Research proposal writers

“A Deep Learning Model for self-management, adherence and Education among Type 1 Diabetes Patients”

CHAPTER ONE

Introduction

This chapter presents the background of the study, the problem statement, purpose, objectives of the study, research questions, study scope, justification of the study, significance.

Background

For 2,000 years’ diabetes has been recognized as a devastating and deadly disease (Mahaffey, 2019), In the first century A.D. a Greek physician, Aretaeus, described the destructive nature of the affliction, which he named “diabetes” from the Greek word for siphon (Mahaffey, 2019).

Physicians in ancient times, like Aretaeus, recognized the symptoms of diabetes but were powerless to treat it effectively (Petersmann et al., 2019). In the 17thcentury a London physician, Dr. Thomas Willis, determined whether his patients had diabetes or not by sampling their urine. If it had a sweet taste he would diagnose them with diabetes mellitus-honeyed diabetes. This method of monitoring blood sugars went largely unchanged until the 20th century (Kato et al., 2019).Before the discovery of the insulin little could be done for patients suffering from diabetes. Low calorie diets prolonged their lives but left them weak and near starvation (Yuan et al., 2019), But in 1921, doctors in Canada treated patients dying of diabetes with insulin and managed to drop high blood sugars to normal levels. Since then, medical breakthroughs have continued to prolong and ease the life of people with diabetes (Lowe et al., 2019). In the ’50s, it was discovered that there were two types of diabetes: “insulin sensitive” (type I) and insulin insensitive (type II). Two thousand years have passed since Aretaeus spoke of diabetes as the mysterious sickness. It has been a long and arduous process of discovery, as generations of physicians and scientists have added their collective knowledge to finding a cure (Sattar et al., 2019). It was from this wealth of knowledge that the discovery of insulin emerged in a small laboratory in Canada. Since then, medical innovations have continued to make life easier for people with diabetes (Pivari et al. 2019).

Global diabetes cases continue to soar from 529 million to 1.3 billion by 2050 (Ong et al., 2023). Prevalence has been rising more rapidly in low- and middle-income countries especially in Asia and Africa than in high-income countries. In Africa, there are currently 24 million adult diabetics; by 2045, that number is expected to increase by 129% to 55 million (WHO analysis, 2024). The incidence of type 1 diabetes varies from 1.5 to 10.1 per 100 000 depending on the nation and age group analyzed, according to estimates from various sub-Saharan African countries (Katte et al., 2023).  According to estimates from the International Diabetes Federation (IDF), around 25,000 African children and adolescents under the age of twenty-one had diabetes in 2019.(Katte et al., 2023). According to the International Diabetes Foundation (IDF), 716,000 persons in Uganda were predicted to have diabetes by 2021; of those, 89% were not aware of their condition or on medication (WHO, 2023).

There has been a significant evolution in system modeling and intelligence since the introduction of early deep learning models. Deep learning methods rapidly emerged and expanded their applications across various scientific and engineering domains (Micikevicius et al., 2022). Health informatics, energy, urban informatics, safety, security, hydrological systems modeling, economics, bioinformatics, and computational mechanics were among the first areas to leverage deep learning techniques. State-of-the-art surveys on data-driven methods and machine learning algorithms indicate that deep learning, along with ensemble and hybrid machine learning methods, represents the future of data science. Further comparative studies show that deep learning models and hybrid machine learning models often outperform conventional machine learning models (Sharma, & Guleria, 2022).

Deep learning methods are rapidly evolving to achieve higher performance. The literature includes numerous review papers on the advancing algorithms in specific application domains such as renewable energy forecasting, cardiovascular image analysis, super-resolution imaging, radiology, 3D sensed data classification, multimedia analytics, sentiment classification, text detection, transportation systems, activity recognition in radar, hyperspectral analysis, medical ultrasound analysis, image cytometry, and Apache Spark (Singh et al., 2022).

Diabetes is a chronic disease that affects how somes’ body processes blood sugar (glucose), which is the primary source of energy for thecells (Foretz, Guigas, &Viollet, 2019),  Insulin is a hormone produced by the pancreas that regulates blood sugar levels, and in people with diabetes, the body either doesn’t produce enough insulin or doesn’t use it effectively (Glovaci et al., 2019), resulting in high blood sugar levels, there are several types of diabetes, including; Type 1 diabetes: This type of diabetes occurs when the immune system attacks and destroys the cells in the pancreas that produce insulin. Type 1 diabetes is usually diagnosed in children and young adults, and people with this type of diabetes need to take insulin injections or use an insulin pump to manage their blood sugar levels; Type 2 diabetes: This type of diabetes occurs when the body becomes resistant to insulin or doesn’t produce enough insulin to maintain normal blood sugar levels. Type 2 diabetes is the most common type of diabetes and is usually diagnosed in adults, although it is increasingly being diagnosed in children and adolescents; Gestational diabetes: This type of diabetes occurs during pregnancy and usually goes away after the baby is born (Zelniker et al., 2019), However, women who have had gestational diabetes are at increased risk of developing type 2 diabetes later in life.Other types of diabetes: There are several other rare types of diabetes, including monogenic diabetes and cystic fibrosis-related diabetes (Kenny, & Abel, 2019).Diabetes can cause a variety of health problems if left untreated, including heart disease, kidney disease, nerve damage, and blindness. However, with proper management, people with diabetes can live long, healthy lives. Treatment usually involves a combination of medication, lifestyle changes (such as a healthy diet and regular exercise), and monitoring blood sugar levels (Cappon et al., 2019).

In Uganda, the annual growth rate of Type 1 Diabetes (T1D) is 10.6%, while Type 2 Diabetes grows at a rate of 3.7%. As a result, T1D patients often present to the health system with challenging-to-treat comorbidities(Type 1 Diabetes Index, 2024). For young people with diabetes, living successfully with Type 1Diabetes (T1D) is particularly hard due to many potential psychosocial and contextual barriers to self-management and adherence (Sawyer et al., 2022).

Scholars state that individuals with type 1 diabetes (T1D) must maintain their insulin levels through insulin pump therapy, multiple daily insulin injections, or the use of an automated insulin delivery system. T1D is a chronic metabolic disorder marked by elevated blood glucose levels, which can cause serious damage to the kidneys, heart, blood vessels, eye , nerves, and kidneys (Lucier and Weinstock, 2023). This condition requires patients to perform critical self- management tasks multiple times per day through the Two key self-management tasks in Type 1 Diabetes are frequent monitoring of blood glucose (BG) and administering insulin (Campbell et al., 2018). These tasks help manage glycemic control to avoid or delay serious short- and long-term consequences, such as retinopathy, neuropathy, and mortality. Mealtimes are a critical time for diabetes self-management (Mulvaney et al., 2019). The management of glucose levels and education is necessary to prevent TID related complications (Campbell et al., 2018). Achieving these glucose targets and awareness is imperative to reduce an increased risk of hypoglycaemia due to the limitations of subcutaneous insulin delivery; hence, diabetes self-management is a critical element of care. Despite the rapid development of science and technology in healthcare, diabetes remains an incurable lifelong illness. Diabetes education aiming to improve self-management skills is an essential way to help patients enhance their metabolic control and quality of life.

Artificial intelligence is the science of making machines that can think like humans (Li et al., 2020, 2022). This AI technology enable such as Medical Image Analysis, Virtual Assistants, Predictive Analytics, Chatbots, Automated Administrative Tasks and Wearable Devices and Sensors have made significant progress in transforming available genetic data and clinical information into valuable knowledge (Li et al., 2020, 2022) . AI also helps to provide insight and guidance for the development of prospective, data-driven decision support platforms for diabetes management, with a focus on individualized patient management (Li et al., 2020).  To identify problems related to self-management, patients, caregivers, and clinicians rely on blood glucose and insulin administration data from devices along with a patient recall of behavioural, emotional, and/or contextual events that could pose barriers to self-management (Zhang et al., 2022) . However, utilizing retrospective memory or recall for events that are days or weeks in the past has been identified as generally unreliable and potentially biased. Unreliable recall of events in diabetes problem-solving could result in modifications to the insulin regimen that are not based on reliable information (Sawyer et al., 2022).

