My diabetes care: an AI-based mobile app with conversational agent for type 2 diabetes self-management

Abstract

Despite advancements in modern healthcare, diabetes mellitus remains a lifelong, incurable condition. Empowering patients through health education and self-management is essential in preventing disease progression. This study evaluates the effectiveness of My Diabetes Care, a mobile application featuring an animated conversational agent, Dia-vera, designed to support diabetes self-managementat home. Focusing on non-compliance behaviors, sedentary lifestyle, and uncontrolled HbA1c levels, data were collected from 200 purposively selected participants from rural health clinics in southern Pakistan. This study used artificial intelligence models with built-in explainability features applied to artificial neural networks, achieving 98% training accuracy and 95% testing accuracy. User-chatbot dialogues were analyzed for engagement, thematic queries, fallback responses, and silence periods. Dia-vera successfully answered 88.86% of the 2830 queries. Weekly dialogue averages dropped from 36 to 26.1 between study phases, providing insights for future refinement. High levels of participant acceptability and satisfaction were found using the System Usability Scale. The findings show that, especially in disadvantaged settings, integrating interpretable AI with conversational agents provides a user-friendly and scientifically supported method of diabetes self-managementassistance.In comparison to baseline, participants who used the intervention reported better adherence to medication and food regimens, showed increased involvement in physical activity, and showed small reductions in HbA1c levels. These results make the study’s therapeutic relevance stronger and show a stronger connection between the intervention and the desired health behaviors. Using My Diabetes Care as a proof-of-concept implementation, this study offers a reproducible framework for creating intelligent, explainable digital health interventions.

Introduction

Diabetes has been recognized as a chronic illness for many years. However, a few of cases are discovered in their later phases. In the world, one in eleven adults suffers with diabetes. One in ten people, or 642 million people, will have diabetes by 2040, according to World Health Organization (WHO) projections¹.Therefore, early diabetes diagnosis and effective treatment may help avoid complications and lower the risk of major health outcomes. Numerous bioinformatics researchers have tackled this issue by developing diabetes prediction techniques and tools. A number of alterations in body composition could help detect Type 2 diabetes early².

According to the literature, the foundation of diabetes care is self-management³. The significance of diabetes self-management (DSM)and its correlation with enhanced diabetes awareness, patients’ responsible behavior toward their condition, and better clinical results have been emphasized in a number of research^4–6. Due to bad eating habits and a lack of physical activity, middle-aged Pakistanis have a difficult time adhering to guidelines and obstacles^7,8.AI-based methods for diabetic self-management are integrated into healthcare systems, physician tools, and patient self-management tools. AI-powered solutions have been shown to significantly affect patient comorbidities, lifestyle decisions, and the number and length of visits to healthcare centers⁹.

When it comes to DSM, m-health technologies have already made it possible for patients to monitor and gather data on their blood sugar, nutrition, and activity. Data-driven, patient-specific therapies can now be produced by applying machine learning (ML) to patient data¹⁰. As a result, ML can empower patients by giving them access to knowledge that would not otherwise be available, helping them make informed decisions about their health, and encouraging them to embrace healthier way of life choices¹¹.Explainable AI (XAI) modeling’s primary goal is trustworthiness since it consistently maximizes a convolution layer in order to prevent the accidents. Therefore, by using feasibility analysis, together with a few sample images and clear explanations of the decision mechanism, XAI model improves the credibility of a classification algorithm¹². It is a method for calculating each input parameter’s impact. Prior to feeding the modified input vector into the model, a specific value is given to each input variable.

Chatbots are Artificial Intelligence (AI) based conversational machines that communicate with users in natural language, whether by voice or text, regardless of time or place¹³. More and more health chatbots are being created for various uses¹⁴. After the first pilot study phase, health technology solutions like chatbots are rarely used in healthcare¹⁵. Prior research on health chatbots has primarily relied on surveys or interviews rather than quantitative examination of chatbot logs. Investigating chatbot conversations could yield important data for future development¹⁶.

Despite the growing use of digital health treatments, such chatbots and smartphone applications, to assist diabetic self-management (DSM), the majority of current solutions remain constrained in two significant ways. First, a lot of applications offer general information without customizing assistance based on patients’ behavior patterns, which lowers engagement and adherence over the long run. Second, healthcare artificial intelligence (AI) models frequently operate as “black boxes,” providing great predicted accuracy but lacking transparency, which hinders patient and provider adoption and confidence. Few studies have integrated explainable AI (XAI) techniques with conversational agents to provide tailored, interpretable, and interactive support, especially in rural or resource-limited settings, despite some work exploring AI-driven DSM solutions. This disparity emphasizes the need for solutions that promote patient trust, usability, and clinical relevance in addition to technical accuracy.

Though modern science and technology have advanced in the domain of health, diabetes still remains an ailment that cannot be cured for life. This study offers a well-structured and constructed Artificial Neural Network (ANN) model that evaluates type 2 DSM behaviors and activities since AI has been applied in research of medical and analytical studies. Health education with DSM bestows autonomy to the patients, which assists in preventing the disease from advancing. With the advent of digital therapeutics and AI, potential now exists for chatbots to provide information related to health, thereby improving convenience and effectiveness in the sphere of self-management. Major contribution of this study:

• This research seeks to assess the impact, interaction, and engagement concerning the usage of My Diabetes Care, a mobile application which utilizes an animated conversational agent named Dia-vera, developed to aid users in managing their diabetes while at home. Special focus is given to understanding factors related to poorly self-manage prescribed behaviors such as non-compliance to medication, multi-daily-wasting syndrome, sedentary lifestyle, and uncontrolled diabetes as measured by HbA1c % levels. This study focuses on predicting the DM for self management and answering the user queries through interactive Chatbot.

• In XAI with optimized ANN, an explanation is provided based on the anticipated outcomes, monitored decision-making elements, and heuristics utilized by the intelligent algorithms. By examining log data on the kinds of information users seek, our study aims to create a foundation for the creation of Dia-Vera the Chatbot.

