this section.The subsection is classified into three interconnected parts.

    4.1.1 Experimental Result

    The classification report that was obtained during model training was addressed in this section.Various classification algorithms were evaluated in this analysis.So, the precision, recall, and F1-score of the algorithms are shown in Tab.3, where the accuracy“P”has been written.The recall is“R”the same way, and the F1-Score is“F1”, confusion matrix measures the accuracy of all types of classification algorithms.It’s consists of four values: True positive, false positive, true negative, and False-negative [30].Type 1 Error is defined as False Positive of the Confusion Matrix, and Type 2 Error is called False Negative [30].To evaluate the accuracy of a model via the Confusion matrix, such approaches are applied.Eqs.(9)-(12) show the formula for finding Precision, Recall, F-1 scores, and Accuracy.Fig.14 shows the visual representation of the accuracy score of Tab.3.

    Figure 12: Block diagram of diagnosis using machine learning

    Precision:The number of correct positive outcomes is divided by the number of correct positive outcomes predicted by the classifier.It is articulated as—

    Recall:The number of positive findings is accurate, split by all the related samples.It is given in mathematical form as—

    F1 score:It is applied to calculate the accuracy of a test.The Harmonic Mean between accuracy and recall is the F1 score.For the F1 score, the range is [0,1].It informs how accurate a classifier is and how robust it is.It is given, mathematically, as—

    Confusion Matrix:It provides a matrix as output and defines the model’s maximum efficiency.

    Figure 13: Working procedure of diabetes prediction using the Health Module

    Accuracy:How accurately the overall model can predict.It is possible to measure the matrix accuracy by taking the average values lying around the main diagonal.It is given as-

    4.1.2 Evaluating Model Performance

    Several elements are required to measure a model’s performance, for instance, ConfusionMatrix,Precision-Recall Curve, ROC Curve, and Cross-Validation.To begin, the Confusion Matrix, often known as the error matrix, is a tool for visualizing the performance of an algorithm.True positive (TP), True negative (TN), False positive (FP), and False negative (FN) are the four parameters.Let us discuss it as an example.Diabetic correctly identified as diabetic is True positive.Healthy people who are accurately identified as healthy are True negative, healthy people who are wrongly identified as diabetic are False positive, and diabetic people who are incorrectly identified as healthy are False-negative.We have already explained the Confusion Matrix in the Experimental Results section and described the mathematical equations of Precision, Recall, and F1-Score in the Eqs.(3)-(6), respectively.Precision quantifies the number of positive predictions for the class that is presently classified as positive.The recall is a measurement of how many positive class predictions were produced from all positive examples in the dataset.F-Measure provides a single score that takes into account both the accuracy and the number of recalls.Plotting the true positive rate (TPR)vs.the false positive rate (FPR) at various threshold values yields theROCCurve example.In another case, Precision-Recall Curve shows the tradeoff between precision and recall for the different thresholds to train and test the algorithm.

    Table 3: Classification report of the machine learning algorithms

    Furthermore, the original sample is randomly partitioned into k equal-sized subsamples in k-fold cross-validation.A single subsample from the k subsamples is kept as validation data for testing the model, while the remaining k-1 subsamples are used as training data.The cross-validation procedure is then performed k times, with each of the k subsamples serving as validation data exactly once.Then, to get a single estimation, the results from each k performance will be averaged.The main goal of repeated random subsampling is to use all observations for both training and validation and to verify each statement only once.There are several advantages of k-fold cross-validation; for example, the different output is available for different folds, so it is known how well the model will function overall, and K fold cross-validation can be used to avoid overfitting.A significant step in the production of a model is model assessment [31].It helps to select the best model for representing our data and forecastingthe selected model’s future performance.However, it is not good to evaluate model output on the data used for training in data science because this can lead to over-optimistic and over fitted models [32].Therefore, it is essential to measure the performance of the model.The above performance indicators are an essential and effective approach in data science, so we have assessed the model’s performance by following these criteria.

