Abstract
Lifestyle influences morbidity and mortality rates in the world. Physical activity, a healthy weight, and a healthy diet are key preventative health behaviours that help reduce the risk of developing type 2 diabetes and its complications, such as cardiovascular disease. A healthy lifestyle has been shown to prevent or delay chronic diseases and their complications, but few people follow all recommended self-management behaviours. This work seeks to improve knowledge of factors affecting type 2 diabetes self-management and prevention through lifestyle changes. This paper describes the design, development, and testing of a diabetes self-management mobile app. The app tracked dietary consumption and health data. Bluetooth movement data from a pair of wearable insole devices are used to track carbohydrate intake, blood glucose, medication adherence, and physical activity. Two machine learning models were constructed to recognise sitting and standing. The SVM and decision tree models were 86% accurate for these tasks. The decision tree model is used in a real-time activity classification app. It is exciting to see more and more mobile health self-management apps being used to treat chronic diseases.
1. Introduction
Diabetes is a life-altering illness that some people find difficult to cope with, which can lead to the emergence of depressive symptoms. People with diabetes are two to three times more likely than nondiabetics to suffer from depression, and only about 25 to 50 percent of those suffering from depression will be detected and treated [1]. Depression can be brought on by diabetes management obligations, but it can also be brought on by outside stressors. Factors like financial stress, food hardship, family separation, and immigration can cause depression in people with Hispanic ancestry who have type II diabetes. A patient may be reluctant to seek therapy because of cultural pressures such as the negative stigma associated with mental health in the Mexican culture. 7.3 percent of Hispanic people used mental health services in 2008–2012, according to the National Survey on Drug Use and Health. This compares to white adults, who used mental health services at a rate of 16.6 percent (Substance Abuse and Mental Health Services Administration) [3]. If a person’s diabetes care is neglected because of depression, it can lead to binge eating, missed doctor visits, and social isolation, all of which can worsen a person’s health even more.
In order to perceive, feel, and act effectively, people need to be healthy. This is why health is important not just for the development of the individual but also for the environment to which they belong. The World Health Organization (WHO) defines health as “a state of complete physical, mental, and social well-being, and not merely the absence of disease or disability.” A healthy body is always changing and adapting to the pressures and changes in the environment, according to WHO, in order to maintain a balance within called homeostasis.
Taking care of one’s health or one’s well being by diagnosing, treating, and preventing sickness or other physical or mental impairments is known as healthcare. There is a study conducted in presenting primary, secondary, and tertiary care as well as in public health that is included [4]. In order to validate the appropriate healthcare distribution based on parameter monitoring and direct provisioning of scientific aid, it is critical to provide sufficient methods and betokens.
Information and communication technologies (ICTs) are being adopted by industries across the globe to improve transportation and increase competitiveness. This rule applies to all aspects of life, including health treatment. ITs have the potential to fundamentally alter the way healthcare is provided and managed. eHealth refers to the application of information and communication technologies (ICTs) in the delivery of fitness care services. Digital verbal interchange and information science are used in the health zone for clinical, educational, and administrative reasons at both the local site and at a distance according to the WHO definition of eHealth [5]. To put it another way, eHealth is the capability of ensuring that the correct fitness statistics are provided to the correct person at the correct time and place in a secure electronic structure in order to optimise fitness care delivery, research, training, and knowledge satisfaction and effectiveness.
Using the newest breakthroughs in information technology, remote tracking and monitoring collect patient data outside of the traditional healthcare setting. Simple remote patient monitoring systems that make use of consumer-grade equipment are the most effective. Most health monitoring systems make use of technology that is designed to be comfortable for patients [6].
2. Background
Diabetes self-management can be greatly improved with the use of smartphones and modern information technology. Exciting results from Kirwan and colleagues’ research show that a diabetes self-management mobile app can reduce HbA1c levels in patients [7]. However, medical professionals, academics, and app developers must collaborate to create a better diabetes self-management mobile app, with consideration for the target audience’s age [8].
Controlling blood sugar at a healthy level is difficult for diabetics for a variety of reasons. As a result of the increased risk of cardiovascular, renal, and neurological illnesses that are associated with high blood sugar levels [9], a study conducted by Basnet et al. [10] looked into the impact on type 1 diabetics using a smartphone app for diabetes self-management. Patients from Diabetes Australia in New South Wales and Queensland nominated participants.
Patients with type 1 diabetes who have had their diagnosis for at least six months and have glycosylated haemoglobin levels (HbA1c) greater than 7.5 percent must be eligible. Participants must also have access to a smartphone. Figure 1 shows an existing method based on the cloud system.
