Diabetes mellitus is one of the major health issues facing today's generation. According to the International Diabetes Federation, about 537 million adults were living with diabetes in 2021, with projections for 2030 and 2045 expected to be 643 million and 783 million, respectively [1]. The disease can be classified into two types: T1D characterized by autoimmune destruction of pancreatic beta cells resulting in absolute insulin deficiency; and T2D characterized by insulin resistance and relative insulin deficiency due to obesity, dyslipidemia, and hypertension [2].

Diabetes mellitus may lead to various complications including microvascular disorders such as diabetic retinopathy, nephropathy, and neuropathy; and macrovascular complications including coronary artery disease, cerebrovascular disease, and peripheral arterial disease [3]. Various studies carried out on the Diabetes Control and Complications Trial (DCCT/EDIC) have shown that controlling the levels of blood glucose (HbA1c<7%) could reduce the risk of cardiovascular disease by 30% and decrease microvascular complications by 35-76%. However, this benefit comes with an increase in the incidence of hypoglycemia [4].

Nonetheless, the cost of diabetes is high. In 2021, more than USD 966 billion was spent yearly on diabetes services globally, representing about 9% of global health care spending amongst adults [1]. In the US, according to the American Diabetes Association, the total annual cost of diagnosed diabetes was estimated to be USD 412.9 billion in 2023, which includes an expenditure of USD 306.6 billion in direct medical costs [5]. Hospitalization caused by the above-mentioned complications makes up most of the financial burden.

Proper management of diabetes mellitus requires early diagnosis, effective treatment options, physiological monitoring, and consistent education, which usually go beyond the scope of conventional ways of providing health care based on periodic visits to clinics, blood sugar measurement through capillaries, and generalized drug prescriptions. Artificial Intelligence can be used to compensate for such inadequacies through machine learning algorithms that allow predicting the onset of hypoglycemia before its occurrence [6], making very precise adjustments to insulin doses [7], forecasting the development of complications many years before their occurrence through traditional biomarkers [8], and classifying different forms of diabetes having similar clinical manifestations through latent variable analysis [9]. Additionally, deep learning algorithms can benefit from data streaming coming from wearable sensors, CGMs, and EHRs to provide personalized decision-making assistance [10].

The above-listed attributes have formed an expanding landscape of AI-driven therapies in the form of CGM devices with integrated hypoglycemia prediction features, “artificial pancreas” devices featuring automated glucose management, remote consultations through telemedicine services, and personalized digital learning resources [11,12]. Studies conducted in various locations have shown benefits of using these technologies in terms of glycemic control indicators, quality of life measures, incidence rates of complications, and health care consumption [13–15].

However, the existing body of evidence includes diverse methodologies, relatively small samples, insufficient validation in external cohorts, and an absence of discussions of equity issues [16-18]. Moreover, a vast majority of scientific publications have been published in connection with high-income countries, where digital health technologies are readily available; this means that the results cannot be generalized to lower-income regions, where there is an increasing number of people with type 1 diabetes.

This systematic review and epistemological meta-analysis presents an analysis of the current evidence base on artificial intelligence and advanced digital technologies in diabetes care through all phases of the care continuum, including diagnosis, treatment, and self-management aspects. We aim to (i) map out the use of AI at each phase of the clinical care pathway; (ii) describe the specific

clinical, behavioral, and educational effects that AI can have; (iii) outline the most effective methods of using AI and potential shortcomings associated with them; and (iv) assess ethical and equity considerations for safe adoption of such innovations.

Figure 1. State of Art of Artificial Intelligence in Diabetic care: An Epistemic Meta-Analysis

Figure 1. Flow chart showing the three steps involved in the meta-analysis of the epistemological framework of artificial intelligence in diabetes management: (1) systematic review of past applications of AI technology in diabetes management; (2) development of a technology roadmap to classify AI technology based on its clinical application, performance indicators, and bioethics; and (3) clinical and ethical convergence.