    Figure 14: Visual representation of the Accuracy score

    In our proposed research, Light Gradient Boosting Machine (LightGBM) is applied, a gradient boosting framework that uses tree-based learning algorithms.LightGBM, based on decision tree algorithms, is a fast, distributed, high-performance gradient boosting method used for ranking,classification, prediction, and many other machine learning tasks [33,34].It can manage large quantities of information and have greater precision than different decision tree gradient boosting models such as eXtreme Gradient Boosting (XGBoosting) [35].LightGBM can be used to address several issues, including binary classification, multi-classification, regression, and several others.The reason for including it in this study is that it has specific functionalities that are not seen in traditional algorithms.Faster training speed, higher performance, capacity to manage large-scale data, support for parallel and Graphics Processing Unit (GPU) learning, lower memory use, and enhanced accuracy are just a few of its characteristics.Since this research is included in the binary classification, the LightGBM has been considered for this, and accuracy has come in handy for its unique features.

    The average score of the ROC-AUC curve over LightGBM is marked as 0.948, and the average accuracy of cross-validation (5 folds) is seen as 0.89.On the other hand, the average score of the ROCAUC curve on KNN is 0.936, and the average accuracy after cross-validation (5 folds) is 0.90.There is no particular formula for calculating K’s value to the best of our knowledge, but it is a good idea to keep 10.A random function is applied to divide data into these many folds For example, suppose we have 10 data points in the data set, and K=5 is specified, then 10/5=2, so there will be 2 points that will be kept for testing for each fold and rest in training.Fig.15a shows the Model performance report of LightGBM (5 folds), Fig.15b shows the Model performance report of LightGBM & KNN.The detailed report of the model is shown in Fig.15.

    Figure 15: (a) Model performance report of light gradient boosting machine(LightGBM) (b) Model performance report of LightGBM & K-nearest neighbor(KNN)

    Tab.4 shows the cross-validation (5 folds) report of the LightGBM & KNN.Reports for LightGBM and KNN are generated individually by K (5) fold cross-validation and mean, and the standard deviation is found at the end of each fold experiment.Mean and standard deviations have also been made by cross-validating the top of the ROC-AUC curve.ROC stands for Receiver Operating Curve, and AUC stands for Area under Curve.Another method of determining how good the performance of different classification models is the ROC-AUC curve.

    Table 4: Cross-validation report of the LightGBM & KNN

    Figs.16 and 17 show the Discrimination threshold plot for LightGBM and Voting Classifier.The threshold of discrimination is a visualization of accuracy, recall, f1 ranking, and queue rate with respect to a binary classifier’s discrimination threshold.The discrimination threshold is the probability or score at which the positive class is chosen above the negative.Vote Classifier is a meta-classifier that uses majority or plurality voting to categorize comparable or conceptually dissimilarmachine learning classifiers.

    Figure 16: Discrimination threshold plot for LightGBM

    Figure 17: Discrimination threshold plot for voting classifier

    4.1.3 Comparative Analysis

    In this section, the result obtained by experimenting with several machine learning algorithms was compared with the previous research using the Pima Indian dataset.Tab.5 shows the comparative study of associated diabetes detection studies with the proposed dataset.The Table is divided based on some criteria such as Method, Accuracy, LightGBM approach, Integrating Mobile Apps, and Deployment Pipeline.We have compared our obtained accuracy with the previous study based on predicting diabetes.In this study, Light Gradient Boosting Machine (LightGBM) & KNN performs efficiently with 90% accuracy.By taking a close look at Tab.5, it can be identified that the performance of the previously published model is comparatively less than the model proposed in this research.In another case, we have indicated that all studies based on diabetes prediction have been completed in the past.They were primarily limited to simulations; however, there is no pipeline on predicting diabetes through mobile applications, and adequate research has not been completed yet.

    Table 5: A Comparative study of associated diabetes detection studies with the Pima Indian dataset

    Through this comparison, we have tried to highlight a few more things, such as the research that has been done in the past on diabetes prediction, especially on this dataset, to find information about their working methods.Many state-of-the-art techniques are currently playing a vital role in data science, so it is essential to find the previous work’s performance so that researchers can develop new solutions to the work in this particular field.

    We have used LightGBM, followed by the novel approach, and obtained 90% accuracy, so the previously published investigation was conducted based on traditional methods.No novel techniques have been used there.In addition to the other things we have analyzed through Comparison, the previous research has worked with a particular algorithm.Still, not all the possible algorithms have been explained in detail.We have designed our research as a benchmark; thus, it will be fruitful for the research community.The details sequence and consequence are shown in Tab.5.