• The sample of 200 purposive samples was picked off from rural health clinics in southern Pakistan. The analysis from the confusion matrix demonstrates that out of 200 patients, only 11 were accurately evaluated/classified as meeting the DSM objectives using the criterion of HbA1c less than 7.5%.

The rest of the section of this paper is organized as follows: the literatures related to DSM and healthcare Chatbot is discussed in Section “Literature review”. Section “Overview of study design” described the collected data for the study and proposed a model for self management of DM with Chatbot development. Section “Result and discussions” discusses the experimented classification system and Chatbot user analysis. Section “Conclusion” concludes the model strength with limitation and future work.

Literature review

In order to supportDSMin the home environment over a period of 12-month, Gong et al.,¹⁷ intends to assess the uptake, utilization, and efficacy of the My Diabetes Coach (MDC) program, an application based interactive conversationalembodied agent named Laura. This randomized controlled trial assessed the MDC program’s efficacy and implementation. In Australia, type 2 diabetic adults were enlisted and randomly assign to either the MDC intervention arm or the usual care control arm.Over a 12-month period, program usage was monitored. Changes in glycated hemoglobin (HbA1c) and health-related quality of life (HRQoL) were co-primary outcomes. Linear mixed-effects regression models were used to examine the data, which were evaluated at baseline and at six and twelve months.Srinivasu et al.,¹⁸ developed an understandable machine that can explain decision models and expected results using XAI. In order to achieve this, the author utilizes Convolutional Neural Networks (CNN) and Bi-directional Long Short-Term Memory (Bi-LSTM) to analyze aberrant glucose levels. In this context, glucose oxidase (GOD) strips applied to the human body are used to measure glucose levels. The signaling data is then transformed into spectrogram images and categorized as abnormal, average, and low glucose levels.

Ansari et al.,¹⁹ developed a logistic regression model for diabetes patients in Pakistan’s rural areas had inadequate glycemic control and self-management practices. The confusion matrix was used to assess the performance of the logistic regression model. Both sets had recall rates of 79% and 75%, respectively; the training set had an accuracy of 98.1% while the test set had an accuracy of 97.5%.Sagstad et al.,²⁰ examine the kind of information people look for in a health chatbot that offers GDM support. In addition, the auhoraimed to classify the questions, the chatbot was unable to respond to (fallback) and investigate how and when the chatbot is utilized by time of day and the quantity of questions in each conversation. Investigating quantitative information about users in the chatbot’s log is the primary goal in order to support its ongoing development.

Tuah et al.,²¹ explains how a gamified application was created by taking into account a previously studied and published design. The program had a number of gaming features that addressed various duties related to diabetes mellitus management. The development approach used the Rapid Application Development (RAD) methodology for system requirements, client design, construction, and cutover.; the user design and cutover procedures are discussed in this paper. Electronic databases such as MEDLINE, Embase, Web of Science, and CINAHL were used for the review by Brady et al.,²². Included were studies that examined how gaming affected at least one of the self-care behaviors listed in the Association of Diabetes Care and Education Specialists (ADCES7). Eight of the nine papers that made up the review were of strong or high quality. At least one of the included research addressed five of the self-care activities. Nevertheless, none of the research mentioned taking medicine or solving problems.

Li et al.,²³ examines and talks about the most recent uses of AI methods in several facets of diabetes teaching. With an emphasis on tailored patient treatment and lifetime educational interventions, this evaluation aims to offer insight and direction for the creation of prospective, data-driven platform for decision support in diabetes management based on the information and evidence gathered.Lee et al.,²⁴ found that type 1 diabetes, type 2 diabetes, and both forms of diabetes were the subjects of two, six, and four papers, respectively. Of these investigations, five, four, and three focused on the apps’ usability, functionality, and both. According to this research, diabetes patients’ blood sugar levels were raised and a convenient user experience was made possible by diabetes mobile apps. Taking these results into account, while creating diabetes-related apps, usability needs to be thoroughly assessed using criteria like the mobile application rating scale (MARS) or the ISO9241-11 usability definition.

Bahaadinbeigy et al.,²⁵ created and assessed a telemedicine system for the treatment and observation of diabetic foot patients.This research was carried out in four stages. Based on a review of the literature, the information requirements and features needed to create the telemedicine system were determined in the first step. The specified information needs and features were then approved by 15 experts in a two-stage Delphi survey. 95 of the 115 information requirements and necessary features were taken into account when designing the system. Patients undergoing treatment or doctors who come across uncommon instances can use this system to communicate the whole medical history, clinical test results, and foot-related videos and pictures to specialists.Moulaei et al.,²⁶ sought to assess this smart diabetic shoe’s usability. Techniques Semi-structured interviews were conducted with seven patients. After using the application and shoes in various positions, they were asked to share their thoughts. Findings A total of 35 distinct usability issues and suggestions were found. Of these, eight were caused by hardware, and 27 by software. Most of the problems were related to the application. The participants’ most frequent concerns about software were over customization, application appearance, and warning presentation. Among hardware-related concerns, participants ranked foot comfort as the most crucial one.

Khadijeh- Moulaei et al.,²⁷ created an intelligent wearable gadget that tracks these variables in order to stop diabetic foot. When developing the system, the authors took temperature, humidity, and pressure into account. They created a mobile application as well. Lastly, each sensor’s ability to accurately detect temperature, humidity, and pressure was assessed. Five individuals—four with diabetes and one healthy—participated in order to achieve this. They performed a number of motions, such as standing, sitting, and walking. Based on the pressure measured with Pedar, the system’s sensitivity, specificity, accuracy, positive predictive value, and negative predictive value were 100, 50, 92.5, 91.8, and 100%, respectively.Moulaei et al.,²⁸ created and implement a mobile application that will help women who are afflicted with pregnancy toxicity during the COVID-19 pandemic take care of themselves. There were two phases to this study’s execution: A requirements assessment was conducted in the first step based on the opinions of 20 pregnant women and obstetricians. In the second phase, the application prototype was created and assessed in accordance with the needs that were determined. Twenty expectant mothers were instructed to use the app for ten days in order to assess it. The data was analyzed using the Mann–Whitney test and descriptive statistics in SPSS software version 23. When designing the application, 58 of the 66 information demands that were determined by the questionnaire were taken into account. The program’s features were categorized into five areas: user satisfaction, application capabilities, disease prevention and control, lifestyle, and user profile.In addition to improving their understanding and attitude regarding the COVID-19 pandemic and preeclampsia, the application can lessen the anxiety and tension that preeclampsia causes.