A PRISMA-guided systematic review was conducted to assess the application of AI throughout the spectrum of diabetic care management, including diagnosis, treatment, monitoring, and self-care. Searches in databases were conducted using PubMed, Web of Science, and Scopus from their inception until April 2026. Publications in the English language and those that were peer-reviewed were considered eligible. Three phases integrated the methodology of the study: (1) systematic identification of evidence and assessment of its quality; (2) development of a classification roadmap of AI technologies; and (3) clinical and ethical consideration for implementation. The overall framework of methodology is shown in Figure 1, while the operationalization of the search strategy is detailed in Figure 2 and the PRISMA flow of records in Figure 3. 2.1. PIO and PRISMA Strategy The application of the PIO (Population, Intervention, and Outcome) model was used to frame the systematic review. The PIO model was chosen in preference to the usual PICO model since the aim of this systematic review is not about comparing artificial intelligence with a different intervention but assessing the effect of artificial intelligence in the entire process of diabetes mellitus management. The key research question guiding this review is: In what ways is artificial intelligence changing diabetes mellitus diagnostics, management, and treatment compared to normal practice, and how are digital interventions being affected by this transformation? 2.2. PIO Strategy Components The components of the PIO framework were defined as follows. First, the Population component (P) was defined by studies involving adults or adults and children having either type 1 or type 2 diabetes mellitus and prediabetes and metabolic syndrome who needed an assessment of the use of AI technologies in the management of their condition. The Intervention component (I) was defined by studies evaluating advanced digital health technologies including machine learning, deep learning, neural networks, reinforcement learning, natural language processing, digital health platforms, telemedicine applications, continuous glucose monitoring systems with integrated analytics, and clinical decision support systems. Finally, the Outcomes component (O) was defined by studies that reported clinical outcomes (HbA1c, percent time in range, hypoglycemia rate, complications), diagnostic test performance metrics (sensitivity, specificity, area under the receiver-operating characteristic curve), behavioral outcomes (adherence, self-efficacy), or educational outcomes (health literacy). Figure 2. PIO Search Strategy Decision Tree Figure 2 shows how the PIO approach was used to organize and include both controlled and uncontrolled vocabulary in the PubMed, Scopus, and Web of Science databases. Use of Boolean operators (AND, OR) and Mesh terms made the research process reliable and consistent. 2.3. Search Strategy The keywords used in the search strategy include those that are synonyms for the terms "diabetes mellitus," "Type 1," "Type 2," "artificial intelligence," "machine learning," "deep learning," "neural network," "continuous glucose monitoring," "closed-loop insulin delivery," "telemedicine," "digital health," "self-management," and "clinical outcomes." The Boolean operators AND/OR and the MeSH terms used were adjusted depending on the bibliographic database consulted. The full search string is presented in Figure 2 below. Temporal Scope and Historical Context of the Search Strategy Taking into consideration the fast development of AI technology and the emergence of novel digital health innovations, no temporal floor was specified within the search strategy. Previous research was included to offer a historical perspective for the process of the development of technological solutions applied to the management of diabetes. Despite potential differences in methodology among various eras, it was deemed important to include past research to trace the technological path that has led to modern technologies. Preference, however, was given to recent papers dating back from 2018. 2.4. Study Selection The records were transferred to a reference management software tool, wherein duplicate studies were automatically de-duplicated and manually verified. Titles and abstracts of records were independently screened by two reviewers before carrying out full-text screening for those deemed potentially relevant. Study inclusion criteria included the following: (i) original studies or well-conducted systematic reviews with a quantitative synthesis; (ii) adult subjects or mixed adult/pediatric subjects with either diabetes mellitus or prediabetes; (iii) AI or digital health solutions as the main topic; and (iv) reporting clinical, diagnostic, and behavioral outcomes. Those studies that assessed the effectiveness of non-AI digital solutions with no AI component whatsoever were