    4.1.4 Feature Importance

    Feature choice is the process of reducing the number of input variables when creating a predictive model.It is advisable to scale the number of input variables to reduce the cost of modeling calculations and, in some cases, to increase the model’s effectiveness.The statistics-based feature selection method uses statistics for each input variable and target variable and selects the input variables with a substantial correlation to the target variable [44,45].Various feature choice techniques are available, such as Univariate Selection, Feature Significance, and Correlation Matrix with Heatmap.Several other online portals from which a wide variety of datasets are commonly available such as Kaggle and UCI Machine learning repository.Feature Importance provides a score to each of the data’s features; the higher the score, the more significant the feature is to the output variable.Tree-Based Classifiers have an inbuilt class called Feature Importance.Feature selection is a fundamental principle in traditionalML that profoundly affects the model’s efficiency [46,47].The data attributes utilized for training machine learningmodels have a big impact on the final output.Model output may be harmed by features that are insignificant or only partially significant.Fig.18 shows the Feature importance of the input variables.

    Figure 18: Feature significance of the input variables

    4.1.5 Exploratory Data Analysis

    Blood sugar and hypertension can be related terms for diabetic patients.It’s necessary to measure the blood sugar level along with blood pressure.There are three ways to increase blood pressure concerning the corresponding glucose level.The first one, when the blood vessels become losing their capability to stretch.The second one, the body’s fluid, will be increased if the diabetes is already affected by the kidneys.The third one is related to insulin resistance, which is also liable to increase hypertension risk.Again, Fig.19 shows a scattered diagram of Glucose levelvs.Body Mass Index(BMI) of our experimental data analysis Fig.20 presents the experimental prediction of the diabetic patient and healthy person.In our experiment, we have defined a threshold value for both glucose level and blood pressure.The experimental data shows 268 patients have a potential risk of diabetes of 92.2%.

    Figure 19: Data analysis based on glucose vs. BMI

    Figure 20: Data analysis based on glucose and blood pressure

    On the other hand, 500 patients have less risk of diabetes than the previous one.Thus, we can mark them as healthy person.

    4.2 Observation of the Proposed System (OPS)

    4.2.1 Survey Data Analysis

    The upper Tab.6 shows the portion of answered questions’responses and analyzing data with different criteria, Necessity and Impactful, which filter survey data.Fig.21 shows a pie shows the result of the survey with the features of Necessity and Impactful.Again, we have calculated votes on criteria, Necessity and Impactful, which filter survey data.Fig.21 shows a pie shows theresult of the survey with the features of Necessity and Impactful.Again, we have calculated votes on 158 responses, we have tracked out that almost 100% of people think about diet plan features.More than 80% of people endorse doctors, and around 80% of people believe in consultation, and so on.We have examined that“track activities”are less popular than any other proposed solution features.Fig.22 shows the corresponding results of the survey.

    Table 6: Analyzed number portion using criteria

    4.2.2 Developed System Interface (DSI)

    The manuscript also ensures the embodiment of the proposed solution.We have first designed the User Interface (UI) design of the proposed mobile application in our development process.After that,we have collected a dataset from the website, and amachine learning design was performed.After that,we have analyzed the experimental data and interpreted it accordingly.Fig.23 shows some interfaces of our proposed developed mobile application.

    Figure 21: Pie chart based on the average response of the survey

    Figure 22: Measuring the feature ranking by conducting the survey

    Figure 23: Interface of the proposed mobile application

    5 Conclusion

    Health is more important than other activities.However, many developing countries suffer from low-quality health services because they allocate a smaller portion of their budget to the health sector.Consequently, the citizen of these countries cannot concentrate on their work because of these difficulties.This work proposes a mobile application based on modern computation, which is straightforward to predict diabetes.The proposed model has been enriched with one of the most emergent technologies, such as a Machine Learning based system to find accurate prediction levels on diabetes.To accomplish this goal, some data preprocessing operations have been interpreted on the dataset.Also, several machine learning algorithms have been utilized in this work to track better accuracy in the diagnosis of diabetes diseases.The proposed model has 90% accuracy on the K nearest neighbors algorithm (KNN) & Light Gradient Boosting Machine (LightGBM).A comparison of the accuracy of the machine learning algorithms has been enumerated with the existing study to ensure this work’s novelty.The research has further performed a survey data analysis on consciousness and awareness of public health-related mobile applications and diabetes.The associated resulting data has also been in this manuscript.Though the proposed work has better accuracy in predicting a patient’s diabetes, the model also has a set of limitations.First of all, the dataset utilized in this study needs to bring many more data preprocessing changes to increase the model’s accuracy.Second, this application has been designed only for educated people on the clinical trial.In the future, this research will overcome these two issues and present a common platform for both educated and uneducated people.Also,this research will present the effectiveness of public health with the developed application through System Usability Scale(SUS).However, this research’s objective is achieved, and the proposed solution can be adjustable in the daily activities of a diabetic patient.