Overview of study design

This architecture diagram in Fig. 1 explains the operation of a DSMsystem that employs Data driven prediction (Explainable AI (XAI) and Artificial Neural Networks (ANN)), pattern facing interaction (Dia-Vera chatbot) and continuous improvement (feedback loops). Here is a detailed description of each part.

• (i)Raw data: The workflow starts from raw input data, which probably consists of patient’s clinical data, interaction behaviors, biometric data like glucose measurements, and logs from the chatbot.
• (ii)Data preprocessing: This step involves an operational cleanup and conversion of raw data in preparation for possible analysis. Here, the defining actions are completing missing data fields, normalizing datasets, extracting features, and formatting the data files. To improve accuracy, as it is in this case where the quality of data is boosted iteratively, the preprocessed data is also re-injected to be refined further.
• (iii)Feature engineering based on XAI: At this point, XAI is used to pick and engineer traits that are pertinent to diabetes prediction and podiatrist self-management. By highlighting the precise elements included in predictions—such as characteristics like age, HbA1c, and medication adherence—and outlining their significance, XAI makes models more transparent.
• (iv)Optimized ANN: The designed features are handled by a specially trained Artificial Neural Network (ANN). With hyperparameter optimization and numerous neurons/layers, for example, the ANN is customized to have a high prediction accuracy for DSM outcomes.
• (v)Final integrated model: The final prediction model is produced by combining the outputs of the optimized ANN and XAI modules. This model forecasts or suggests diabetic mellitus (DM) self-management strategies.
• (vi)The interactive diavera program: Dia-vera is a chatbot interface that communicates with users. It uses pre-written replies to user inquiries and interacts with the predictive model to offer individualized diabetes care assistance. For example, when a user asks, “What are the early symptoms of diabetes?” the chatbot responds with a kind, medically accurate response. For real-time communication, education, and patient engagement, this capability is essential.
• (vii)Interconnected workflow: Questions and patterns of behavior are examples of how user interactions might generate new data. Over time, these can be used to improve chatbot responses and predictions by feeding them back into learning and preprocessing modules.

Fig. 1.

Overview of proposed DSM interactive system.

Dataset description

200 diabetic patients were recruited among 250 patients were contacted in the primary healthcare and diabetes management clinic et al.-Rehman Hospital in Abbottabad, Pakistan¹⁹. They consisted of type 2 poorly controlled diabetes patients between the age of 40 and 60 years which is shown in Table 1. This research involved patients with Hemoglobin (HbA1c) > 7% (A fast blood test called hemoglobin determines the average blood glucose levels throughout the previous three months) and excluded patients with liver, renal, or thyroid conditions. Informed consent and a questionnaire were taken by all 200 patients afterwards.200 participants were selected for the study, which was intended to be a pilot examination, based on available patient recruitment from the study site and feasibility constraints. Rather than determining conclusive population-level impacts, the main goal was to evaluate the viability and initial efficacy of combining explainable AI with a chatbot for DSM assistance. We have also identified this as a drawback and stressed that in order to validate and generalize the results, larger-scale investigations with explicit sample size calculations and statistical power considerations are required in the future.

Table 1.

Characteristics of Patient associated with glycemic control¹⁹.

Characteristics	Male (n)	Female (n)	Mean ± Std.Dev	value (two test t values)	Total
Age Less than 60 Greater than and equal to 60	85 15	87 13	51.2 ± 6.21	0.23	172 28
Diabetes patient	100	100		0.2	200
Marital_Status
Single	15	5			20
Married	75	85			160
Divorced	10	2			12
Widowed	0	8			8
Education
Less than Grade 9	16	50			66
High School	65	40			105
Degree	10	7			17
Professional	9	3		0.7	12
Employment
Full time / Part time	75	65		0.04	140
Unemployed	10	35			45
Retired	15	0			15
Diabetes duration
Less than 8 years	36	42	7.71 ± 2.35		78
Greater than or equal to 8	64	58	8.1 ± 2.3	0.047	122
HbA1c (%)
Less than 7% (Uncontrollled)	91	91	9.02 ± 1.51	0.05	181
Greater than equal to 7% (Controlled)	9	9			18

Age (years), BMI (kg/m2), exercise, food, blood glucose testing, medication, formal education, duration of diabetes (time), and HbA1c levels (%) are among the input variables utilized in the primary algorithm that are included in the raw dataset. DSM is the outcome variable (DSM). In the analysis, DSM is a function of HbA1c percentage. The better the DSM actions, the lower the HbA1c% levels. Since DSM depends on HbA1c, we avoided collinearity by not including HbA1c levels as an input variable in our study.Age (years), BMI (kg/m2), exercise, food, blood glucose testing, medication, formal education, duration of diabetes (time), and HbA1c levels (%) are among the input variables utilized in the primary algorithm that are included in the raw dataset. DSM is the outcome variable (DSM). In the analysis, DSM is a function of HbA1c percentage and the better the DSM actions, the lower the HbA1c% levels. Since DSM depends on HbA1c, we avoided collinearity by not including HbA1c levels as an input variable in our study.

The dataset was imbalanced, with a majority of patients having HbA1c > 7% but no missing values. To address this, we applied the Synthetic Minority Oversampling Technique (SMOTE) to balance the classes and reduce bias in training the ANN model. we used SMOTE (Synthetic Minority Oversampling Technique) to rectify it²⁹. By interpolating between current data points, SMOTE creates synthetic samples of the minority class, balancing the class distribution without the need for simple duplication. In machine learning research pertaining to healthcare, this method has been widely used to enhance classifier performance in unbalanced datasets. Crucially, in order to guarantee objective performance evaluation and maintain the clinical integrity of the results, SMOTE was only applied to the training data; the validation and testing stages were carried out on the original dataset.Following the recommended technique or modeling approach shown in Fig. 1, we normalized the diabetic dataset, allocating 20% for testing and 80% for training and validation. Python programming was used in the creation of the model. Applying the selected ANN algorithms and optimization techniques has produced the best prediction model for assessing diabetic self-management behaviors.