Ecological momentary assessment (EMA) techniques have been created and effectively applied in a variety of health disorders such as diabetes, heart illnesses, and hypertension in order to solve the limits of human recall and bias in health behaviour research (Zhang et al., 2022). In contrast to traditional assessment methods, EMA utilizes more frequent and in-vivo ambulatory assessment of factors that impact health behaviors and decision-making (Zhang et al., 2022). EMA techniques offer a more direct and frequently more precise technology-mediated way to track and evaluate the situations, personal experiences, and decision-making processes that surround health decisions in day-to-day living (Zhang et al., 2022). Specifically, EMA techniques that yield richer data and more frequent and relevant observations per individual have been developed for therapeutic benefits in diabetes management. These techniques help identify novel correlates for interventions like continuous glucose monitoring (CGM) and improve the accuracy of health behaviour correlation assessments. (Friedman et al., 2023). Smartphone apps to log daily events and calculate bolus insulin are increasingly being adopted to successfully reduce the daily burden associated with T1D self-management (Friedman et al., 2023). Other wearables, such as wristbands, have been used in recent literature to estimate physical activity for T1D subjects (Zhu et al., 2022). However, there are currently no integrated platforms synchronized with wearable technologies and apps for decision making and clinical efficacy(Zhu et al., 2022). The use of these technologies continue to yield a substantial amount of granular data and have boosted machine learning-based algorithms.

Despite their exceptional performance across a growing range of problems, deep learning models that excel on average often make systematic errors, struggling with semantically coherent subsets of data. A landmark example is the Gender Shades study, which revealed that vision models for gender recognition have disproportionately high error rates for images of black women. AI systems have also demonstrated poor performance for marginalized groups in areas such as object recognition, speech recognition, mortality prediction, and recruiting tools (Alanazi et al., 2022), Other systematic errors can be harder to anticipate. For instance, medical imaging classifiers may be sensitive to changes in imaging hardware; essay scoring software might give high scores to long, poorly written essays; and visual question-answering systems can fail when questions are rephrased. Recognizing and mitigating these errors is crucial to avoid designing systems that exhibit discriminatory or systematically unreliable behavior (Wang et al., 2022).

These challenges have led the community to develop better tools for testing model performance, establish clearer standards for reporting model biases, and create various methods for training more equitable or robust models. Even when it is difficult or impossible to make repairs, identifying and flagging edge cases where systems fail can help expert users work around an algorithm’s flaws (Saba et al., 2022).

However, these methods require practitioners to recognize and label well-defined groups in their datasets ahead of time, which inevitably overlooks semantically related sets of inputs not identified in advance. While it is essential for practitioners to explicitly assess model performance on sensitive subpopulations, it is extremely challenging to anticipate all types of inputs on which models might systematically fail, such as specific age groups, poses, backgrounds, or lighting conditions (Gupta, & Agrawal, 2022).

Furthermore, Previous literature explored several classic machine-learning approaches for the prediction of glucose levels or glycemic events using prediction horizons between 15- and 60-min (Zhu et al., 2022). However, the performance of physiological and rule-based prediction models is still limited by the influence of various external factors and high inter and intra-subject variability on glucose dynamics (Zhu et al., 2022) . It is well-accepted that self-management is essential to reduce the risks of chronic complications in diabetes patients (Li et al., 2020, 2022). Diabetes education is a system-wide and individualized way to ensure the effectiveness of patient self-management (Li et al., 2020, 2022). Considering the variety of individual needs, goals, and life experiences, patients should be educated with fine-tuned knowledge and skills that fit their situation. Mobile health (mhealth) and Electronic health (eHealth) tools have been embraced in chronic disease management (Mugabirwe et al., 2021). In fact, health technology with different models, frameworks and e-health tools such as such as PositiveLinks Interventions (Mugabirwe et al., 2021) and short message service (SMS) texting(Asiimwe et al., 2011; Siedner et al., 2012) have been embraced to help in the self-management of non-communicable diseases. Moreover, by 2017, more than 318,000 mobile health applications were available to consumers worldwide with Diabetes apps accounted for 16% of the total number of disease-specific apps available to consumers (Alkawaldeh, Choi and Jacelon, 2020). These Diabetes apps vary in the functions they provide, including tracking blood glucose measurements, nutrition database and carbohydrate tracking, physical activity and weight tracking, sharing data with clinicians or peers, social support, messaging, and reminders (Alkawaldeh, Choi and Jacelon, 2020). However, most of these technologies address type 2 diabetes, and there fewer literatures in low resource countries like Uganda especially on TID self-management and educative messages using health technologies.

AI techniques have been used for a number of diabetes-related activities, including blood glucose classification and screening. Deep learning is particularly effective for applications like image analysis and natural language processing because it uses hierarchical layers to interpret input data.

(Tahir and Farhan, 2023). This research therefore proposes a deep learning model that is able to synchronize the real-time data from TID patients’ data to improve decision support towards enabling timely self-management and educative messages of TID.

Diabetes mellitus, particularly Type 2 Diabetes Mellitus (T2DM), is a significant and escalating non-communicable disease in Uganda and across sub-Saharan Africa, contributing to a substantial disease burden among older adults (Priyadarshani et al., 2023). T2DM patients face a doubled risk of coronary heart disease and have poor prognoses, with many cases going undiagnosed and untreated. This silent killer necessitates active prevention and management strategies to curb its rising prevalence (World Health Organization, 2023).

In Uganda, the situation is further compounded for Type 1 Diabetes (T1D) patients, who face severe challenges in managing their condition (Bulemela et al., 2023). Access to insulin and necessary medical supplies has deteriorated, resulting in decreased insulin injection frequency, blood glucose monitoring, and meal frequency, which complicates adherence and effective management. The lifelong commitment required for insulin therapy involves meticulous monitoring and frequent administration of insulin, which is mentally and physically demanding (Ashique et al., 2024). These challenges lead to heightened stress, anxiety, and a decline in quality of life, especially in home settings with limited medical support.

Current diabetes interventions, such as continuous glucose monitoring and digital health applications, are often inaccessible to patients in low-resource settings due to their complexity and cost. These gaps in accessibility and the lack of tailored self-management support exacerbate the difficulties faced by individuals managing diabetes (Surlari et al., 2023). Furthermore, disparities in healthcare resources in low- and middle-income nations intensify these challenges, making it imperative to develop solutions that empower patients and their caregivers with enhanced tools for remote management (Ashique et al., 2024).

Therefore, this study aims to design a Deep Learning Model for self-management of Type 1 Diabetes. The model will enhance remote management, adherence, and education of individuals dealing with T1D through a robust system that integrates wearable-generated data. By addressing the specific needs and challenges of T1D patients in Uganda and similar low-resource settings, the model seeks to improve glycemic control, reduce the risk of complications, and ultimately enhance the quality of life for these individuals.

Problem statement

Diabetes is one of the top non-communicable diseases in Africa, contributing to the increasing disease burden among the old adults. In Uganda, diabetes mellitus is a silent major non communicable disease and the T2DM is a major and growing problem in the country (Gatimu et al., 2016). Individuals with T2DM have a two times higher danger of acquiring coronary illness than whatever remains of the populace, and their prognosis is very poor (Tuomilehto et al., 1997). T2DM today, is found in nearly all population and epidemiological suggestion suggests that without active prevention and management agendas, the prevalence will remain to upsurge worldwide (Unwin et al., 2009). In Gatimu et al.2010, work it was stated that in sub-Saharan Africa, 21.5 million individuals are living with diabetes prompting roughly a large portion of a million diabetes-related passing’s in 2013 (IDF, 2013). Diabetes mellitus is one of the still most important non-communicable disease and the T2DM is a key rising problem in the country. T2DM is a silent killer as several of the cases go undiagnosed (de-Graft Aikins, 2007).

In Uganda, Type 1 Diabetes patient’s access to insulin syringes has significantly worsened, and Insulin injection frequency, blood glucose monitoring and meal frequency significantly decreased causing adherence and management challenges (Sseguya et al., 2023). Controlled and self-managed Type 1 Diabetes to reduce the risks of chronic complications in diabetes patients through self-management, adherence and education in an individualized way to ensure the effectiveness of patient is very key (Maina, Pienaar and Reid, 2023). Considering the variety of individual needs, goals, and life experiences, patients should be educated with fine-tuned knowledge and skills that fit their situation.

Patients with type 1 diabetes must follow a number of treatment guidelines, including food, insulin dosage schedules, and blood glucose monitoring, among others, since these all improve the course of the condition (Kyokunzire and Matovu, 2018). However, the lifelong commitment of Type 1 diabetes patients to insulin therapy is very complex, because individuals must meticulously monitor their blood glucose levels and administer insulin multiple times a day to maintain glycemic control (Janež et al., 2020; Zhu et al., 2022). This routine, though essential, can be mentally and physically demanding, and lead to heightened stress and anxiety levels, affecting their quality of life. This becomes even more burdensome especially in home settings where access to immediate medical support may be limited (Janež et al., 2020; Zhu et al., 2022). These challenges contribute to the persistent escalation of diabetes-related complications and mortality, as well as the lack knowledge, and encounter obstacles related to timely decision-making support.