    Funding Statement:The authors received no specific funding for this study.

    Conflicts of Interest:The authors of this study declare that they have no conflict of interest.

    99re在线观看精品视频| 九色国产91popny在线| 十八禁网站免费在线| 波多野结衣高清无吗| ponron亚洲| 黑人巨大精品欧美一区二区mp4| 一边摸一边做爽爽视频免费| 白带黄色成豆腐渣| 在线观看66精品国产| 天堂影院成人在线观看| 欧美人与性动交α欧美精品济南到| 国产精品久久久人人做人人爽| 12—13女人毛片做爰片一| 激情在线观看视频在线高清| 女性被躁到高潮视频| 精品不卡国产一区二区三区| 欧美日韩黄片免| 18禁观看日本| 两人在一起打扑克的视频| 宅男免费午夜| ponron亚洲| 又紧又爽又黄一区二区| 亚洲一区中文字幕在线| 91字幕亚洲| 国产免费av片在线观看野外av| 91av网站免费观看| 中出人妻视频一区二区| 岛国在线观看网站| 精品国产一区二区三区四区第35| 69av精品久久久久久| 国产亚洲欧美98| 一级毛片女人18水好多| 一本综合久久免费| 香蕉丝袜av| 久久精品aⅴ一区二区三区四区| 精品国产美女av久久久久小说| 欧美黑人欧美精品刺激| 亚洲av日韩精品久久久久久密| 日韩欧美免费精品| 久久天躁狠狠躁夜夜2o2o| 日本一区二区免费在线视频| 精品欧美一区二区三区在线| 国产亚洲欧美精品永久| 欧美中文日本在线观看视频| 久久九九热精品免费| 久久性视频一级片| 精品无人区乱码1区二区| 亚洲av电影不卡..在线观看| 老司机深夜福利视频在线观看| 亚洲 国产 在线| 波多野结衣高清作品| 99riav亚洲国产免费| 国产成人影院久久av| 亚洲欧美激情综合另类| 国产亚洲精品综合一区在线观看 | 亚洲最大成人中文| 男女下面进入的视频免费午夜 | 中文字幕久久专区| 亚洲电影在线观看av| 色综合亚洲欧美另类图片| 久久久久久久精品吃奶| 在线播放国产精品三级| 国产视频一区二区在线看| 成人三级做爰电影| 成人国产综合亚洲| 18禁国产床啪视频网站| 午夜精品在线福利| 国产成人精品久久二区二区免费| 国产三级在线视频| 亚洲七黄色美女视频| e午夜精品久久久久久久| 在线国产一区二区在线| 亚洲自拍偷在线| 亚洲精品国产精品久久久不卡| 1024手机看黄色片| 日日爽夜夜爽网站| 免费看a级黄色片| 国产精品自产拍在线观看55亚洲| 狠狠狠狠99中文字幕| 国产视频内射| 99国产精品一区二区三区| 久久精品人妻少妇| 亚洲熟妇中文字幕五十中出| 青草久久国产| 国产1区2区3区精品| 少妇粗大呻吟视频| 90打野战视频偷拍视频| 可以免费在线观看a视频的电影网站| 久久久久免费精品人妻一区二区 | 成人永久免费在线观看视频| 又紧又爽又黄一区二区| 夜夜躁狠狠躁天天躁| 亚洲国产欧洲综合997久久, | 久久精品国产亚洲av高清一级| 性色av乱码一区二区三区2| 久久精品成人免费网站| 日本一区二区免费在线视频| 亚洲男人的天堂狠狠| 99精品在免费线老司机午夜| 黄色a级毛片大全视频| 黄片大片在线免费观看| 性欧美人与动物交配| 最新美女视频免费是黄的| 亚洲av成人一区二区三| 国产99白浆流出| www.