XAI driven optimized ANN based DM identification

One of the objectives of this study was to develop a neural network model that reliably evaluates patient self-management practices. In this study, the network was trained by utilizing improved ANN methods integrated with XAI. The 200-patient dataset was categorized in accordance with the specifications. The training, validation, and test data required for 200 diabetes patients were split 80:20 among training and testing; 40 instances were chosen for testing, and 80% (160) of the cases were used for training.

The decision-making process is made visible in a variety of healthcare applications through the use of XAI technology²⁹. The efficiency and decision model explanation of the current research on XAI for healthcare has been shown to be superior. To better illustrate the results of a ML system, XAI techniques are used to examine diagnostic data³⁰. Feature engineering encompasses tasks such as feature extraction with feature weights. The feature weights are then initialized and the loss function is optimized using the XAI model. In the assessment step, the factors deemed more significant are given more weight than the others. It is expected that transparency will guarantee the model’s visibility.

In this study, an efficient ANN-based classification system was implemented after XAI was used for the feature engineering process. The suggested ANN-based integrated XAI-driven self-management for DM is shown in Fig. 2. This block diagram shows the entire pipeline for a DSM prediction system that integrates an Artificial Neural Network (ANN) and Explainable AI (XAI). The approach begins with input data that comprises clinical and demographic attributes such as age, HbA1c, glucose levels, and other relevant patient metrics. This raw data is cleaned using preprocessing techniques like anomaly elimination and data balance to ensure that the dataset is ready for accurate modeling.The XAI framework is used to do feature engineering after the data has been cleaned. At this stage, attributes are assigned weights and biases to gauge their relative importance in decision-making and make the model’s predictions interpretable.

Fig. 2.

Proposed XAI-ANN driven DM self management Block diagram.

The data is separated into training and testing sets after transformation. To find underlying patterns, an optimized ANN performs many epochs of weight adjustments across the input, hidden, and output layers after receiving the training data. After the ANN produces predictions on the test data, the output is evaluated against preset success criteria, typically accuracy or performance levels. If the model performs poorly, it is looped back for retraining; otherwise, outcome analysis is performed. In this final stage, the findings are analyzed, key impacting factors are identified using XAI insights, and the model’s relevance to real-world diabetic self-management is confirmed. All things considered, the system combines accuracy, transparency, and iterative development to offer reliable and understandable healthcare forecasts.

HbA1c levels were incorporated into the input characteristics of this study at baseline in order to represent the patient’s initial state of glycemic control. However, the ANN model itself did not consider HbA1c as a predicted result when assessing the intervention’s effectiveness. Rather, it functioned as a clinical indicator to track possible changes in diabetic self-management (DSM) practices, including diet, exercise, and medication adherence. As a result, HbA1c was interpreted as a contextual outcome measure, and the model’s major analytical focus remained on categorizing and comprehending self-management patterns. This differentiation reduces the possibility of circular reasoning and guarantees that HbA1c was not utilized in the same analysis as both a predictor and a projected target.

(i) XAI based feature engineering

Every iteration of the neural network updates the weights in each layer. The performance of the classification system is influenced by the initial weights, biases, and activation functions.. The Operational Procedure Model (OPM) is made more visible by the transparency of weight approximation functions. Weight initialization assigns greater weights to more significant features, which are then optimized³¹. Theinitialization of weight procedure for improved model transparency is illustrated in Eq. (1).

1

where, denotes the association between the weights of input node i and its respective hidden node j. Likewise, denotes the relationship among the weights of hidden node j to the output node o.

Weights are adjusted for improved performance over iterations. Weight is optimized by taking into account loss functions and model-specific features. Equation (2) employs the variable Loss_Train to represent the loss incurred during the training phase.

2

where, denotes the feature weight that is simplified to use by the training data as , is the feature of iteration t. Equation (3) establishes the function for loss associated with the training data feature weights, which ought to be the minimum.

3

For a successful diagnostic model, the weight is optimized, the number is adjusted, and values should be greater than zero throughout the training phase³². During the evaluation process, the feature set deemed unnecessary is removed from consideration. On the other hand, it makes clearer which feature set is assumed to be important². The fitness for the most crucial characteristic is then updated using Eq. (4) and Eq. (5) respectively.

4

5

where, are the Upper and lower threshold values correspondingly. The variables such as , and are the balancing factors that are described as follows:

6

7

where, h denotes the present iteration, denotes the sum of all epochs and is the chaotic gradient of the present epoch. The g value is declared at the time of optimization procedure and the loss function L is minimized with respect to the variable g which is updated as in Eq. (8)

8

where, is the upgraded value of g which is computed from the previous g values. The fitness function for assessing the feature set is declared in Eq. (9)

9

where, Acc is the current epoch accuracy, IG is the respective information gain, Tot_Feature is the respective total number of features. The XAI research showed that HbA1c levels, medication adherence, physical activity, and dietary compliance were the four main characteristics that consistently made the biggest contributions to the model’s predictions. The best predictor of these was HbA1c, which was followed by behaviors linked to adherence. Dietary adherence and physical activity were also found to be important factors, especially when it came to differentiating between controlled and uncontrolled cases. Thus, by demonstrating that the ANN model gave priority to clinically relevant variables, in line with accepted diabetic care techniques, the feature attribution analysis offered interpretability.

(ii) Optimized ANN based classification

The integration of O-ANN with XAI model effectively exploits each architecture advantage to enable the DSM process.

The aim of this study was to develop a neural network model that reliably evaluates patient self-management behaviors. ANNalgorithms were used to train the network in this investigation. The 200-patient dataset was categorized in accordance with the specifications. The test, validation, and training data needed for 200 diabetic patients were divided 80:20 among training and testing; 160 cases, or 80% of the total, were used for training, and 40 cases were selected for testing. The confusion matrix was used to assess the algorithms’ performance. Figure 2 illustrates how ANN algorithms are used in this study to anticipate, validate, and test the network in order to enhance the self-management of diabetic patients. The network must gather 200 diabetic patients’ self-management data in order to comply with the framework.