There were two groups of participants. Using an iPhone app for diabetes self-management and help from a Certified Diabetes Educator (CDE) using the data generated by the app is one option for intervention. The other is a control group that continues to get the same level of treatment as before [11]. It is called “Glucose Buddy,” and it has useful features like letting you manually enter your blood glucose levels, insulin dosages, nutrition, and viewing a personalised health information graph.
It is still a new field of healthcare, even though mobile health (mHealth) offers so many advantages to patients and healthcare providers alike. Many difficulties and challenges are still unresolved or unclear. There are many issues that need to be addressed, such as how reliable and efficient the service will be, how good the quality of the service will be, and how full the promised functionality will be for patients. In addition, as mHealth handles private information from patients, the topic of privacy security comes up frequently [12].
3. Related Work
Patients with diabetes who have been diagnosed by a doctor are said to have morbidity, but diabetics who have not been diagnosed do not. Adults aged 20 and over had a morbidity rate of 15% between 2013 and 2016, with Hispanic or Latino people of Mexican heritage having a morbidity rate of 19%. 83,564 people died in 2017 as a result of diabetes in the United States, or 25.7 fatalities per 100,000 persons, according to a National Center for Health Statistics report released in 2019 [13]. It is estimated that the Hispanic population has an average life expectancy of 81.8 years, which has remained stable over time due to underreporting because they are the most uninsured of any racial group and less likely to seek medical attention.
On the traditional and cloud computing environments, an overview of the IERF framework is shown in Figure 2 to improve classification accuracy and provide improved resource allocation. The University of California at Irvine (UCI) website was used to collect data for testing. There are two diabetes datasets included in this study: one from the Pima Indian community in Arizona and the other from the United States [14]. There is a description of the characteristics provided. The PID dataset includes 768 instances with 8 different attributes. It was given to the university as a gift in 1990.
4. Methodology
It combines the advantages of the genetic algorithm and rough set theory’s relative reduct algorithm in a single model. Conditional and decision qualities are the two categories into which the attributes fall. The value of the variable R is set to 0 at the beginning of the programme. And the value 0 is assigned to the variable best. T contains R and another temporary variable, tmp, and contains the value that is now the best. Each attribute is removed one at a time, and the dataset is evaluated for consistency in the decision table. Attributes are eliminated, and the smaller dataset is saved if the decision table is consistent. Once the classification accuracy of the conditional attributes has been verified, the result is compared to the decision attributes [15]. The first generation of children is generated if the classification accuracy of the conditional attributes acquired is higher than the classification accuracy of the decision attributes. Attributes are chosen from among the resulting children. The traits that have been chosen are subject to mutation and cross over.
T holds the resulting reduct set. A selection of the best decision attributes is kept in a variable called best. T gives R the reduct set it received, and R uses it. Once the optimal dataset is found, the process is repeated indefinitely. The algorithm comes to an end once the optimal set of data has been obtained. The reduct set R is the final result.
After collecting and preprocessing the data, the dataset will be trained using some supervised learning algorithms like KNN, decision trees, and ANN. Then, the accuracy of each model’s predictions will be compared, and the best model will be chosen to implement on the dataset in order to predict future occurrences [16].
There should be the provider side’s key objective to create a straightforward approach for providers to monitor the education and communication with their patients from many angles. The features on the provider’s end give medical professionals a slew of useful options [17–21]. As a result, patients can benefit from high-quality services and successful therapy. With smartphones as the primary device of both health care providers and patients, both sides of the equation require an internet connection. A user’s data is transferred to the backend server when they utilise services such as adding a meal or uploading their blood pressure. The data are then returned to the frontend server when the server has finished processing it.
Clients and servers can exchange data by utilising the Hypertext Transfer Protocol (HTTP). Next, the server will handle HTTP requests and use SQL to query the database (Structured Query Language). Our applications frequently make use of GET, POST, PUT, and DELETE methods. Use the GET function to get data from a particular resource. Figure 3 shows an example of a client using the GET method to receive the day’s meal information in the second phone screen capture. The server receives the GET service and processes it on its own before using SQL to retrieve the relevant data from our database’s Meal table and present it to the client. The data are created and updated using POST and PUT methods. Think about when people add or change their daily meals, blood pressure, or sugar levels. The POST or PUT method is used to send data from the client to the server. With POST, you can produce data numerous times, however, with PUT, you can only make one copy of the same thing. That is the sole difference between them. If you are not familiar with SQL, the DELETE command is simpler to grasp because it simply removes specified data from the database.