自偷自拍.com| 亚洲av美国av| 国产成人精品无人区| 人成视频在线观看免费观看| 日韩视频一区二区在线观看| 国产精品久久久久久人妻精品电影| 淫秽高清视频在线观看| 18禁国产床啪视频网站| 亚洲成人国产一区在线观看| 十八禁网站免费在线| 69av精品久久久久久| 精品熟女少妇八av免费久了| 一本综合久久免费| 男人舔女人下体高潮全视频| a在线观看视频网站| 人人妻人人看人人澡| 黄色 视频免费看| 亚洲av成人av| 在线看三级毛片| 一本综合久久免费| 丰满的人妻完整版| 一进一出抽搐动态| 欧美国产精品va在线观看不卡| 久久国产精品男人的天堂亚洲| 日本a在线网址| 国产精品久久电影中文字幕| 老汉色∧v一级毛片| 国产黄a三级三级三级人| 欧美日韩福利视频一区二区| 欧美性长视频在线观看| 中文字幕人成人乱码亚洲影| 在线观看免费日韩欧美大片| 精品久久久久久久久久免费视频| 亚洲精品在线观看二区| 亚洲黑人精品在线| 九色国产91popny在线| 亚洲精品中文字幕在线视频| 国产精品av久久久久免费| 特大巨黑吊av在线直播 | 国产精品精品国产色婷婷| 狂野欧美激情性xxxx| 欧美乱码精品一区二区三区| 日韩欧美一区视频在线观看| 日韩大码丰满熟妇| 啦啦啦观看免费观看视频高清| 日韩免费av在线播放| 午夜福利免费观看在线| 亚洲国产中文字幕在线视频| 亚洲人成网站在线播放欧美日韩| 好男人电影高清在线观看| 女人高潮潮喷娇喘18禁视频| 成人永久免费在线观看视频| 悠悠久久av| 波多野结衣av一区二区av| 婷婷六月久久综合丁香| 性色av乱码一区二区三区2| av超薄肉色丝袜交足视频| 欧美丝袜亚洲另类 | 欧美日韩亚洲综合一区二区三区_| 高清毛片免费观看视频网站| 久久精品国产99精品国产亚洲性色| 啦啦啦韩国在线观看视频| 亚洲五月色婷婷综合| 国产精品久久久av美女十八| 18禁裸乳无遮挡免费网站照片 | 日韩 欧美 亚洲 中文字幕| 视频在线观看一区二区三区| 免费观看精品视频网站| 久久欧美精品欧美久久欧美| 午夜福利在线观看吧| 一级毛片女人18水好多| 中文字幕人成人乱码亚洲影| av福利片在线| 免费无遮挡裸体视频| 欧美激情 高清一区二区三区| 在线观看日韩欧美| 国产精品一区二区精品视频观看| 久久久久久大精品| 欧美国产日韩亚洲一区| 国产精品久久久久久亚洲av鲁大| 国产精品 国内视频| 黑人欧美特级aaaaaa片| 波多野结衣av一区二区av| av超薄肉色丝袜交足视频| 欧美一级a爱片免费观看看 | 18禁美女被吸乳视频| 国产av一区在线观看免费| 每晚都被弄得嗷嗷叫到高潮| 免费人成视频x8x8入口观看| 国产1区2区3区精品| www日本黄色视频网| 丝袜人妻中文字幕| 搡老熟女国产l中国老女人| 日韩欧美一区视频在线观看| 99热6这里只有精品| 成人国产综合亚洲| 国产精品二区激情视频| 成人三级做爰电影| 久久精品人妻少妇| 一本久久中文字幕| 亚洲自偷自拍图片 自拍| 91麻豆精品激情在线观看国产| 久久九九热精品免费| 韩国av一区二区三区四区| 日韩欧美一区视频在线观看| 国产视频内射| 麻豆久久精品国产亚洲av| 国产精品精品国产色婷婷| 成年免费大片在线观看| 中文字幕最新亚洲高清| 91国产中文字幕| 99久久综合精品五月天人人| 热re99久久国产66热| 成年版毛片免费区| 欧美日本亚洲视频在线播放| 国产一区二区在线av高清观看| 亚洲五月婷婷丁香| 91大片在线观看| 一边摸一边做爽爽视频免费| 一a级毛片在线观看| 欧美日韩亚洲综合一区二区三区_| 精品久久久久久久末码| 国产一区二区激情短视频| 老鸭窝网址在线观看| 欧美成人免费av一区二区三区| 成年版毛片免费区| 午夜免费观看网址| 91国产中文字幕| 国产成人av激情在线播放| 日本黄色视频三级网站网址| 国产一卡二卡三卡精品| 男人的好看免费观看在线视频 | 