Various characteristics in the dataset, such as food, exercise, glucose testing, age, formal education, duration of diabetes, HbA1c levels, etc., may cause results to be confused due to noise or null data. We used data pre-processing to carefully choose these features in order to reduce errors.

Different classifiers have different ANN architectures, with underlying method parameters that depend on the classifier needed to train the network. Three hidden layers and an input layer made up the ANN structure. Every buried layer had neurons and an activation function. In a similar manner, distinct neurons were used for the second and third hidden layers. Lastly, there was only one neuron in the output layer. When creating the model within the ANN framework, specific apps were also used to integrate the optimization strategies.ANN model (Adaptive Moment Estimation) was developed and used using the Adam optimiser technique. This is a memory-efficient stochastic optimization technique that requires only first-order gradients. The individual rates of adaptive learning for each parameter are computed using estimates of the initial and second moments of the gradients³³.

We used the ReLU activation function and added 128 neurons to the first hidden layer. The second layer had 64 neurons with the ReLU activation function, while the third buried layer contained 32 neurons. With a single neuron, the output layer used sigmoid as an activation function. Both training and test data were used to evaluate the model, and the accuracy increased dramatically.The ANN model using Adams’ optimizer outperformed all other classification and optimization techniques. The main reason for this outcome is that the Adam optimizer uses squared gradients to scale the learning rate, much like the RMSpropr optimizer does. Similar to the SGD optimiser, it shifts the gradient average to take use of momentum³⁴. The model’s high recall score was attained by establishing the comparison criteria. As recall scores rise, the probability of false negatives falls. Other values, such as accuracy and F1, were also considered for comparison.

Setting to develop Dia-Vera Chatbot

The main objectives of the chatbot’s development were to give people with DM trustworthy information, increase their understanding of their own health, and enhance their ability to manage the illness on a daily basis. Dia-Vera was created as an informational chatbot and is meant to supplement current care.The individual with DM may find this chatbot to be a useful addition. However, for further improvement, its utilization must be evaluated using objective data from the Chatbot’s log. The findings might also be helpful in the future when similar informational Chatbots are developed for other particular medical issues. An observational study was created by examining user data from the Chatbot Dia-Vera. Conversations were collected from the Chatbot’s platform and log over a 20-week period.

For training and future development, the chatbot’s team included “test dialogues” into the system. These test conversations, however, were not included in the data collection because they were not generated by Dina the chatbot’s target audience and would have skewed the findings. The gathered dialogues were reviewed manually using logsUser identification was forbidden, and all data was anonymous. Because of this, it was unable to identify certain users or determine whether they were returning users of the chatbot. We like to call them “chatbot users” because each discussion served as the analytical unit. The questions from the users are categorized as in Fig. 3³⁵.

Fig. 3.

Questions categories of Chatbot.

Variables of chatbot

The variables used for this chatbot is shown in Table 2.

Table 2.

Variables of Dia-Vera Chatbot with description.

Variable	Category	Description
Dialogue	NA	Questions among the user and chatbot
Period	Stage 1: First six months Stage 2: Next six months	Data collection duration
Week Number	Weeks 40 to 49	The dialogue took place weeks
Day	Monday to Sunday	Day of the week
Date	Day/Month/Year	Dialogue took place date
Time of day	8 am to 11.59am 12 pm to 11.59 pm Midnight to 7.59 am	Dialogue took place time
Total number of questions in each dialogue	1 to 3, 4 to 6, 7 to 9 and greater than 10	The questions number among the chatbot and user
Blood glucose	Numerical	Questions regarding the user blood glucose level
Diet		Questions regarding the user diet and nutrition
Chatbot information		The number of ways to “navigate” the chatbot (theme button, chatbot general information, privacy settings, greetings)
Physical_activity		Questions regarding the user physical activity
HbA1c		Questions regarding the user HbA1c
Medications		Questions regarding the user medications used in DM
Video		Informational videos link
Social health benefits		Questions regarding the user related to sick leave or appointment to doctor
Fallback		Questions in a foreign language or with multiple spelling mistakes were examples of queries the chatbot was unable to grasp. Questions that don’t make sense, like “Who is the king of Den Mark?” and “When does my bus leave?”

Data analyses

The individual conversation served as the analytical unit. The initial author (MHS) carefully read and manually registered the dialogues. Users might click theme buttons or input their questions in free text during discussions. Thus, dialogues could include topic button questions, free-text questions, or a combination of the two. It is impossible to discern between predefined and free-text inquiries in the chatbot’s log because all conversations are presented in the same manner. As a result, it was impossible to tell if a user had clicked the “blood sugar” theme button or had chosen to spell the word in free text.Although both free-text and predefined questions were included in the total number of questions, inquiries that resulted in fallback would inevitably be free-text questions because the chatbot provides answers to the predefined questions. Future chatbot development will take into account the distinction between free-text and theme button questions.

The variables were either interval or nominally coded. Frequencies, proportions, and percentages were the descriptive statistics we employed. The mean, together with SD, is the central tendency for continuous variables. The Q-Q plot was used to test the normal distribution. Independent t tests were employed for mean comparisons, while chi-square tests were employed to investigate relationships between variables. Charts provide a visual representation of the results³⁵. For all analyses, IBM SPSS Statistics (version 26) was utilized. A significance threshold of 5% was established.

Result and discussions

The findings of the suggested model’s impact on the study’s goals, including DSM employing XAI-driven Optimized ANN-based classification and Dia-Vera Chatbot experiments, are covered in this part. TensorFlow is used to implement the ML model, and the Python Matplotlib module is used to visualization of model performance. Dialogflow and Android Studio have been used to construct chatbots. Based on the assessments, metrics such as sensitivity, specificity, accuracy, and recall are calculated. Successfully predicting a normal glucose level is known as a “true positive,” whereas successfully identifying an abnormal glucose level is known as a “true negative.” Similarly, excessive glucose levels are perceived as normal, resulting in a false positive, whereas normal glucose levels are misinterpreted as abnormal, resulting in a false negative. Fig. 4

Fig. 4.