only included if the study aim was the development/evaluation of algorithms for diabetes management. All disagreements between reviewers were resolved by discussion and consensus. Figure 3. PRISMA flow diagram Figure 3. PRISMA flow diagram illustrating the stepwise selection of studies from initial identification (n = 5,847) through screening, eligibility assessment, and final inclusion (n = 22) in the epistemic meta-analysis. Population Stratification and Inclusion Criteria Eligible subjects were divided into two groups to maintain clarity in concepts: (i) adult studies, which included participants aged ≥18 years suffering from either type 1 or type 2 diabetes mellitus or had pre-diabetes or metabolic syndrome; and (ii) mixed-aged studies, which included both adolescent and adult participants. Adult studies were included without any limitations while assessing the effects of artificial intelligence on diagnosing, monitoring, treating, predicting, and educating patients with diabetes. Mixed-aged studies were included only if (a) separate outcomes for pediatric and adult patients were described or (b) the aim was the development/validation of AI techniques applicable in the field of adult diabetes management. Figure 4. Stratification framework for population selection in diabetic care studies Figure 4. The framework provides criteria to make decisions about age, type of diabetes, clinical context, use of AI or digital health technology, and relevance to outcomes, thereby ensuring that only articles relevant to the methods and clinical application of diabetic care among adults and mixed ages are synthesized. 2.5. Quality Appraisal The evaluation of each paper was done via a 27-item quality assessment checklist based on PRISMA criteria. The criteria included the following: (a) clarity of objectives (20%); (b) compatibility between the research problem and methodology (20%); (c) appropriateness of methods utilized given the nature of the study design (20%); (d) appropriateness of outcome measures (20%); and (e) adequacy of results reported (20%). A score of 80% or more was deemed the minimum for acceptance into the epistemic meta-analysis. Twenty-two papers met this criterion (Figure 3, Table 1). 2.6. Data Extraction and Synthesis For each study, we identified the study design, characteristics of the participants, sample size, artificial intelligence method used, comparison group, main outcomes (such as changes in HbA1c, percentage in range, accuracy of prediction, compliance, and complications), performance measures, and issues regarding equity and ethics. Due to the high degree of heterogeneity in terms of design, intervention, target groups, and endpoints, a meta-analysis was not possible. Results are therefore presented in narrative synthesis and tables, focusing on trends. 2.7. Classification of Artificial Intelligence Models Studies that have used specific AI-based algorithms have been reviewed by categorizing the AI models according to their algorithmic type: (i) supervised machine learning models consisting of logistic regression, random forests, support vector machines, gradient boosted tree algorithms, and decision tree algorithms; (ii) deep learning models comprising CNNs, RNNs, LSTM networks, transformer models, and deep belief network; (iii) reinforcement learning models mainly used for insulin dosing; (iv) unsupervised models like hierarchical clustering and k-means clustering used for disease phenotyping; and (v) hybrid models using physiological knowledge together with data-based models. Digital health platforms without the application of any predictive algorithms have also been segregated separately from the above categories. 2.8. Integral Epistemic Meta-Analysis As per the suggested epistemic meta-analysis by Ayala et al. [19], we conducted an epistemic meta-analysis that takes into account the theories used, the methods employed, and the validation process behind every single study that was included in this meta-analysis. Such an approach is especially relevant to a dynamically changing field of research, such as using AI to improve the quality of life of people suffering from diabetes, where the methodologies involved differ considerably, as do the outcomes considered. Unlike traditional meta-analysis that calculates average effect sizes, an epistemic meta-analysis highlights common trends within certain important variables (reduction in HbA1c levels, improvement in time in range, prediction capabilities, patient compliance with the treatment plan). Epistemic meta-analysis also helps identify existing knowledge gaps and assess how mature each technique is when it comes to implementation in a real-life environment. An epistemic meta-analysis can additionally be augmented with a hermeneutic analysis, which is one of the key features of any narrative systematic review, although in our case we preferred directional integration over thematic coding.