亚洲天堂国产精品一区在线| 国产午夜福利久久久久久| 色综合站精品国产| 精品一区二区三区四区五区乱码| 黑人操中国人逼视频| 欧美日本亚洲视频在线播放| 国产精品自产拍在线观看55亚洲| 女人被狂操c到高潮| 日韩国内少妇激情av| 搡老岳熟女国产| 俄罗斯特黄特色一大片| 欧美中文日本在线观看视频| 欧美精品啪啪一区二区三区| 精品国产美女av久久久久小说| 欧美日韩亚洲综合一区二区三区_| 午夜激情福利司机影院| 欧美大码av| 亚洲三区欧美一区| 宅男免费午夜| 欧美乱妇无乱码| 夜夜躁狠狠躁天天躁| 国产99白浆流出| 99精品欧美一区二区三区四区| 99热6这里只有精品| 女同久久另类99精品国产91| 国产伦人伦偷精品视频| 久久国产亚洲av麻豆专区| 少妇被粗大的猛进出69影院| 婷婷丁香在线五月| 日本五十路高清| 男人的好看免费观看在线视频 | 男人舔奶头视频| 日韩大码丰满熟妇| 757午夜福利合集在线观看| 亚洲av成人av| 美女扒开内裤让男人捅视频| 免费看a级黄色片| 午夜精品在线福利| 亚洲人成网站在线播放欧美日韩| 欧美在线一区亚洲| 欧美乱码精品一区二区三区| 亚洲中文av在线| 日本在线视频免费播放| 波多野结衣av一区二区av| 午夜精品久久久久久毛片777| 女性被躁到高潮视频| 欧美日韩中文字幕国产精品一区二区三区| 精品第一国产精品| 久久精品aⅴ一区二区三区四区| 免费在线观看亚洲国产| 999久久久国产精品视频| 欧美黑人巨大hd| 亚洲午夜理论影院| 女性生殖器流出的白浆| 日韩大尺度精品在线看网址| 91麻豆精品激情在线观看国产| 久久久久久亚洲精品国产蜜桃av| 97碰自拍视频| 一级作爱视频免费观看| 国产99白浆流出| 啪啪无遮挡十八禁网站| 中文字幕精品免费在线观看视频| 一个人免费在线观看的高清视频| 日日爽夜夜爽网站| 免费高清视频大片| 亚洲人成网站高清观看| 丰满的人妻完整版| 国产精品久久电影中文字幕| 国产真人三级小视频在线观看| tocl精华| 日韩中文字幕欧美一区二区| 人成视频在线观看免费观看| 少妇熟女aⅴ在线视频| 久久久久久久午夜电影| 国产激情欧美一区二区| 国产精品二区激情视频| 精品第一国产精品| 国产精品免费一区二区三区在线| 亚洲第一电影网av| 久久精品国产亚洲av香蕉五月| 在线av久久热| 国产野战对白在线观看| 日韩一卡2卡3卡4卡2021年| 老汉色∧v一级毛片| 一边摸一边抽搐一进一小说| 国语自产精品视频在线第100页| 久久中文字幕一级| 国产亚洲精品综合一区在线观看 | 制服诱惑二区| 亚洲精品在线美女| 大香蕉久久成人网| 熟女电影av网| 十八禁人妻一区二区| 欧美激情久久久久久爽电影| 日日干狠狠操夜夜爽| 欧美日本视频| videosex国产| 免费一级毛片在线播放高清视频| 亚洲精品粉嫩美女一区| 两个人看的免费小视频| 亚洲,欧美精品.| svipshipincom国产片| 亚洲国产精品999在线| 无人区码免费观看不卡| 天天躁夜夜躁狠狠躁躁| 久久性视频一级片| 久久久久免费精品人妻一区二区 | 天天一区二区日本电影三级| 亚洲 欧美 日韩 在线 免费| 色综合站精品国产| 亚洲成国产人片在线观看| 欧美精品亚洲一区二区| 90打野战视频偷拍视频| 精品第一国产精品| 免费看a级黄色片| 亚洲aⅴ乱码一区二区在线播放 | 99热6这里只有精品| www日本在线高清视频| 宅男免费午夜| 免费在线观看黄色视频的| 亚洲专区字幕在线| 欧美激情 高清一区二区三区| bbb黄色大片| 欧美激情久久久久久爽电影| 欧美黑人欧美精品刺激| 最新在线观看一区二区三区| 熟女电影av网| 天天添夜夜摸| 久久久久精品国产欧美久久久| 