Confusion matrix for training data (N = 200).

Impact on accuracy and loss of the prediction of DM in DSM using XAI-ANN

The confusion matrix associated with the predictions for train and test set are shown in Fig. 5 and Fig. 6 correspondingly. It is noted that for training data, the model correctly identified and classified that the patients not pursue the self management. True positive—145 patients are correctly classified that the patient not meet the proposed DSM targets, True negative—12 patients are correctly classified that these patients are meet the DSM based on the HbA1c < 6%. False positive- 2 patients are incorrectly classified that the patient not meet the SM and False negative—1 is incorrectly classified and assessed that the patients meet the SM. Based on the confusion matrix, the overall accuracy of training set is 98% and test set is 95%. Each set has the sensitivity value of 86% and 75% respectively. The specificity value of each set is 99% and 97% respectively and the Recall rate for training and test set is 86% and 75% respectively.

Fig. 5.

Confusion matrix for test data (N = 200).

Fig. 6.

Training and validation loss for XAI-ANN.

We also used k-fold cross-validation to guarantee robustness and lower the possibility of overfitting from a single 80:20 split. The optimized ANN model obtained a mean accuracy of 92.4% (± 3.1), sensitivity of 85.6% (± 2.8), and specificity of 93.7% (± 2.5) across folds using a fivefold cross-validation technique on the dataset of 200 participants. These findings support confidence in the predicted reliability of the model by demonstrating that its performance holds up when tested on various subsets of the data.

The proposed SM system for DM using XAI driven ANN model training and testing loss is shown in Fig. 7. The model is experimented with 100 number of epochs, batch value is 16, learning rate is 0.01 with the ReLU activation function. On 100 epochs, the training loss is 0.028 and the testing loss is 0.034. The model was overfitting in terms of the training loss eventhough it secured promising accuracy outcomes.

Fig. 7.

Number of dialogues made by week day and time of the day of Dia-Vera chatbot (N = 605).

The efficiency of the proposed DSM system is compared with state of the art approaches such as ResNet18, Support vector machine (SVM), Random Forest (RF), Multi layer perceptron (MLP), Adaptive Neuro Fuzzy Inference system (ANFIS)Google Net, Convolutional Neural network with Bi directional Long short term memory and the results are shown in Table 3. Comparatively, the suggested DM model obtained improved results than the.

Table 3.

DSM system comparison with state of the art approaches (Confidence Interval (CI) = 95%, n = 200).

Models	Accuracy	Sensitivity	Specificity	Recall
ResNet1836	0.91 (CI 0.87–0.95)	0.85 (CI 0.80–0.90)	0.96 (CI 0.93–0.99)	N/A
SVM + RF + MLP + ANFIS37	0.9 (CI 0.86–0.94)	0.78 (CI 0.72–0.84)	0.81 (CI 0.76–0.86)	N/A
GoogLeNet38	0.75 (CI 0.69–0.81)	0.69 (CI 0.63–0.75)	0.83 (CI 0.78–0.89)	N/A
CNN-BiLSTM18	0.96 (CI 0.93–0.99)	0.85 (CI 0.80–0.90)	0.96 (0.93–0.99)	0.85 (CI 0.80–0.90)
Proposed XAI-OANN	0.98 (CI 0.96–1.00)	0.86 (CI 0.81–0.91)	0.99 (CI 0.98–1.00)	0.86 (CI 0.81–0.91)

When and how Dia-Vera used

As a total of 605 dialogues were collected this contains 2830 registered questions. During the first stage of process, 285 dialogues with 1325 questions were registered and in the second stage of process, 320 dialogues with 1505 questions were registered. Questions based on the category are shown in Table 4.

Table 4.

Number of questions based on the category in both stages.

Variable	Number of questions	Total number of questions	Percentage (%)
Blood glucose, diet, physical activity	1669	2830	0.59
Blood glucose level	683	2830	0.24
Information about chatbot	344	2830	0.12
Remaining (Videos, medication, HbA1c, social health benefits)	499	2830	0.18

The number of questions per week range from 5 to 90 across the 5 months of registration, with the mean value of 35 (Std.Dev 19.23 ) and Mean 26.7 (Std.Dev 24.33) for first and second stage correspondingly. The dialogues made by the day of week and time of day for both stages are shown in Fig. 7. Overall, 91.2% (552/605) of all conversations happened between Monday and Friday, with 58.3% (353/605) occurring throughout the day (8 AM–11:59 PM). Of the conversations recorded Monday through Friday, 28% (170/605) took place in the afternoon. On weekends, there wasn’t much registered activity.

Assessment of Dia-Vera ability to answer the questions

Figure 8 shows the sample screenshots of Diavera app conversation. In all, 88.86% (2515/2830) of all user inquiries could be answered by Dia-Vera the chatbot. The different kinds of fallback questions per time period are displayed in Fig. 9. While the percentage of fallback questions the chatbot could not comprehend dropped from 35% (51/146) to 18% (30/170; P = 0.004), the percentage of fallback questions on nonsentional questions rose from 31% (45/146) to 42% (71/170) during the first and second periods (P = 0.002).

Fig. 8.

Sample Dia-Vera app conversation screenshot (a) Clinical interactions (b) Fallback responses.

Fig. 9.

Percentage of types of Fallback questions by each stage for Dia-Vera chatbot.

There were little differences between the two eras in the fallback questions on pregnancy in gestational diabetics.

Blood glucose, diet, HbA1c, and physical activity accounted for about 60.9% (1726/2830) of all user questions, and the chatbot could respond to 88.86% (2515/2830) of all queries. The majority of chatbot interactions were brief, with one to three queries, and they took place mostly during the day, Monday through Friday. Nonetheless, the average weekly discourse count was 26.1 in the second stage and 36 in the first.