3.1. State of the Art: Diabetes and Artificial Intelligence

The total number of records obtained from PubMed (n = 2,981), Web of Science (n = 2,193), and Scopus (n = 673) was 5,847 according to the predetermined PIO approach and PRISMA protocol (Figures 2 and 3). Following duplication removal through an automated and manual process, there were 2,614 duplicates removed, resulting in 3,233 unique articles. Records that were irrelevant to the research topic were filtered out based on their titles and abstracts, leading to the exclusion of 2,890 records. From the 343 articles that passed the screening, 321 were removed because they did not satisfy the inclusion criteria, which include non-AI digital technology without an algorithmic element (n = 124), non-diabetic subjects (n = 87), non-English publication (n = 34), conference papers without peer-reviewed full texts (n = 46), and poor quality rating (n = 30). Twenty-two studies met all eligibility and quality criteria and were included in the final epistemic meta-analysis. The summarized quality scores are presented in Table 1.

Table 1. PRISMA-derived quality appraisal of included studies. Each domain contributes 20% to the overall score; a minimum threshold of 80% was required for inclusion.

3.1.1. Geographical Distribution and Digital Infrastructure

Geographical analysis of the provided data indicated a strong inclination for cluster formation in the context of wealthy and technologically developed countries. The top contributing country was the USA (36.4%), followed by China (22.7%), the UK (9.1%), and individual contributions from Italy, Spain, South Korea, Norway, Singapore, Kuwait, Saudi Arabia, India, and Croatia. This kind of distribution clearly indicates that there is always the existing North-South technology divide, considering that the vast majority of technological breakthroughs in the field of artificial intelligence-based diabetes management occur in places with already established digital infrastructure and device density, along with the relevant regulatory environment. Resource-limited areas, particularly those located in Sub-Saharan Africa and South Asia, where the highest rates of diabetes incidence can be observed, remain underrepresented.

3.1.2. Epistemic Meta-Analysis Outcomes

Consistent with the meta-analytical approach of Ayala et al. [19] based on epistemological reasoning, our synthesis process was divided into five key stages, including: (1) identification of credible evidence concerning the research problem (Table 1); (2) analysis of implicit and explicit knowledge in the use of AI in the treatment of diabetes (Tables 2 and 3); (3) identification of critical knowledge gaps (Section 4.1); (4) examination of knowledge transfer and exchange processes within the field; and (5) knowledge integration innovation concerning AI in diabetic care (Figures 6 and 7).

The 22 studies reviewed above have demonstrated that AI technology is changing the paradigm of diabetes via the following five interrelated fields: (i) risk stratification and early diagnosis, where machine learning and deep learning algorithms based on EHR and biomarkers enable to detect patients at risk prior to meeting the clinical criteria; (ii) real-time monitoring and glycemic control, where the use of CGM combined with prediction models/closed-loop system enables to provide superior glucose control compared to conventional methods; (iii) complications diagnostics and early diagnosis, where the application of CNN technology and prediction models enable to diagnose retinopathy, nephropathy, neuropathy, and CVD with high accuracy; (iv) digital self-care and education, where the use of conversational AI, gamification, and telemedicine promotes compliance, health literacy, and independence; and (v) personalized insulin delivery, where reinforcement learning and bolus advisory improve insulin dose.

Study Architecture and Sample Range

The 22 selected studies were grouped according to four main themes: (i) study design; (ii) intervention type; (iii) technological use; and (iv) target population. The above categorization served as an integration structure for reviewing the current use of AI and digital technology applications in diabetes treatment (Figure 5). Concerning the design of studies, the included literature used different types of research, namely randomized controlled trials [12,14,20,23,27,31], observational studies [24,25,28], systematic and narrative reviews [11,13,21,26], and computational studies [10,15,16,17,18,22,29,30,32]. As for intervention type, various kinds of digital solutions were addressed, such as ML prediction, deep learning image analysis, closed-loop algorithms, telemedicine applications, conversational AI systems, and digital diabetes self-management interventions. Finally, regarding the target population, adults with T1D, adults with T2D, mixed adult T1D, mixed T1D/T2D groups, and hospitalized patients were included.