国产精品综合久久久久久久免费| av视频在线观看入口| 看免费av毛片| 99re在线观看精品视频| 在线观看日韩欧美| 婷婷精品国产亚洲av| 哪里可以看免费的av片| www国产在线视频色| 久久香蕉精品热| 岛国视频午夜一区免费看| 精品人妻1区二区| 黄频高清免费视频| 精品一区二区三区四区五区乱码| 亚洲午夜精品一区,二区,三区| 男女做爰动态图高潮gif福利片| 久久久精品欧美日韩精品| 美女高潮喷水抽搐中文字幕| 俺也久久电影网| 亚洲 欧美一区二区三区| 日本免费一区二区三区高清不卡| 久久中文字幕一级| 九色国产91popny在线| 免费看a级黄色片| 高清毛片免费观看视频网站| 91九色精品人成在线观看| 久久精品国产清高在天天线| xxxwww97欧美| 亚洲精品在线美女| 一夜夜www| 免费观看精品视频网站| 国产不卡一卡二| 欧美日韩福利视频一区二区| 777久久人妻少妇嫩草av网站| 欧美激情极品国产一区二区三区| 亚洲avbb在线观看| 伊人久久大香线蕉亚洲五| 久久精品91无色码中文字幕| 看片在线看免费视频| 男女做爰动态图高潮gif福利片| 高清毛片免费观看视频网站| 一区二区三区激情视频| 1024手机看黄色片| 成人国产一区最新在线观看| 国产97色在线日韩免费| 久久国产精品男人的天堂亚洲| 欧美成人免费av一区二区三区| 99re在线观看精品视频| 在线播放国产精品三级| 亚洲,欧美精品.| 黄色a级毛片大全视频| 亚洲精品美女久久久久99蜜臀| 亚洲在线自拍视频| 国产精品日韩av在线免费观看| 99国产精品99久久久久| 久久香蕉国产精品| 亚洲国产日韩欧美精品在线观看 | 高清在线国产一区| 一进一出好大好爽视频| 怎么达到女性高潮| 亚洲av日韩精品久久久久久密| 亚洲国产欧美网| 久久中文字幕人妻熟女| 日韩国内少妇激情av| 日韩av在线大香蕉| 麻豆成人av在线观看| 久久久久国内视频| 亚洲国产欧美日韩在线播放| 午夜激情福利司机影院| 丁香欧美五月| 国产成人啪精品午夜网站| 一进一出好大好爽视频| 免费在线观看日本一区| 嫩草影院精品99| 又黄又粗又硬又大视频| 欧美黑人巨大hd| 熟女电影av网| 免费在线观看黄色视频的| 日韩精品中文字幕看吧| 精品久久久久久久毛片微露脸| 精品国产国语对白av| 亚洲九九香蕉| 欧美激情久久久久久爽电影| 18禁裸乳无遮挡免费网站照片 | 亚洲国产精品成人综合色| 国产精品一区二区精品视频观看| 国产亚洲av嫩草精品影院| 一进一出抽搐gif免费好疼| 三级毛片av免费| 午夜影院日韩av| 午夜福利成人在线免费观看| 免费电影在线观看免费观看| 啦啦啦观看免费观看视频高清| 成人免费观看视频高清| 不卡av一区二区三区| 亚洲av第一区精品v没综合| 国产乱人伦免费视频| 亚洲五月婷婷丁香| www.自偷自拍.com| 91麻豆av在线| 精品乱码久久久久久99久播| 一区二区三区国产精品乱码| 免费女性裸体啪啪无遮挡网站| 亚洲色图av天堂| 成人手机av| 日本免费一区二区三区高清不卡| 国产成人欧美在线观看| 久久精品国产亚洲av香蕉五月| 国产精品一区二区免费欧美| 露出奶头的视频| 亚洲最大成人中文| 桃红色精品国产亚洲av| 此物有八面人人有两片| 精品欧美一区二区三区在线| 国产精品,欧美在线| 看免费av毛片| 欧美日韩乱码在线| 国产亚洲av高清不卡| 99精品久久久久人妻精品| av中文乱码字幕在线| 亚洲av电影不卡..