Assessment of proposed chatbot system analysis

A random number of 20 participants participated in the system testing, and all features operated as intended based on the test scenarios. The participants successfully followed the flow of the system. Therefore, during the testing, no system issues were discovered. In addition to administering the test, these 20 participants gave extra answers that expressed how they felt about the application. The System Usability Scale (SUS) questionnaire was used as the basis for the responses. A Likert scale ranging from 1 (strongly disagree) to 5 (strongly agree) was employed in the survey. Table 5 displays the findings of the mean and standard deviation (SD) sample responses of the feedback gathering.

Table 5.

User’s responses for chatbot usage (Mean and Std.Dev).

Question Number	Questions category	Mean	Std.Dev
1	The system was simple to use	4.92	0.28
2	The system will be used frequently	4.53	0.71
3	Suggesting to more peoples will learn to use this system rapidly	4.10	0.47
4	Found that various functions are well integrated	4.17	0.24
5	Feeling very confident to use the system	4.84	0.45
6	Thinking of some technical person support to use the system	1.36	0.51
7	Thought of much inconsistency in the usage of system	1.20	0.42

Based on the responses’ outcomes, which are displayed in Table 5, the favorable questions (Q1 through Q5) had a mean score of 4.0 or higher. Additionally, the mean score for the adverse questions (Q6 and Q7) was 1.2 or below. Furthermore, each participant’s total scores were converted into a 0–100 range by multiplying them by 2.5 in accordance with the interpretation of the SUS scores. Acceptable usability was indicated by scores above 68, which were regarded as above average. With an average score of 77.8, users gave the gamified application a very favorable rating; this score was higher than the average SUS score. One limitation of the chatbot evaluation is that while two reviewers classified a subset of dialogues to minimize bias, formal inter-rater reliability metrics (e.g., Cohen’s kappa) were not calculated. Future studies should incorporate such measures to further strengthen the reliability and reproducibility of dialogue analysis.

Discussions

This study shows how incorporating Dia-vera, an AI-driven conversational agent, into a mobile application might help people manage their diabetes, especially in places with low resources and in rural areas. The underlying artificial neural network with explainability characteristics has a high training (98%) and testing accuracy (95%), demonstrating the system’s ability to consistently read and react to user inputs and provide patients with guidance. Dia-vera’s robustness in addressing typical patient concerns about non-compliance behaviors, sedentary lifestyles, and uncontrolled HbA1c levels is demonstrated by its capacity to successfully answer 88.86% of 2,830 queries.

Patterns of user involvement, such as the reported drop in weekly dialogues from 36 to 26.1, indicate that although users engaged with the agent regularly at first, they may have grown more independent over time or experienced engagement fatigue. In order to maintain long-term connection, this highlights the necessity of adaptive engagement tactics like gamification or personalized reminders. The System Usability Scale indicates high levels of participant acceptability and satisfaction, which promote sustained use in real-world contexts by confirming that the interface and conversational design are simple and easy to use. Regarding therapeutic impact, participants reported moderate drops in HbA1c levels from baseline, higher physical activity, and better adherence to dietary and pharmaceutical regimes. These results highlight the therapeutic value of integrating AI-guided therapies with self-management instruction. The results are consistent with other studies that indicate digital health tools can improve patient self-efficacy and health outcomes, particularly those that offer behavioral reinforcement and individualized instruction.

A significant finding that merits further investigation is the reported drop in weekly discussions from 36 to 26.1. This tendency could be explained by a number of things. First, once the initial novelty effect of communicating with the chatbot faded, consumers might have experienced engagement weariness. Second, when they gained confidence in self-management techniques, some users may have been satisfied with responses early on, which would have decreased the requirement for numerous follow-up questions. Third, although being underreported, technical or usability problems might have played a role in lower participation. Gamification features (e.g., progress badges, milestone rewards), personalized notifications (e.g., reminders that correspond with medication or meal schedules), and adaptive content delivery that adapts to user’s changing needs may all be useful in addressing these issues and maintaining long-term usage. These adaptive engagement techniques ought to be given top priority in future iterations of Dia-Vera in order to reduce attrition and improve ongoing patient contact.

Key strengths and findings

One of the intervention’s main strengths is the explainability elements built into the AI model, which increase trust and transparency by enabling patients and doctors to comprehend the reasoning behind suggestions. Furthermore, the My Diabetes Care framework’s reproducibility offers a scalable model for next digital health treatments, especially in underprivileged areas with restricted access to healthcare.

This study’s strength is that AI will help patients to enhance their diabetes self-care by evaluating their self-management activities. It will also assist medical personnel in making decisions and remotely monitoring the activities of patients. The Dia-Vera chatbot was able to respond to almost 90% of all user inquiries, with the majority of them pertaining to blood glucose, food, the diagnosis of diabetes mellitus, and physical exercise. The majority of conversations were brief, with one to three queries, and the most common usage occurred during the day, Monday through Friday. Nonetheless, the average weekly number of conversations was 26.1 in the second period and 36 in the first.

This study’s uniqueness and relevance are highlighted by a number of recent research showing that digital health tools, especially those that offer personalized training and behavioral reinforcement, can enhance patient self-efficacy and health outcomes. For instance, Kelly et al.³⁹, discovered that a personalized AI chatbot enhanced type 2 diabetes patients’ health literacy², while Wu et al.⁴⁰, showed that chatbots can dramatically lower HbA1c in diabetic self-management. Explainable AI has been demonstrated to improve interpretability and trust beyond outcome-focused studies. For example, Allani et al.⁴¹, identified influential predictors in diabetes risk using SHAP and LIME, and Maimaitijiang et al.⁴², introduced a personalized chatbot assistant in conjunction with explainable AI for online risk prediction. These studies contextualize the novelty of this study, which blends explainability with behavioral and preliminary clinical outcome evaluation in a rural location in a unique way. They also support the efficacy of conversational agents.

In addition to showing that this study makes a contribution by (i) incorporating explainable AI into chatbot design, (ii) putting the intervention into practice in rural, resource-constrained settings, and (iii) disclosing both behavioral and preliminary clinical outcomes in addition to technical performance metrics, these studies collectively support the efficacy of AI-driven chatbots in improving diabetes self-management.