Classification of included studies by design, intervention, application, and target population

Figure 5. Classification of the 22 included studies according to study design, type of intervention, technological application, and target population. The concept map visualizes the methodological and thematic diversity of the included evidence and the relationships between technological approaches and clinical purposes.

The distribution of study populations, AI and digital approaches, reported clinical outcomes, performance metrics, and key methodological limitations across these categories is synthesized in Table 4.

Table 2. Implementation, learning models, and validation of digital platforms and AI systems in diabetic care.

Table 3. Differential clinical roles of artificial intelligence in type 1 and type 2 diabetes.

Intervention Typology

The identified interventions were grouped into six groups based on technological basis, purpose, and level of user engagement:

1. Prediction using artificial intelligence and risk stratification. Roughly 36.4% of studies implemented machine learning and deep learning techniques for glucose prediction, risk prediction for complications, or type determination [10,16,17,21,25,28,30]. The interventions yielded reliable performance improvements measured in terms of AUROC with the range of 0.84–0.98 related to hypoglycemia, diabetic retinopathy, CKD progression, and cardiovascular events.

2. Integrated monitoring and closed-loop management with CGM. Predictive algorithms implemented in CGM and hybrid or closed-loop insulin delivery systems increased TIR by 11–15 percentage points and decreased HbA1c by 0.4–0.6% as opposed to pump or MDI regimens [12,14,31].

3. Self-management and education digitally assisted. Applications, chatbots, and other digital educational tools increased adherence by 24–35%, and improved self-efficacy and health literacy in 60–70% of patients [23,26,27].
4. Telemedicine and remote monitoring. Video conferencing, asynchronous communication platforms, and CGM sharing software facilitated increased interaction between patients and clinicians and demonstrated that they have the capability to maintain their glycemic levels despite disruptions in healthcare, such as the COVID-19 pandemic [24].
5. Computer vision-based complications screening. CNN models for screening diabetic retinopathy and diabetic foot ulcers via fundus photographs, wound images, and ECG signals achieved similar or even better results than those made by specialists in those fields [15,28,29].

6. Data enrichment and decision support based on NLP models. NLP models created using transformers increased the number of variables contained within risk factor data sets from medical records significantly [18].

Over 60% of the included studies showed a statistically significant or clinically relevant improvement in primary outcome measures. The common limitations found in the methodology of studies included the use of a single center, small number of participants, lack of standardization in evaluation, and short follow-up periods of 12-24 weeks on average. Furthermore, there was no external validation of AI models used in most cases.

Table 4. Overview of clinical and methodological aspects of artificial intelligence and digital interventions in diabetic care.

3.1.3. The Digital Shift in Diabetic Care: Results Across 22 Studies

The analysis of the cross-article demonstrates how rapidly diabetes treatment is becoming digitized and automated, and researchers are shifting their interests from episodic care towards automated monitoring systems, predictive risk assessment, and telemedicine applications. The largest part of publications concerns the use of computational modeling and algorithm validation followed by randomized control trials and systematic review addressing the feasibility and efficiency of novel diabetes management techniques in the clinical practice. Advanced ML algorithms process large amounts of data provided by CGM sensors, insulin pumps, EHRs, fundus imaging, and wearables along with cloud computing solutions in order to provide personalized recommendations [13,22,30].

It should be noted that among the 22 papers under analysis, 12 employed AI-based predictive models, including neural networks, gradient boosting algorithms, reinforcement learning agents, and transformers, whereas the other 10 articles referred to physiological measures, educational interventions, and telemedicine applications [12,14,23,24]. The purposes that drive the innovations mentioned above can be stated as follows: prediction and prevention of hypoglycemia, optimization of glycemic levels through remote visits, improvement of adherence and self-management, screening for complications using low-cost POC tests, and evidence-based decision-making grounded on digital literacy [15,21,32].

The use of AI technologies showed higher sensitivity and specificity in six tests, implying that the time for the integration of such technologies into the health care system has come [10,15,16,21,25,28]. The digital literacy was one of the most important moderators of effectiveness in intervention as patients and their families with better skills of using technological devices achieved significant results due to the application of AI technologies [26]. The self-management education program has identified several communication barriers between patients and doctors which need to be overcome to make personalized treatment possible [23,27].