在线观看| 亚洲精品色激情综合| 国产成人精品久久二区二区免费| 大型av网站在线播放| 日本黄色视频三级网站网址| 精品无人区乱码1区二区| 国产一区在线观看成人免费| 人人澡人人妻人| 亚洲人成77777在线视频| 欧美激情久久久久久爽电影| 欧美乱妇无乱码| 99国产极品粉嫩在线观看| 校园春色视频在线观看| 美女高潮喷水抽搐中文字幕| 中文资源天堂在线| 麻豆av在线久日| 久久久久久人人人人人| 91麻豆av在线| 亚洲国产欧洲综合997久久, | 男女视频在线观看网站免费 | 亚洲av成人av| 无限看片的www在线观看| 18美女黄网站色大片免费观看| 麻豆久久精品国产亚洲av| 桃色一区二区三区在线观看| 他把我摸到了高潮在线观看| 波多野结衣巨乳人妻| 一夜夜www| 亚洲电影在线观看av| 在线看三级毛片| 久久精品国产亚洲av香蕉五月| 脱女人内裤的视频| 免费女性裸体啪啪无遮挡网站| 女同久久另类99精品国产91| 美女 人体艺术 gogo| 国产精品香港三级国产av潘金莲| 中文字幕久久专区| 亚洲欧美一区二区三区黑人| 久久久久久免费高清国产稀缺| e午夜精品久久久久久久| 国产精品久久久人人做人人爽| 黄色毛片三级朝国网站| 99国产精品一区二区蜜桃av| 日韩欧美 国产精品| 精品久久蜜臀av无| 美女高潮喷水抽搐中文字幕| 国产伦在线观看视频一区| 这个男人来自地球电影免费观看| 久久欧美精品欧美久久欧美| 国内久久婷婷六月综合欲色啪| 国产一区二区在线av高清观看| 欧美精品啪啪一区二区三区| 最近在线观看免费完整版| 波多野结衣av一区二区av| 国产成人啪精品午夜网站| 国产免费av片在线观看野外av| 正在播放国产对白刺激| 久久中文看片网| 男女那种视频在线观看| 看免费av毛片| 亚洲,欧美精品.| 亚洲欧洲精品一区二区精品久久久| 在线永久观看黄色视频| 午夜久久久在线观看| 亚洲欧美日韩无卡精品| 中出人妻视频一区二区| 精华霜和精华液先用哪个| 国产成人啪精品午夜网站| 亚洲片人在线观看| 欧美另类亚洲清纯唯美| 精品国产一区二区三区四区第35| 欧美成人免费av一区二区三区| 97人妻精品一区二区三区麻豆 | 日本一区二区免费在线视频| 日韩精品青青久久久久久| 成人三级黄色视频| 欧美日韩中文字幕国产精品一区二区三区| 国产伦在线观看视频一区| 色播亚洲综合网| 国产黄片美女视频| 日韩高清综合在线| 动漫黄色视频在线观看| 在线观看免费午夜福利视频| 村上凉子中文字幕在线| 精品日产1卡2卡| 草草在线视频免费看| 国内精品久久久久精免费| 亚洲国产毛片av蜜桃av| 亚洲欧美精品综合久久99| 一a级毛片在线观看| 少妇的丰满在线观看| 亚洲国产毛片av蜜桃av| 亚洲avbb在线观看| 久热这里只有精品99| 一卡2卡三卡四卡精品乱码亚洲| 久久国产精品男人的天堂亚洲| 亚洲av电影不卡..在线观看| 精品国产一区二区三区四区第35| 99热这里只有精品一区 | 1024香蕉在线观看| 九色国产91popny在线| 欧美黑人欧美精品刺激| 精品国产乱子伦一区二区三区| 母亲3免费完整高清在线观看| 亚洲精品一卡2卡三卡4卡5卡| 又大又爽又粗| 妹子高潮喷水视频| 亚洲成人久久爱视频| 亚洲精品在线观看二区| 国产v大片淫在线免费观看| 国产蜜桃级精品一区二区三区| 成人精品一区二区免费| 国产成人影院久久av| 欧美av亚洲av综合av国产av| 99精品在免费线老司机午夜| 久久天躁狠狠躁夜夜2o2o| 2021天堂中文幕一二区在线观 | 欧美性猛交╳xxx乱大交人| 国产亚洲av高清不卡| www国产在线视频色| 国产一区二区在线av高清观看| 国产精品免费视频内射| 十分钟在线观看高清视频www| 老汉色av国产亚洲站长工具| 成人国产综合亚洲| 老司机午夜十八禁免费视频| 深夜精品福利| 久久久久国产精品人妻aⅴ院| 免费在线观看黄色视频的| 91麻豆精品激情在线观看国产|