Current diabetes interventions, including continuous glucose monitoring, digital health apps, and AI-driven recommendations, face  gaps in accessibility to patients in low resource setting, because many need resources and are complex in nature, and barely sustainable (Kumbara et al., 2023), as they do not address self-management, adherence and inadequate knowledge regarding management of TID.  Furthermore, disparities in healthcare resources, especially in low- and middle-income nations, intensify these challenges. Limited access to advanced monitoring technologies and specialized medical care exacerbates the difficulties faced by individuals managing diabetes, making it imperative to develop solutions that empower them and their caregivers with enhanced tools for remote management. This study, therefore, aims to design a Deep Learning Model for self-management for Type 1 Diabetes to enhance remote management, adherence and education of individuals dealing with type 1 diabetes through a robust model that integrates wearable-generated data. This technology will enable patients self-manage, adhere to their medication and therapy as well as have access to educative information ultimately improving the overall outcomes and quality of life for those affected by type 1 diabetes.

Main Objective

The study aims at contributing towards improving Type 1 Diabetes (T1D) self-management by designing a deep learning model embedded into an application for self-management, adherence and education among TID patients.

Specific Objectives

1. To analyze deep learning models in order to find out their strength and weaknesses during self-management of type 1 diabetes and inform the design of deep learning model.
2. To design and develop a deep learning model incorporated into an app for self-management, adherence and education among Type 1 Diabetes patients.
3. To assess the preliminary impact of the model on self-management, adherence, and knowledge among Type 1 Diabetes patients using the UTAUT model, it is important to delve deeper into the concepts of adherence and knowledge in the context of diabetes management.

Research Question

1. Type 1 diabetes (T1D) is a chronic condition characterized by the autoimmune destruction of insulin-producing beta cells in the pancreas, leading to lifelong dependency on exogenous insulin. Managing T1D is complex and requires continuous monitoring of blood glucose levels, carbohydrate intake, and physical activity, along with regular insulin administration. This intricate balancing act poses significant challenges for patients, often leading to both acute and chronic complications if not managed properly. Consequently, effective management strategies are crucial to maintaining glycemic control and preventing complications, therefore it is worth examining What are the strengths and weaknesses of deep learning models in management of type 1 diabetes?

1. Self-management is paramount in the treatment of Type 1 Diabetes, as it empowers patients to take control of their condition through informed decision-making and proactive health behaviors. Successful self-management involves regular blood glucose monitoring, insulin dosage adjustments, dietary planning, and physical activity regulation. Education on these aspects is essential, enabling patients to understand how different factors influence their blood glucose levels and how to respond to fluctuations. Adherence to these self-management practices can significantly reduce the risk of complications and improve overall quality of life for individuals with Type 1 Diabetes, therefore How can a deep learning model incorporated into an app for self-management, adherence and education among Type 1 Diabetes patients be designed and developed?

The advent of technology, particularly deep learning models, has opened new avenues for enhancing T1D management. Deep learning models excel at analyzing large datasets and identifying patterns, making them suitable for predicting blood glucose levels, optimizing insulin dosing, and providing personalized recommendations. These models can process continuous glucose monitoring (CGM) data, dietary intake, and physical activity information to offer real-time insights and predictive analytics, thus aiding in more precise and timely decision-making, what is the preliminary impact of the model on self-management, adherence and knowledge among Type 1 Diabetes patients?

Significance of the study

Diabetes has been a recognized medical challenge for over 2,000 years, with its destructive nature first described by the Greek physician Aretaeus in the first century A.D. Despite significant advancements in diabetes treatment and management, it remains a leading cause of morbidity and mortality globally, particularly in low- and middle-income countries (Mahaffey, 2019; Petersmann et al., 2019). The discovery of insulin in 1921 marked a major breakthrough, transforming diabetes management and enabling many patients to live longer and healthier lives (Yuan et al., 2019; Lowe et al., 2019). However, the burden of diabetes continues to escalate, with global cases projected to rise from 529 million to 1.3 billion by 2050, disproportionately affecting regions such as Africa and Asia (Ong et al., 2023; WHO, 2024).

In Uganda, the diabetes epidemic is particularly concerning, with a significant increase in both Type 1 and Type 2 Diabetes cases. The annual growth rate of Type 1 Diabetes (T1D) in Uganda is 10.6%, leading to severe complications and comorbidities due to inadequate self-management and healthcare infrastructure (Type 1 Diabetes Index, 2024). The challenges faced by T1D patients in Uganda are exacerbated by limited access to insulin, essential medical supplies, and effective diabetes education, resulting in poor disease management and a decline in quality of life (Ashique et al., 2024).

Recent advancements in artificial intelligence (AI) and deep learning offer promising solutions to improve diabetes management. Deep learning techniques, particularly those involving real-time data integration from wearable devices, have shown potential in enhancing decision support systems, enabling timely self-management, and providing personalized educative messages for diabetes patients (Li et al., 2020; Zhang et al., 2022). These technologies can overcome the limitations of traditional diabetes management approaches, which often rely on unreliable patient recall and retrospective data analysis.

This study aims to develop a deep learning model specifically designed for the self-management of Type 1 Diabetes in Uganda. By leveraging real-time data from wearable devices, this model seeks to enhance glycemic control, adherence, and patient education. The proposed model addresses the unique challenges faced by T1D patients in low-resource settings, offering a cost-effective and accessible solution to improve diabetes management and reduce the risk of complications.

The significance of this study lies in its potential to transform diabetes care in Uganda and similar low-resource settings. By providing a robust, data-driven decision support system, the deep learning model can empower patients with the tools and knowledge needed for effective self-management. This, in turn, can lead to improved health outcomes, reduced healthcare costs, and a better quality of life for individuals living with Type 1 Diabetes. Furthermore, this research contributes to the broader field of health informatics by demonstrating the practical applications of AI and deep learning in chronic disease management, paving the way for future innovations in healthcare technology.

CHAPTER TWO

Literature review

2.0 Introduction

This section presents the literature of the study in line with the study objectives;

2.1 To analyze deep learning models in order to find out their strength and weaknesses during self-management of type 1 diabetes and inform the design of deep learning model.

Diabetes mellitus (DM) is a chronic, metabolic, clinically heterogeneous disorder in which prevalence has been increasing steadily worldwide (Chen et al., 2012). It is estimated that 366 million people had DM in 2011; by 2030, this will have risen to 552 million (Cho et al., 2018). DM is characterized by persistent hyperglycemia, which may be due to impaired insulin secretion, resistance to insulin’s peripheral actions, or both, which eventually leads to pancreatic beta-cell failure (Padhy et al., 2019). People living with DM are more vulnerable to various forms of both short- and long-term complications due to metabolic aberrations that can cause damage to various organ systems, leading to the development of disabling and life-threatening health complications, the most prominent of which are microvascular (retinopathy, nephropathy, and neuropathy) and macrovascular complications (Lotfy et al., 2015).

Deep learning belongs to the broad family of machine learning methods, Unlike traditional neural networks-based classifiers, deep learning builds classifiers with many hidden layers, aiming to identify the salient low-level features of an image (Goodfellow et al., 2016). In the context of deep learning, transfer learning is a technique that exploits the usage of features that were learned by a network over a given problem to solve a different challenge in the same domain. Transfer learning has many advantages. First, it saves computational time because, instead of training a new model from scratch, it uses the already available information from the last training process. Second, it extends the knowledge it acquired from previous models, and third, transfer learning is beneficial when the size of the new training dataset is small. Transfer learning promises valuable contributions to the fields of computer vision, audio classification, and natural language processing.

There have been many attempts to automatize the image classification task to facilitate the process or make it more accurate. One of the earliest attempts was the convolutional neural network (CNN), which LeCun et al. (1989) introduced for the image classification task. In 2012, thanks to Krizhevsky et al. (2012), CNN became the most popular technique for addressing the image classification problem. The authors achieved state-of-the-art performance in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) competition (Russakovsky et al., 2015), outperforming other commonly used machine learning techniques. CNN can be used in image classification and natural language processing (Collobert & Weston, 2008; Dos Santos & Gatti de Bayser, 2014; Kalchbrenner et al., 2014) and time series analysis (Wang et al., 2017; Zhao et al., 2017). In all of these cases, training the deep network weights from scratch requires a substantial amount of time and massive datasets (hundreds of thousands of images). These requirements make.

Deep learning algorithms very challenging in the context of medical images where, typically, only a limited number of images are available. A lot of time and experience are required to annotate medical images. That is where transfer learning can play a significant role: It allows for the use of a pre-trained architecture that was previously fitted to images of the same domain. Thus, transfer learning is particularly suitable for addressing the DR classification domain, where there is a lack of images to accurately train a CNN from scratch. Several studies have been done to classify DR by using CNN, either by using transfer learning or by introducing novel architectures (Nørgaard & Grauslund, 2018), but to the best of our knowledge, there have not been any reviews that survey the existing transfer learning techniques to classify DR images. To answer this call, in this paper, we discuss state-of-the-art DR image classification models that use the transfer learning of deep CNNs. Moreover, we discuss some essential open questions to apply transfer learning in the DR domain better.