3.1.4. Two Paradigms of Digital Care: Platforms versus Predictive Engines

The 22 included studies reveal a clear bifurcation in the landscape of AI-enabled diabetic care, organized into two intersecting paradigms. The first paradigm, called platform oriented, comprises around 45% of all research papers, and includes digital health solutions for visualizing patterns of glucose level, analysis of health information, telemedicine, and education of patients and their families [11,23,24,26,27]. By providing knowledge about glycemic patterns to everyone involved in the process (patients, families, doctors), such systems facilitate adjustment of therapy via collaboration and reduce the burden of diabetes management. In addition, the platforms are equipped with various reinforcement methods, game elements, and intelligent conversational systems to engage and modify user behaviors, especially among patients with T2D.

The second paradigm—that of prediction-based, comprising ~55% of the literature—relies on ML and DL methods for addressing more difficult challenges such as predicting episodes of hypoglycemia several hours ahead of time, independently controlling insulin administration, quantifying 10-year risk for diabetic retinopathy or nephropathy, discovering new diabetes subgroups with different patterns of complications, and/or extracting risk-associated insights from free-form clinical notes [10,12,14,15,16,17,18,20,22,25,28,29,30,31,32]. Such advanced technologies are mostly known for their substantial improvement in clinically relevant outcomes of HbA1c levels, diagnostic performance, TIRs, and absence of complications. Some also provide additional evidence for enhanced patient well-being and decreased disease-induced psychological stress. Unfortunately, methodological soundness varies greatly, with model validation encompassing prospective independent cohorts, simulation studies, or retrospective single-center evaluations. Very few models have been granted regulatory approval and incorporated into clinical workflows.

3.1.5. Methodological Landscape of AI-Driven Diabetes Tools

Viewed through a technical lens, the 22 studies reveal a broad and rapidly maturing toolkit of algorithms, data streams, and validation strategies. The elements - ML class, features, targets, and limitations are listed in Table 2. The overwhelming majority of the teams employed supervised learning for classification of states and prediction of future events [10,16,21,22,25,32]. Some teams experimented with deep learning techniques for modeling complex physiological processes: CNNs for analysis of fundus and wound images [15,29]; LSTM and transformers for modeling glucose levels [13,22,28,30]; and transformers for performing natural language processing tasks, like risk factor extraction [18]. Innovative approaches include incorporation of physiological reasoning along with machine learning techniques, such as neural networks for insulin absorption that improve predictions, and closed-loop control using MPC with pharmacokinetics/pharmacodynamics modeling [12,14,31].

Whereas data provenance was not considered crucial, the focus of the study was rather on organizing data and making sense of it. Fixed clinical variables such as age, gender, HbA1c level, insulin dose, and body mass index are combined with frequent behavioral and physiologic data like physical activity or circadian rhythm [11,18,23,30]. Handling such complex and sequential data would have likely made the researchers opt for deep neural networks capable of handling long-term relationships. The algorithms address five clinical goals: (i) glucose prediction [13,22,30]; (ii) closed-loop personalized insulin administration [12,14,20,31,32]; (iii) prediction of side effects [15,16,21,28]; (iv) early diagnosis by means of biomarkers and images [17,18,29]; and (v) tailoring educational and psychological interventions promoting patient compliance [23,26,27].

Nevertheless, despite its complexity, implementation becomes difficult. Most algorithms were evaluated using smaller than 100-patient cohorts, simulations, or no external validation at all [22,29,32], thus limiting their application range. Infrastructure challenges such as lack of connectivity, costly equipment, and poor digital literacy further restrict adoption in developing countries. Whereas the progress made in science is obvious, there is also an urgent need to develop more robust data sets, ethics and regulation policies, and infrastructure in order to turn these technologies into reality.