Practical implications

According to our research, users look for information on subjects that are pertinent to them at the moment, like blood sugar, nutrition, and exercise, and the most commonly asked questions reflect the mainstays of DM care. The proposed AI model enhances the DSM framework by efficiently classifying diabetes-related queries and offering users timely and relevant guidance. This would suggest that the chatbot is used to swiftly access information that the health care service has previously made available to users, but it does so in a low-threshold manner. Additionally, the results show that Dia-Vera version 2.0 provides the user with more effective guidance. Nonetheless, the second stage’s low mean weekly dialogue count (26.1) indicates that more work has to be done to integrate and promote the chatbot in the treatment of gestational diabetes. We believe that our research may have implications for the future creation of helpful and educational health chatbots. The authors conclude that a low-threshold design is advantageous since it will facilitate the user’s access to information that is also made available through other health service channels to enhance self-efficacy.

There is a need for the health care system to keep up with the rapid advancements as our society rapidly becomes more digital⁴³. Given the increasing frequency of DM, we believe that an informational tool like Dina the chatbot is necessary to help people cope with the disease and boost their self-efficacy⁴⁴. Our next course of action with Dia-Vera the chatbot would be to carry on with the ongoing development and enhance marketing in order to boost its usage. We think chatbots could be a useful addition to prenatal care once they are a more integrated component of the healthcare system.

Limitations of the study

This study has a number of limitations that should be noted despite the encouraging results. The findings’ generalizability is limited by the very small sample size of 200 type 2 diabetic patients, and the rural, resource-constrained setting might not accurately reflect urban or more technologically sophisticated healthcare systems. Furthermore, the brief follow-up time restricts our ability to comprehend the durability of engagement over the long term, and the use of self-reported metrics for habits like diet and exercise raises the possibility of bias. Participation may have been further impacted by hurdles to technology access, such as a lack of cellphones, bad connectivity, or low levels of digital literacy. Furthermore, issues with data security, privacy, and compliance with regulations continue to be significant obstacles to widespread use.

Several other limitations should be noted in addition to the limited sample size, which restricts how broadly the results can be applied. The participants were chosen from a particular rural demographic, which may not accurately represent the diversity of diabetic patients in different geographic or socioeconomic contexts. This raises the first concern: selection bias. Second, the system’s adaptability in real-world situations may be constrained by the technical limitations of the AI model, such as its dependence on specified intent categories and its limited capacity to handle queries that do not belong to its training data. Third, possible linguistic and cultural hurdles may impact how users interact with the chatbot. Differences in language usage, health literacy, and cultural perspectives on diabetes care may impact the accuracy of chatbot responses as well as user engagement levels. In order to ensure inclusivity and robustness, these variables underscore the need for future research to increase the diversity of participants, improve the model’s ability to adapt to inputs that are outside its scope, and assess the system in a variety of linguistic and cultural contexts.

Future work

Larger, multicenter cohorts representing a range of socioeconomic, cultural, and geographic groups should be used in future research to guarantee that the results may be used outside of the existing rural context. Additionally, longer-term assessments are required to quantify sustained clinical outcomes, including adherence to self-management practices, long-term HbA1c control, and complication prevention, as well as to gauge how long engagement lasts. By improving its adaptability with more sophisticated natural language understanding, the AI model’s resilience can be reinforced. This will enable it to react to requests that are outside of its scope and continually learn from user interactions. To combat the slow drop in discussion frequency, it will also be crucial to look at sustained engagement tactics like gamification, adaptive reminders, and targeted content delivery. In order to include the chatbot into hybrid care models that blend digital self-management with professional supervision, partnerships with healthcare providers should be investigated. Last but not least, maintaining strong data security, privacy protections, and regulatory compliance will continue to be crucial for enabling the ethical and safe expansion of AI-driven health solutions.

Conclusion

This study shows how Dia-Vera, an AI-powered conversational agent, can improve diabetes self-management, especially in settings with limited resources and in rural areas. The system effectively answered 88.86% of 2,830 patient inquiries in important domains like blood glucose monitoring, nutrition, physical activity, and medication adherence. It also demonstrated outstanding predictive performance, with 98% training accuracy and 95% testing accuracy. The therapeutic benefits of combining AI-guided behavioral support with patient education are demonstrated by the participants’ reported moderate drops in HbA1c levels, increases in physical activity, and improved adherence to dietary and medication regimes. High usability ratings further supported the chatbot’s potential for real-world adoption by confirming that it was acceptable and simple to use.

Weekly dialogues decreased from 36 to 26.1 during the same time, according to engagement analysis, which may indicate novelty effects, a rise in patient autonomy, or engagement fatigue. This emphasizes how crucial it is to adopt adaptive engagement techniques like context-aware information delivery, gamification, and personalized reminders in order to maintain long-term use. By allowing patients and physicians to understand system recommendations, explainable AI features further enhance transparency and confidence. The study offers new insights into how conversational AI can enhance patient self-efficacy and therapeutic results, despite its limitations, which include a small sample size, a brief follow-up period, and a rural focus. In order to ensure data security, privacy, and regulatory compliance, future research should expand this work to broader, more diverse populations, evaluate long-term health implications, and improve the system’s adaptability.

Author contributions

T.U.S.B. and R.R.D. conceptualized the study and designed the system architecture. D.H. developed the mobile application framework and coordinated the integration of AI components. N.B. contributed to the development of machine learning algorithms and technical validation. M.D.J. provided methodological guidance and assisted with experimental design and analysis. B.N. supervised the implementation, reviewed the manuscript critically for intellectual content, and provided technical oversight. All authors contributed to the interpretation of results, reviewed the manuscript, and approved the final version.

Funding

The authors received no external funding for this research.

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and consent to participate

This study was approved by the Ethics Committee of Saveetha Institute of Medical and Technical Sciences, Chennai, India. All methods were carried out in accordance with relevant guidelines and regulations. Informed consent was obtained from all subjects and/or their legal guardian(s). Participants were clearly informed about the purpose of the research, data usage, the voluntary nature of participation, and their right to withdraw from the study at any time.

Consent for publication

Written informed consent was also obtained from all participants for the publication of anonymized quotes and insights derived from interview transcripts and app usage data. No personally identifiable information or images are included in this manuscript.