Type 1 diabetes (T1D) is a chronic condition requiring continuous self-management, including monitoring blood glucose levels, administering insulin, and maintaining a balanced diet. The advent of deep learning (DL) has brought new possibilities for improving the self-management of T1D by providing personalized predictions, recommendations, and automated insulin delivery systems. This review analyzes various deep learning models used in T1D management, focusing on their strengths, weaknesses, and implications for future model design.

Deep learning models, particularly convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short-term memory (LSTM) networks, have been employed in managing T1D. These models can predict blood glucose levels, detect patterns, and optimize insulin dosing. The integration of continuous glucose monitoring (CGM) data with DL models has shown promising results in improving glycemic control.

DL models, especially LSTMs, are known for their ability to capture temporal dependencies in glucose level data, leading to high predictive accuracy. This is crucial for preventing hypo- or hyperglycemic events. DL models can be trained on individual patient data, allowing for personalized diabetes management. This personalization improves the relevance and effectiveness of recommendations provided by the model.

Some DL models are integrated with insulin pumps to create closed-loop systems, automating insulin delivery based on real-time glucose readings. This reduces the burden on patients and improves adherence to treatment protocols. Once trained, DL models can be scaled across large populations, making them a valuable tool for public health interventions in diabetes management (Rane, Choudhary, & Rane, 2024).

DL models require large amounts of high-quality data for training, which may not be readily available for all patients. This limits the model’s applicability in diverse populations. One of the major challenges of DL models is their “black-box” nature, making it difficult for healthcare professionals to understand how predictions are made. This lack of transparency can hinder trust and adoption among clinicians and patients (Baduge et al., 2022).

Training DL models is computationally intensive, requiring significant processing power and time. This can be a barrier to real-time applications in diabetes management. DL models are prone to overfitting, especially when trained on small datasets. Overfitting can lead to poor generalization, where the model performs well on training data but fails to predict accurately on new data.

The use of patient data in DL models raises concerns about privacy, data security, and the potential for bias in model predictions. These ethical considerations must be addressed to ensure equitable and responsible use of DL technologies in T1D management. The strengths and weaknesses of current DL models provide valuable insights for future model design. To enhance the effectiveness of DL in T1D management, the following strategies are recommended: Incorporating synthetic data generation and data augmentation techniques can address the challenge of limited datasets, improving model robustness and generalizability (Asif et al., 2024).

Developing interpretable DL models, such as using attention mechanisms or hybrid models combining DL with traditional machine learning, can increase trust and adoption among users. Streamlining the computational requirements of DL models through techniques like model pruning and quantization can make them more feasible for real-time applications. Implementing robust ethical frameworks that prioritize patient privacy, data security, and fairness in model predictions is essential for responsible AI in healthcare (Pichler, & Hartig, 2023).

Deep learning models hold significant potential for improving the self-management of type 1 diabetes by providing accurate predictions, personalized recommendations, and automated insulin delivery. However, challenges related to data requirements, interpretability, computational complexity, and ethical considerations must be addressed. By focusing on these areas, future DL models can be better designed to meet the needs of patients and healthcare providers, ultimately enhancing the management of T1D.

2.2 OBJECTIVE TWO; Designing and developing a deep learning model for an app aimed at Type 1 Diabetes (T1D) management involves several key steps:

Designing a deep learning model involves several key steps, each critical to building an effective and efficient model.

The first step in designing a deep learning model involves defining the Problem

Designing deep learning models involves several intricate challenges, and the identification of problems at the initial stages is crucial for developing effective models. A well-conducted literature review highlights key issues and provides insight into addressing these challenges.  One of the most critical problems in designing deep learning models is related to data quality and availability. Deep learning models require large amounts of high-quality data to perform effectively. However, data can be noisy, incomplete, or imbalanced, which can significantly affect the model’s performance. Incomplete or missing data can lead to biased models that do not generalize well to real-world scenarios. Data preprocessing techniques such as data augmentation, imputation, and normalization are essential steps, but they add complexity to the model design process. The availability of labeled data is another issue, especially in fields where obtaining labeled data is expensive or time-consuming. Unlabeled data can lead to the exploration of unsupervised or semi-supervised learning, but these approaches are less mature and pose their own set of challenges (LeCun, Bengio, & Hinton, 2015).

The complexity of deep learning models can lead to problems such as overfitting, where the model performs well on training data but fails to generalize to new, unseen data. Overfitting is a significant challenge in deep learning, particularly when the model is too complex relative to the amount of training data available. Techniques such as regularization, dropout, and cross-validation are often employed to mitigate overfitting, but they require careful tuning and can introduce additional computational overhead (Srivastava et al., 2014). Additionally, finding the right balance between model complexity and performance remains a significant challenge, as overly complex models can also lead to increased training times and difficulty in deployment.

Underfitting occurs when a model is too simple to capture the underlying structure of the data, leading to poor performance on both training and test data. This problem often arises from insufficient model capacity, inadequate training data, or overly simplified feature representations. The identification of underfitting requires careful monitoring of model performance metrics during the training process. Techniques such as increasing the model’s complexity, improving feature engineering, and augmenting the training data are commonly used to address underfitting. However, these solutions must be implemented cautiously to avoid the risk of overfitting (Goodfellow, Bengio, & Courville, 2016).

Deep learning models, especially deep neural networks, are often criticized for being “black boxes” due to their lack of interpretability. Identifying and addressing the interpretability problem is crucial, particularly in domains like healthcare, finance, and autonomous driving, where understanding the model’s decision-making process is essential. While techniques like LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) have been developed to improve model interpretability, they often require additional computational resources and do not always provide complete transparency (Ribeiro, Singh, & Guestrin, 2016). Designing interpretable models from the outset is an ongoing research challenge that involves trade-offs between model accuracy and transparency (Shoaib et al., 2023).

Deep learning models are computationally intensive, requiring significant resources for training and inference. The problem of identifying appropriate computational resources is critical, as insufficient resources can lead to prolonged training times, reduced model accuracy, and challenges in scaling the model for real-world applications. The advent of specialized hardware like GPUs and TPUs has mitigated some of these issues, but the cost and accessibility of such resources remain barriers for many researchers and organizations. Additionally, scalable deep learning models require careful consideration of distributed computing techniques and efficient parallelization, which add complexity to the model design process (Dean et al., 2012).

Another emerging problem in the design of deep learning models is the identification and mitigation of biases that may arise from the data or the model itself. Bias in deep learning models can lead to unfair or discriminatory outcomes, particularly in sensitive applications like criminal justice, hiring, and lending. Ethical concerns must be addressed early in the model design process, through techniques such as fairness-aware learning and bias auditing. However, these approaches are still evolving, and their integration into the deep learning pipeline requires careful consideration of ethical principles and regulatory guidelines (Barocas, Hardt, & Narayanan, 2019).

The identification of problems in designing deep learning models is a complex and multifaceted challenge that requires careful consideration of data quality, model complexity, interpretability, computational resources, and ethical concerns. By addressing these issues early in the design process, researchers and practitioners can develop more robust, reliable, and ethically sound deep learning models. Ongoing research and advancements in the field will continue to refine the tools and techniques available for identifying and mitigating these problems, ultimately leading to more effective and trustworthy deep learning applications (Shoaib et al., 2023).

Deep learning (DL) has revolutionized various fields by enabling the development of models that can learn from vast amounts of data and perform complex tasks with remarkable accuracy. However, the effectiveness of these models largely depends on the clarity with which the problem they are designed to solve is defined (Soori et al., 2023).  A well-defined problem sets the foundation for the entire model development process, influencing everything from data collection and preprocessing to model selection and evaluation (Ahmed et al., 2023).

A clearly defined problem provides a roadmap for the entire deep learning model design process. According to Bengio et al. (2015), the problem definition stage is crucial as it helps in identifying the specific objectives, constraints, and expected outcomes of the model. This clarity ensures that the model is tailored to address the exact needs of the application, leading to more efficient and targeted solutions.

The choice of data is directly influenced by how well the problem is understood. Data quality, relevance, and volume are critical factors that determine the success of a deep learning model (Goodfellow et al., 2016). A well-defined problem helps in selecting appropriate datasets that are representative of the real-world scenario the model is expected to operate in. Additionally, it guides the preprocessing steps, ensuring that the data is cleaned, normalized, and augmented in a manner that aligns with the model’s objectives.