5. Results
SVM and decision tree models used ratios of 10 : 90, 15 : 85, 20 : 80, 25 : 75, and 30 : 70, respectively. The decision tree and support vector machine models were tested on several ratios of tests to training data. Finally, two models were built. A sensor and accelerometer data produced it. For each sensor, the mean and median time differences were used to measure the accuracy of sensor-based activity detection. These 16 sensor data features correctly detected activity. To detect non-foot-movement activities, only sensor-based data were employed, but no accelerometer data were employed (e.g., sitting).
Various test train ratios were attempted during model training to obtain the most accurate model. 15% and 20% test to train yielded the best results (accuracies of 90–100 percent). Two sets of training sessions were used to build the machine learning models: seven activities employing sensor data and accelerometer.
The system correctly detected walking and running in 31 situations, One walking sample and twenty-six running samples were taken Each test train ratio had five classification model observations on average.
Figures 4 and 5 show the results for four test train ratios (20, 5, 10, and 90). The decision tree and support vector machine models were put to the test using different test-to-training data ratios. In Figures 6 and 7, it is possible to state that two models were created. They were created using data from a sensor and an accelerometer. Each sensor has its own set of parameters. The accuracy of accelerometer and sensor data categorising walking and running was 96%. In 31 cases, the technology identified walking and running, one walking and 26 running samples.
The 15 sensor readings were 80% accurate.
Three of the seven samples correctly identified stair descent, three stair climbings, and one walking. The model had no false positives but missed several stair descending samples (1 : 0.43) for elliptical use. The samples were labelled as standing, walking, and sitting. Precision-recall ratio was 0.80 : 0.76 for non-elliptical use.
That means, the precision-to-recall ratio is 0 to 0. Because sitting does not create pressure sensor data, the analysis relies on accelerometer data. Without accelerometer data, sedentary movement cannot be recognised, as shown in Table 1.
The test data generated four decision-tree models. Among all 43 features, these models had the best walking and running detection results (100%). Activities detected with 90% to 100% accuracy with 43 characteristics. Using only sensor data, the 2-activity and 7-activity models obtained accuracy rates of 80–90% and 29%, respectively. Figure 6 shows the investigation’s results integrated with a decision tree.
Based on model quality assessments, BeticTrack incorporated the MLM created from all 43 features. The model was then put to the test in real time to classify things. This categorisation was based on time. For example, a person performed the seven tasks over time, and the classification result was saved in the database. We tallied the total time spent on each activity to ensure that the categories were accurate. Figure 7 shows the classification after a two-minute stroll. The MLM classed most walking as elliptical (over the course of 1 minute and 45 seconds). This is a 10 second stroll upstairs.
The model correctly identified standing data, as shown in Figure 7, with a 4.1% error rate. This could be due to the dataset’s size. The model’s classifications could be enhanced by collecting more data from more people over time.
The shoe sole distinguishes between sitting and running with near-perfect precision. Sitting has no sensor data compared to the other 7 activities. During application execution, running data contain the truest events. If the sensor feels pressure, it accepts it. It simply has two values: 0 and 1. If pressed on the ground while running, it returns 1. For example, standing, walking, or using stairs yields zero data from the pressure sensor. That is because the classification attempt failed. The use of elliptical data and classification of walking data overlaps. Eliminating elliptical data from classification may improve the classification of walking activity because walking and elliptical use are similar. This study’s purpose was to collect data for five seconds straight to build features. Additional research can help construct a better categorisation model by identifying the best time period for this research.
6. Conclusion
The Patient and physician data usage are ensured through this study. There are many diabetes management smartphone apps. However, we built the system based on physician feedback and tried to precisely gather all diabetic treatment data. Patient self-management tools and doctor remote patient monitoring were the goals of this study. Wearable technology was employed to capture patient activity data. The wearable IoT insole records activity data with no further user effort. App learns sitting, walking, and running using SVM and decision trees. Mobile health and self-management solutions provide remote monitoring and contact without regard to time, location, or expense. The herb also aids diabetes. Each day’s meal can be entered and reviewed by the patient. It enables patients to self-manage and track their development. These tools assist patients comprehend the importance of their choices. The long-term goal of this research is to test the application’s effectiveness on type 2 diabetic patients. Based on trial results and patient feedback, future designs can be improved. Depending on the results of a pilot study with real patients, this application may become a standard tool for diabetic care. Assuming its works, this model could potentially treat other chronic diseases.
Data Availability
The data underlying the results presented in the study are available within the manuscript.
Conflicts of Interest
The authors declare that they have no conflicts of interest regarding the publication of this paper.
Acknowledgments
We deeply acknowledge Taif University for supporting this study through Taif University Researchers Supporting Project Number (TURSP-2020/344), Taif University, Taif, Saudi Arabia. It was performed as a part of the Employment of Institutions.