The design of the neural network architecture and the tuning of hyperparameters are heavily dependent on the problem definition. Different problems may require different model architectures such as convolutional neural networks (CNNs) for image processing or recurrent neural networks (RNNs) for sequential data (LeCun, Bengio, & Hinton, 2015). A clear understanding of the problem allows for the selection of the most appropriate architecture and aids in the efficient tuning of hyperparameters, which are crucial for optimizing model performance.

A poorly defined problem can lead to overfitting, where the model performs well on training data but fails to generalize to new, unseen data. This is often due to the model being trained on data that is not representative of the broader application domain (Zhang et al., 2016). Clear problem definition helps in identifying the appropriate boundaries of the problem space, ensuring that the model learns patterns that are generalizable and not just specific to the training dataset.

When the problem is not clearly defined, it becomes challenging to establish the correct evaluation metrics. This can lead to situations where the model is optimized for the wrong criteria, resulting in suboptimal performance in the actual application. For instance, in a classification problem, the choice between accuracy, precision, recall, or F1-score as the evaluation metric should be driven by a thorough understanding of the problem (Sokolova & Lapalme, 2009). A well-defined problem ensures that the evaluation metrics align with the model’s intended use.

Deep learning models, especially complex ones, are often viewed as “black boxes” due to their lack of interpretability (Lipton, 2016). When the problem is not clearly defined, it exacerbates the difficulty in interpreting the model’s decisions, as there is no clear reference point to assess whether the model is making the right predictions for the right reasons. A clear problem definition can help in designing models with better interpretability, by aligning the model’s internal representations with the underlying structure of the problem.

In healthcare, the stakes are particularly high, and the importance of problem definition cannot be overstated. For instance, Esteva et al. (2017) demonstrated the importance of a well-defined problem in developing a deep learning model for skin cancer classification. By clearly defining the problem identifying specific skin conditions to diagnose and determining the relevant image data the researchers were able to develop a model that achieved dermatologist-level accuracy. This case highlights how a clear problem definition can lead to impactful and reliable models in critical domains.

In autonomous driving, the need for precise problem definition is crucial for safety and reliability. Chen et al. (2015) discussed how the complexity of the driving environment requires a clear problem definition to ensure that the model can handle various driving scenarios, such as lane detection, obstacle avoidance, and traffic sign recognition. By clearly defining each sub-problem and integrating them into a cohesive model, the researchers were able to enhance the overall performance and safety of autonomous vehicles.

The importance of clearly defining the problem in designing deep learning models cannot be overstated. It serves as the foundation for the entire model development process, guiding data selection, model design, and evaluation, without a clear problem definition, deep learning models risk being inefficient, difficult to interpret, and prone to overfitting. As deep learning continues to be applied in increasingly complex and high-stakes domains, the need for precise problem definition becomes even more critical

In the context of developing deep learning models for healthcare, particularly for chronic diseases like Type 1 Diabetes (T1D), the processes of data collection and preprocessing are critical. These stages are foundational to the success of the model as they directly influence the accuracy, reliability, and generalizability of the predictions made by the model. This review explores the significance of data collection and preprocessing in designing a deep learning model for self-management, adherence, and education among T1D patients.

In T1D management, various data types are essential, including blood glucose levels, insulin dosage, carbohydrate intake, physical activity, and patient-reported outcomes (e.g., mood, stress levels). Data can be gathered through continuous glucose monitors (CGMs), insulin pumps, food diaries, wearable devices, and mobile health applications. Each data type contributes uniquely to the model, providing a comprehensive view of the patient’s health status.

Data for deep learning models can be sourced from clinical trials, electronic health records (EHRs), patient self-reports, and real-time monitoring devices. Real-time data from CGMs and insulin pumps are particularly valuable as they provide continuous, high-frequency inputs that are critical for accurate prediction and recommendation. However, the diversity in data sources necessitates a robust integration process to ensure that the model can effectively utilize the information.

Collecting high-quality data poses several challenges, including ensuring patient compliance with data logging, dealing with missing or incomplete data, and managing the variability in data due to individual differences in T1D management. Additionally, privacy and ethical concerns must be addressed, particularly when handling sensitive health information. Before data can be fed into a deep learning model, it must be cleaned to remove inaccuracies, inconsistencies, and outliers. This involves dealing with missing data, which can be handled through techniques like imputation or by removing affected data points. Ensuring that the data is accurate and reliable is crucial for model performance (Rubin-Falcone, Lee et al. 2023).

Normalization and scaling are critical steps in preprocessing, especially when dealing with heterogeneous data sources. For instance, blood glucose levels and insulin doses are on different scales, and normalization ensures that no single variable disproportionately influences the model’s predictions. Common techniques include min-max scaling and Z-score normalization.

Data augmentation techniques can be employed to artificially increase the size of the dataset, which is particularly useful when dealing with small datasets. This can include generating synthetic data or applying transformations to existing data, such as shifting or scaling time series data. Augmentation helps in reducing overfitting and improving the generalizability of the model. Feature engineering involves selecting, modifying, or creating new features from raw data to improve model performance. In T1D management, this might involve creating features that capture trends in glucose levels over time or interactions between insulin doses and carbohydrate intake. Effective feature engineering can significantly enhance the model’s predictive power (Wang, Kurth-Nelson et al. 2016).

Properly splitting the data into training, validation, and test sets is essential for evaluating the model’s performance. Care must be taken to ensure that the splits are representative of the entire dataset, and techniques like cross-validation can be employed to maximize the use of available data. The quality and quantity of data, as well as the robustness of the preprocessing steps, have a profound impact on the performance of deep learning models. Poorly collected or preprocessed data can lead to inaccurate predictions, reduced model reliability, and increased risk of overfitting. Conversely, well-collected and preprocessed data contribute to more accurate, reliable, and generalizable models that can effectively support T1D self-management, adherence, and education (Joseph, & Vadasseril, 2022). Data collection and preprocessing are indispensable components of designing deep learning models for T1D management. By carefully addressing the challenges in these areas, researchers can develop models that provide valuable insights and support for patients managing T1D. Future work should focus on developing standardized protocols for data collection and preprocessing in this domain, ensuring the robustness and applicability of deep learning models across diverse patient populations (Fitzgerald, Perez-Concha et al. 2023) .

Type 1 Diabetes (T1D) is a chronic condition requiring continuous monitoring and management, which can be challenging for patients. The advent of deep learning has opened new avenues for developing systems that can assist in the self-management, adherence, and education of T1D patients. The choice of an appropriate model architecture is crucial in designing a deep learning model that effectively addresses the unique challenges of T1D management (Domanski, Ray et al. 2024).

Convolutional Neural Networks (CNNs) are widely used in image processing but have also shown potential in healthcare applications, particularly in analyzing time-series data, such as glucose levels. CNNs can automatically extract features from input data, making them suitable for detecting patterns in glucose fluctuations. Research by Xie et al. (2020) demonstrated the effectiveness of CNNs in predicting glucose levels, contributing to better adherence to treatment plans. However, CNNs may require substantial data preprocessing and may not fully capture the temporal dependencies in T1D management data (Rubin-Falcone, Lee et al. 2023).

RNNs, particularly LSTM networks, are designed to handle sequential data, making them a natural choice for modeling the time-series nature of glucose monitoring data. LSTM networks address the vanishing gradient problem associated with standard RNNs, allowing for the capture of long-term dependencies, which is critical in understanding trends in glucose levels, LSTMs can accurately predict future glucose levels, helping patients maintain adherence by anticipating hypo- or hyperglycemic events. However, LSTMs can be computationally expensive and may require careful tuning to avoid overfitting(Domanski, Ray et al. 2024).

Generative Adversarial Networks (GANs) have been explored for generating synthetic data that can augment training datasets, particularly in scenarios where real patient data is limited. GANs can help in creating diverse training examples that improve the generalizability of deep learning models. For instance, Ziyabari et al. (2022) utilized GANs to generate realistic glucose level data, which improved the performance of prediction models. However, the complexity of GANs and the risk of generating unrealistic data are challenges that need careful management.

Transformers, known for their success in natural language processing, have recently been applied to time-series data in healthcare. Their ability to capture long-range dependencies without the limitations of sequential processing, as in RNNs, makes them suitable for T1D management. Transformer-based models, such as the one proposed by shown promise in accurately predicting glucose levels and providing insights for patient education. However, the high computational requirements and the need for large datasets are potential drawbacks (Rubin-Falcone, Lee et al. 2023).

Combining different model architectures can leverage the strengths of each. For example, hybrid models that integrate CNNs with LSTMs can capture both spatial and temporal features of glucose monitoring data, hybrid models could improve prediction accuracy and provide more personalized recommendations for T1D management. The main challenge with hybrid models is the increased complexity in model design and the need for extensive hyperparameter tuning(Porumb, Griffen et al. 2020).

Autoencoders, typically used for unsupervised learning, can be effective in feature extraction and dimensionality reduction, which are essential in handling the high-dimensional data often encountered in T1D management. They can help in identifying the most relevant features for predicting adherence and providing educational insights, the use of auto encoders in reducing the noise in glucose monitoring data, leading to more accurate predictions. However, the risk of losing critical information during the dimensionality reduction process is a concern that needs to be addressed(Alhassan, Budgen et al. 2019).

Selecting an appropriate model architecture for a deep learning model aimed at T1D self-management, adherence, and education is a complex task that requires balancing various factors such as data availability, computational resources, and the specific needs of patients, while CNNs and LSTMs offer strong foundations for time-series analysis, emerging architectures like Transformers and hybrid models provide new opportunities for more accurate and personalized solutions (Chib, & Rossmann, 2020).  Ultimately, the choice of architecture should be guided by the specific requirements of the application, including the nature of the data, the need for real-time predictions, and the goal of enhancing patient adherence and education (Tucker et al., 2020).

Designing the Model

Input Layer

The model takes various inputs such as continuous glucose monitor (CGM) readings, insulin doses, carbohydrate intake, physical activity levels, and patient demographics.

Data Preprocessing

The input data is normalized to ensure consistency in the input range, while Derived features like trends in glucose levels, time since the last meal, are calculated.

Convolutional Neural Network (CNN) Layers

Extract spatial features from the input data.

The Architecture will include.  Conv1D layer, 64 filters, kernel size of 3, ReLU activation.

MaxPooling1D, Pool size of 2.

Conv1D layer: 128 filters, kernel size of 3, ReLU activation.

MaxPooling1D: Pool size of 2.

Recurrent Neural Network (RNN) Layers:

Capture temporal dependencies in the time-series data.

Architecture:

LSTM layer: 100 units, return sequences=True.

LSTM layer: 50 units, return sequences=False.

Attention Mechanism:

Purpose: Focus on relevant time steps and features.

Architecture:

Dense layer for computing attention scores.

Softmax activation to generate attention weights.

Weighted sum of LSTM outputs based on attention weights.

Fully Connected (Dense) Layers:

Purpose: Combine features extracted from CNN and RNN layers.

Architecture:

Dense layer: 256 units, ReLU activation.

Dropout layer: 0.5 dropout rate to prevent overfitting.

Dense layer: 128 units, ReLU activation.

Dropout layer: 0.5 dropout rate.

Output Layer:

Purpose: Provide predictions or classifications.

Architecture:

For self-management: Regression output predicting future glucose levels.

For adherence: Binary classification output (adherent/non-adherent).

For education: Multiclass classification to suggest educational content or actions (e.g., adjusting insulin dose, dietary advice).

Training Process, and Loss Functions.

For self-management (regression): Mean Squared Error (MSE).

For adherence (classification): Binary Cross-Entropy.

For education (multiclass classification): Categorical Cross-Entropy.

Optimization: Adam optimizer with learning rate decay.

Evaluation: Model evaluated using metrics like Mean Absolute Error (MAE) for regression, accuracy, precision, and recall for classification tasks.

User Interface (UI):

Purpose: Present predictions and educational content to patients in an understandable way.

Features: Visualizations of predicted glucose levels, adherence tracking, personalized educational recommendations.

Data Collection: The patient inputs data through a mobile app or wearable device.

Prediction: The model predicts future glucose levels, adherence likelihood, and suggests educational content.

Feedback Loop: The patient receives real-time feedback and suggestions, which they can act upon to manage their condition better.

Supervised learning involves training a model on a labeled dataset, where the input data is paired with the correct output. In the context of T1D management, this could involve training models using historical data of blood glucose levels, insulin dosages, dietary intake, and other relevant variables.

Supervised learning models can achieve high accuracy when trained on large and diverse datasets. For example, studies have demonstrated the effectiveness of supervised learning algorithms, such as support vector machines (SVMs) and random forests, in predicting blood glucose levels based on historical data, these models can be designed to provide interpretable predictions, which is crucial for patients and healthcare providers to understand the reasoning behind recommendations. The performance of supervised learning models is heavily dependent on the quality and quantity of the labeled data. In T1D management, obtaining a comprehensive dataset that accurately reflects the diversity of patient experiences can be challenging (Gupta, Mendonca et al. 2018).

Supervised models are prone to overfitting, especially when trained on small datasets, leading to poor generalization to unseen data. Unsupervised learning involves training models on data that is not labeled, with the goal of discovering hidden patterns or structures within the data. In T1D management, unsupervised learning can be used for clustering patients with similar characteristics or behaviors, which can inform personalized treatment plans (Botvinick, Ritter et al. 2019).

Unsupervised learning models, such as clustering algorithms (e.g., k-means) and dimensionality reduction techniques (e.g., principal component analysis), can uncover patterns in patient data that may not be immediately apparent. These patterns can help in segmenting patients into groups for tailored interventions. Since unsupervised learning does not require labeled data, it can be advantageous in situations where labeled data is scarce or expensive to obtain (Wang, Kurth-Nelson et al. 2016).

Unsupervised models can be less interpretable than supervised models, making it difficult for patients and healthcare providers to understand the reasoning behind the clusters or patterns identified, while unsupervised learning can reveal useful patterns, translating these patterns into  actionable recommendations for T1D management can be challenging. Reinforcement learning (RL) involves training a model to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties. In T1D management, RL models can learn to optimize insulin dosing strategies by simulating interactions with patient data.(Li 2023)

RL is well-suited for problems involving sequential decision-making, such as adjusting insulin dosages throughout the day. Models like Q-learning and deep Q-networks (DQNs) have shown promise in learning personalized treatment strategies based on continuous feedback . RL models can adapt to changing patient conditions, such as varying insulin sensitivity or lifestyle changes, by continuously learning from new data (Gibb).

RL models are complex to train and require a carefully designed reward structure to ensure the model learns appropriate behaviors. This complexity can lead to challenges in model deployment and interpretation. In the context of healthcare, there are concerns about the safety of RL models, as inappropriate recommendations could have serious consequences for patients. Ensuring the safety and reliability of RL models in T1D management is a critical consideration (Schmidt, Lu et al. 2022).

Hybrid approaches combine elements of supervised, unsupervised, and reinforcement learning to leverage the strengths of each method. For example, a hybrid model might use supervised learning to predict short-term blood glucose levels and reinforcement learning to optimize long-term insulin dosing strategies (Fisher-Grace, 2021).

Hybrid models can provide a more comprehensive approach to T1D management by integrating multiple learning paradigms. This can lead to more accurate and personalized recommendations. Hybrid approaches can be tailored to different aspects of T1D management, such as short-term glucose prediction and long-term treatment optimization (Xie, Housni et al. 2023).

The integration of multiple learning paradigms increases the complexity of the model, making it more challenging to train, validate, and deploy. Additionally, the interpretability of hybrid models may be reduced due to their complexity. Transfer learning involves using a pre-trained model on a related task and fine-tuning it for the specific problem at hand. In T1D management, transfer learning can be used to adapt models trained on general health data to the specific requirements of T1D patients (Cristello Sarteau, Ercolino et al. 2024).

Transfer learning can significantly reduce the amount of data and time required to train a model, as it leverages knowledge gained from related tasks. This is particularly useful in T1D management, where labeled data may be limited, By starting with a model pre-trained on a large and diverse dataset, transfer learning can improve the generalization of the model to new patients and scenarios(Cha, Saxena et al. 2022).

The effectiveness of transfer learning depends on the similarity between the source and target domains. If the pre-trained model is based on data that is too different from T1D data, the benefits of transfer learning may be limited. Fine-tuning a pre-trained model for T1D management requires careful selection of hyperparameters and training strategies to avoid overfitting or underfitting (An, Yeh et al. 2023).

The training of models for self-management, adherence, and education among Type 1 diabetes patients involves a variety of approaches, each with its own strengths and limitations. Supervised learning provides accuracy and interpretability but is dependent on labeled data, while unsupervised learning excels at pattern discovery but faces challenges in direct application. Reinforcement learning offers dynamic decision-making capabilities but raises concerns about complexity and safety. Hybrid and transfer learning approaches provide flexibility and efficiency but introduce additional complexity and the potential for domain mismatch (Davies, Aroda et al. 2022).

2.2 OBJECTIVE THREE; To assess the preliminary impact of the model on self-management, adherence, and knowledge among Type 1 Diabetes patients using the UTAUT model, it is important to delve deeper into the concepts of adherence and knowledge in the context of diabetes management.

The management of Type 1 Diabetes (T1D) involves complex self-care routines that are critical to maintaining optimal health. Adherence to prescribed treatment regimens and a solid understanding of diabetes management are crucial components of effective self-management(Farthing, Bally et al. 2022). The Unified Theory of Acceptance and Use of Technology (UTAUT) model provides a framework or assessing the impact of technological interventions on these aspects. Adherence in diabetes management refers to the extent to which patients follow their prescribed treatment regimens, including medication intake, dietary restrictions, physical activity, and regular blood glucose monitoring. Poor adherence can lead to suboptimal glycemic control, increasing the risk of complications such as cardiovascular disease, neuropathy, and retinopathy(Kirkman 2023).

Several factors influence adherence in T1D management. These include patient-related factors (e.g., age, education level, psychological factors), therapy-related factors (e.g., complexity of the regimen, side effects), and healthcare system-related factors (e.g., access to care, provider-patient communication)(Vitale, Asher et al. 2024). Studies have shown that adherence is often challenging due to the demanding nature of T1D management, which requires continuous monitoring and adjustment of insulin doses (Xie 2023).

Technological interventions, including mobile apps and wearable devices, have been shown to improve adherence by providing real-time feedback, reminders, and educational content. The UTAUT model, which examines factors such as performance expectancy, effort expectancy, social influence, and facilitating conditions, can be applied to assess the adoption and effectiveness of these technologies. Preliminary research suggests that when these technologies are user-friendly and well-integrated into daily routines, they can significantly enhance adherence (AlBurno 2024).

Knowledge about diabetes management is another critical factor in effective self-care. This includes understanding the disease process, recognizing symptoms of hypo- and hyperglycemia, and knowing how to adjust treatment based on blood glucose levels. Knowledge empowers patients to make informed decisions about their care and to recognize when to seek medical advice (AlBurno 2024).

Education is a cornerstone of diabetes management. Structured diabetes education programs have been shown to improve knowledge and self-efficacy, leading to better glycemic control. These programs often cover a wide range of topics, including carbohydrate counting, insulin administration, and lifestyle modifications. The effectiveness of education in improving diabetes outcomes underscores the need for continuous patient education, particularly in the context of evolving treatment options and technologies (Lampert-Okin 2023).

Technology plays a pivotal role in enhancing patient knowledge. Digital platforms can deliver personalized educational content, interactive learning modules, and access to peer support networks. The UTAUT model can be used to evaluate the impact of these technological tools on knowledge acquisition. Factors such as perceived ease of use, perceived usefulness, and social influence are key determinants of whether patients will engage with and benefit from these technologies. Preliminary studies suggest that technology-enhanced education can lead to significant improvements in knowledge and, consequently, in self-management behaviors (Núñez-Baila, Gómez-Aragón et al. 2024).

The UTAUT model provides a comprehensive framework for understanding the factors that influence the adoption of technology in diabetes management. By assessing performance expectancy (the belief that using the technology will improve outcomes), effort expectancy (the ease of using the technology), social influence (the impact of others’ opinions), and facilitating conditions (the availability of resources and support), researchers can evaluate the preliminary impact of a deep learning model on self-management, adherence, and knowledge among T1D patients (Romero-Castillo, Pabón-Carrasco et al. 2022).

In the context of T1D management, performance expectancy refers to the extent to which patients believe that the deep learning model will improve their self-management and health outcomes. Studies have shown that when patients perceive that a technology will lead to better glycemic control, they are more likely to adopt and adhere to it (Wamucii 2023).

Effort expectancy pertains to the ease of using the technology. For T1D patients, a model that is intuitive and seamlessly integrates into their daily routines is more likely to be adopted. The design of the user interface, the simplicity of data input, and the clarity of feedback are all critical factors (MUKAMARARA 2022).

Social influence involves the impact of peers, family, and healthcare providers on the patient’s decision to adopt the technology. Support from these groups can enhance the likelihood of technology adoption, as patients may be more motivated to use a tool that is endorsed by others in their support network (Sewell 2024).

Facilitating conditions refer to the availability of resources and support to use the technology effectively. This includes access to technical support, educational resources, and integration with existing healthcare systems. Ensuring that these conditions are met can significantly impact the successful implementation of the deep learning model (OJO, OJO et al. 2023).

The preliminary impact of a deep learning model on self-management, adherence, and knowledge among T1D patients can be effectively assessed using the UTAUT model. By understanding the factors that influence adherence and knowledge in diabetes management, as well as how these factors interact with technological interventions, researchers can better evaluate the potential benefits of the model. The integration of such models into diabetes care has the potential to improve patient outcomes, but their success will depend on careful consideration of the factors outlined in the UTAUT framework (Gonzalez, Tanenbaum et al. 2016).

This review provides an overview of the critical concepts of adherence and knowledge in diabetes management and the application of the UTAUT model in assessing the impact of technological interventions. Further research is needed to explore the long-term effects of these interventions on patient outcomes and to refine the models used to support T1D self-management (Planalp, Kliems et al. 2022).

References

African region tops world in undiagnosed diabetes: WHO analysis (2024) WHO | Regional Office for Africa. Available at: https://www.afro.who.int/news/african-region-tops-world-undiagnosed-diabetes-who-analysis (Accessed: 19 March 2024).

Alkawaldeh, M.Y., Choi, J. and Jacelon, C.S. (2020) ‘A diabetes self-management tablet-based application for older adults with type 2 diabetes: The ASSISTwell pilot study’, Geriatric Nursing, 41(6), pp. 839–845. Available at: https://doi.org/10.1016/j.gerinurse.2020.05.004.

Asiimwe, C. et al. (2011) ‘Use of an innovative, affordable, and open-source short message service-based tool to monitor malaria in remote areas of Uganda’, The American Journal of Tropical Medicine and Hygiene, 85(1), pp. 26–33. Available at: https://doi.org/10.4269/ajtmh.2011.10-0528.

Campbell, F. et al. (2018) ‘Follow-Up Support for Effective type 1 Diabetes self-management (The FUSED Model): A systematic review and meta-ethnography of the barriers, facilitators and recommendations for sustaining self-management skills after attending a structured education programme’, BMC Health Services Research, 18(1), p. 898. Available at: https://doi.org/10.1186/s12913-018-3655-z.

Friedman, J.G. et al. (2023) ‘Use of Continuous Glucose Monitors to Manage Type 1 Diabetes Mellitus: Progress, Challenges, and Recommendations’, Pharmacogenomics and Personalized Medicine, 16, pp. 263–276. Available at: https://doi.org/10.2147/PGPM.S374663.

Janež, A. et al. (2020) ‘Insulin Therapy in Adults with Type 1 Diabetes Mellitus: a Narrative Review’, Diabetes Therapy, 11(2), pp. 387–409. Available at: https://doi.org/10.1007/s13300-019-00743-7.

Katte, J.C. et al. (2023) ‘The phenotype of type 1 diabetes in sub-Saharan Africa’, Frontiers in Public Health, 11, p. 1014626. Available at: https://doi.org/10.3389/fpubh.2023.1014626.

Kumbara, A.B. et al. (2023) ‘Impact of a Combined Continuous Glucose Monitoring–Digital Health Solution on Glucose Metrics and Self-Management Behavior for Adults With Type 2 Diabetes: Real-World, Observational Study’, JMIR Diabetes, 8, p. e47638. Available at: https://doi.org/10.2196/47638.

Kyokunzire, C. and Matovu, N. (2018) ‘Factors associated with adherence to diabetes care recommendations among children and adolescents with type 1 diabetes: a facility-based study in two urban diabetes clinics in Uganda’, Diabetes, Metabolic Syndrome and Obesity: Targets and Therapy, 11, p. 93. Available at: https://doi.org/10.2147/DMSO.S156858.

Li, J. et al. (2020) ‘Application of Artificial Intelligence in Diabetes Education and Management: Present Status and Promising Prospect’, Frontiers in Public Health, 8. Available at: https://doi.org/10.3389/fpubh.2020.00173.

Li, Z. et al. (2022) ‘Temporal trends in the burden of non-communicable diseases in countries with the highest malaria burden, 1990–2019: Evaluating the double burden of non-communicable and communicable diseases in epidemiological transition’, Globalization and Health, 18(1), p. 90. Available at: https://doi.org/10.1186/s12992-022-00882-w.

Lucier, J. and Weinstock, R.S. (2023) ‘Type 1 Diabetes’, in StatPearls. Treasure Island (FL): StatPearls Publishing. Available at: http://www.ncbi.nlm.nih.gov/books/NBK507713/ (Accessed: 10 January 2024).

Maina, P.M., Pienaar, M. and Reid, M. (2023) ‘Self-management practices for preventing complications of type II diabetes mellitus in low and middle-income countries: A scoping review’, International Journal of Nursing Studies Advances, 5, p. 100136. Available at: https://doi.org/10.1016/j.ijnsa.2023.100136.

Mugabirwe, B. et al. (2021) ‘Acceptability and feasibility of a mobile health application for blood pressure monitoring in rural Uganda’, JAMIA Open [Preprint], (ooaa068). Available at: https://doi.org/10.1093/jamiaopen/ooaa068.

Mulvaney, S.A. et al. (2019) ‘A mobile app identifies momentary psychosocial and contextual factors related to mealtime self-management in adolescents with type 1 diabetes’, Journal of the American Medical Informatics Association: JAMIA, 26(12), pp. 1627–1631. Available at: https://doi.org/10.1093/jamia/ocz147.

Ong, K.L. et al. (2023) ‘Global, regional, and national burden of diabetes from 1990 to 2021, with projections of prevalence to 2050: a systematic analysis for the Global Burden of Disease Study 2021’, The Lancet, 402(10397), pp. 203–234. Available at: https://doi.org/10.1016/S0140-6736(23)01301-6.

Sawyer, B. et al. (2022) ‘Barriers and Strategies for Type 1 Diabetes Management Among Emerging Adults: A Qualitative Study’, Clinical Medicine Insights. Endocrinology and Diabetes, 15, p. 11795514221098389. Available at: https://doi.org/10.1177/11795514221098389.

Siedner, M.J. et al. (2012) ‘High acceptability for cell phone text messages to improve communication of laboratory results with HIV-infected patients in rural Uganda: a cross-sectional survey study’, BMC medical informatics and decision making, 12, p. 56. Available at: https://doi.org/10.1186/1472-6947-12-56.

Sseguya, W. et al. (2023) ‘Type 1 diabetes patient experiences and management practices during the COVID-19 pandemic in rural Uganda’, Journal of Diabetes & Metabolic Disorders, 22(2), pp. 1129–1137. Available at: https://doi.org/10.1007/s40200-023-01222-4.

Tahir, F. and Farhan, M. (2023) ‘Exploring the progress of artificial intelligence in managing type 2 diabetes mellitus: a comprehensive review of present innovations and anticipated challenges ahead’, Frontiers in Clinical Diabetes and Healthcare, 4. Available at: https://www.frontiersin.org/articles/10.3389/fcdhc.2023.1316111 (Accessed: 2 February 2024).

Uganda (no date) Type 1 Diabetes Index. Available at: https://www.t1dindex.org/countries/uganda/ (Accessed: 2 February 2024).

Uganda Casts the Spotlight on Diabetes; Adopts Global Diabetes Targets (2023) WHO | Regional Office for Africa. Available at: https://www.afro.who.int/countries/uganda/news/uganda-casts-spotlight-diabetes-adopts-global-diabetes-targets (Accessed: 2 February 2024).

Zhang, P. et al. (2022) ‘Using Momentary Assessment and Machine Learning to Identify Barriers to Self-management in Type 1 Diabetes: Observational Study’, JMIR mHealth and uHealth, 10(3), p. e21959. Available at: https://doi.org/10.2196/21959.

Zhu, T. et al. (2022) ‘Enhancing self-management in type 1 diabetes with wearables and deep learning’, NPJ Digital Medicine, 5, p. 78. Available at: https://doi.org/10.1038/s41746-022-00626-5.

AlBurno, H. (2024). “Ehealth intervention to optimize adherence in late adolescents and young adults with type 1 diabetes: content development and preparing adoption.”

Alhassan, Z., et al. (2019). Collaborative denoising autoencoder for high glycated haemoglobin prediction. Artificial Neural Networks and Machine Learning–ICANN 2019: Workshop and Special Sessions: 28th International Conference on Artificial Neural Networks, Munich, Germany, September 17–19, 2019, Proceedings 28, Springer.

An, L., et al. (2023). “An online structured diabetes self-management education and support programme improves glycaemic control in young adults with type 1 diabetes.” Journal of Multidisciplinary Healthcare: 2641-2654.

Botvinick, M., et al. (2019). “Reinforcement learning, fast and slow.” Trends in cognitive sciences 23(5): 408-422.

Cha, Y. J., et al. (2022). Transitioning toward independence: enhancing collaborative self-management of children with type 1 diabetes. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems.

Cristello Sarteau, A., et al. (2024). “Nutritional Status, Dietary Intake, and Nutrition-Related Interventions Among Older Adults With Type 1 Diabetes: A Systematic Review and Call for More Evidence Toward Clinical Guidelines.” Diabetes care: dci230099.

Davies, M. J., et al. (2022). “Management of hyperglycemia in type 2 diabetes, 2022. A consensus report by the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD).” Diabetes care 45(11): 2753-2786.

Domanski, P., et al. (2024). “Advancing blood glucose prediction with neural architecture search and deep reinforcement learning for type 1 diabetics.” Biocybernetics and Biomedical Engineering 44(3): 481-500.

Farthing, P., et al. (2022). “Type 1 diabetes management responsibilities between adolescents with T1D and their parents: An integrative review.” Journal for Specialists in Pediatric Nursing 27(4): e12395.

Fitzgerald, O., et al. (2023). “Continuous time recurrent neural networks: overview and benchmarking at forecasting blood glucose in the intensive care unit.” Journal of Biomedical Informatics: 104498.

Gibb, C. E. A thesis of clinical research and practice: Part A: Development of a Video Intervention on Self-Management of Type 1 Diabetes in Young Adults Part B: Evaluation of a Video Intervention on Self-Management of Type 1 Diabetes in Young Adults Part C: Summary of clinical practice and assessments, University of Surrey.

Gonzalez, J. S., et al. (2016). “Psychosocial factors in medication adherence and diabetes self-management: Implications for research and practice.” American Psychologist 71(7): 539.

Gupta, A., et al. (2018). “Meta-reinforcement learning of structured exploration strategies.” Advances in neural information processing systems 31.

Kirkman, M. S. (2023). Medical Management of Type 1 Diabetes, American Diabetes Association.

Lampert-Okin, S. L. (2023). The Role of Patient-Provider Symmetry in Influencing Self-Management and Psychosocial Functioning of Emerging Adults With Type 1 Diabetes, Rosalind Franklin University of Medicine and Science.

Li, S. E. (2023). Deep reinforcement learning. Reinforcement learning for sequential decision and optimal control, Springer: 365-402.

MUKAMARARA, I. (2022). Self-care of adolescents with type 1 diabetes mellitus (T1dm) attending rural outpatient medical clinics in Rwanda, University of Rwanda.

Núñez-Baila, M.-Á., et al. (2024). Lifestyle in Emerging Adults with Type 1 Diabetes Mellitus: A Qualitative Systematic Review. Healthcare, MDPI.

OJO, I. C., et al. (2023). “Diabetes self-management education: Benefits and challenges.” Journal of Integrative Nursing 5(2): 93-101.

Planalp, E. M., et al. (2022). “Development and validation of the self-management Barriers and Supports Evaluation for working-aged adults with type 1 diabetes mellitus.” BMJ Open Diabetes Research and Care 10(1): e002583.

Porumb, M., et al. (2020). “Nocturnal low glucose detection in healthy elderly from one-lead ECG using convolutional denoising autoencoders.” Biomedical Signal Processing and Control 62: 102054.

Romero-Castillo, R., et al. (2022). Diabetes management after a therapeutic education program: a qualitative study. Healthcare, MDPI.

Rubin-Falcone, H., et al. (2023). Forecasting with sparse but informative variables: a case study in predicting blood glucose. Proceedings of the AAAI Conference on Artificial Intelligence.

Schmidt, M., et al. (2022). “Learning experience design of an mHealth self-management intervention for adolescents with type 1 diabetes.” Educational technology research and development 70(6): 2171-2209.

Sewell, N. W. (2024). “Discharge Preparedness for the Type 1 Diabetic Patient.”

Vitale, R. J., et al. (2024). “Transition to self-management among emerging adults with type 1 diabetes: a mixed methods study.” Frontiers in Clinical Diabetes and Healthcare 5: 1332159.

Wamucii, E. G. (2023). Self-care knowledge and practices among persons living with Type 2 Diabetes Mellitus attending an outpatient diabetes clinic at Thika level 5 hospital, Kenya, JKUAT-CoANRE.

Wang, J. X., et al. (2016). “Learning to reinforcement learn.” arXiv preprint arXiv:1611.05763.

Xie, L. F. (2023). Design, evaluation, optimization, and adaptation of a web application for type 1 diabetes self-management, McGill University (Canada).

Xie, L. F., et al. (2023). “Adaptation of an adult web application for type 1 diabetes self-management to youth using the behavior change wheel to tailor the needs of health care transition: qualitative interview study.” JMIR diabetes 8(1): e42564.