Figure 6. Five-phase roadmap for the integration of artificial intelligence in diabetic care

Figure 6. The transition map shows the evolution of the management approach from episodic/reactive management through the steps of digitalization, predictive AI, intelligence, integration and adaptability in care to AI-driven precision medicine.

3.1.6. The AI Care Pipeline: Linking Base Data to AI Outputs

The maximum effects ML technologies have can be observed in glucose prediction, insulin adjustment, and patient compliance [12,14,20,30,31,32]. Telemedicine enables physicians to monitor the patients remotely through video sessions and continuous monitoring; hence, rural residents will gain from fewer visits, lower costs, and instant modification of their dosages [13,23,24]. Educative programs continue to widen the digital health landscape. Patients gain access to smartphone applications, motivational training programs, and serious games that help improve their dietary behavior and encourage physical activity by providing instant communication options for healthcare professionals [23,26,27].

The most advanced research projects aim to create digital twins based on clinical, genomic, and metabolic information, which will predict the future disease profile of each individual. In conjunction with artificial intelligence-based analysis tools, these will detect retinopathy, nephropathy, and cardiovascular conditions several years in advance of existing markers, heralding a new dawn of prevention medicine [10,15,16,28]. But there are marked differences in the extent of research being done in this field. Research on smaller sample sizes, use of simulators for testing, and vast differences in availability of technology restricts the availability of such innovations to just a select few. It is extremely important for researchers to find means of validating their findings across a wider demographic, implementing effective reimbursement systems, and conducting digital literacy programs.

3.1.7. Digital Equity and Ethical Considerations in AI Implementation

The continuing lack of access to health care may overshadow the possible positive impacts that could arise from the use of AI for managing diabetes. Countries with established digital infrastructures (e.g., the US, UK, Germany and, to some extent, parts East Asia) have seen implementation of various technologies while, in countries that have fewer resources, none of these opportunities have existed [26]. In addition to access issues among low resource areas, low levels of digital literacy exist among patients and their caregivers in developed economies. Patients and caregivers with low literacy also experience barriers to using well-designed intervention programs created for them, particularly those living in impoverished communities. Lastly, an ethical framework is yet another obstacle that still needs significant improvement. There are well defined expectations regarding the provision of data protection standards, obtaining informed consent, and auditing algorithms; however, the implementation of these expectations varies throughout the world.

3.1.8. The Endocrinology–AI Binomial: Driving Innovation in Modern Medicine

With the incredible amount of data involved with endocrinology and its longitudinal nature, endocrinology is well suited for the use of AI applications. The use of AI allows for the examination of high frequency (time component), multimodal datasets that will identify latent patterns and build individualized and dynamic predictions that can assist with clinical decision-making in a timely and accurate manner [10,16,25]. Use of AI is evident in the management of diabetes and how it can expand the current synergy between endocrinology and AI. To optimally manage diabetes, patients will require constant adjustments (of insulin), continual monitoring, and continuous education and behavior modification interventions [12,14,23]. AI applications, through machine learning and deep neural networks, can manage many of these tasks, such as optimizing insulin delivery, minimizing user errors, and stabilizing blood sugar levels [12,14,31]. Additionally, AI's utility in endocrinology is in the area of preventive medicine. Algorithms and models developed from longitudinal clinical datasets and supplemented by genomics, metabolites, and other omic profiles will be able to accurately stratify risk and predict the occurrence of disease complications years before their actual clinical manifestations [16,17,28].

Figure 7. Transition toward digital medicine and artificial intelligence in diabetic care.

Figure 7. The figure represents the evolutionary process from conventional medicine to the current AI-based education-led approach in healthcare, identifying important enablers for each stage of development.

3.1.9. Data Privacy, Integration, and User-Centered Outcomes in AI-Enabled Diabetic Care

The application of AI technologies and digital platforms in diabetes management presents several concerns with regards to data privacy when it comes to its utilization and incorporation within algorithms. All AI implementations described in the selected studies employ data that is gathered in real time via wearable devices and applications that transfer their data to clouds that are connected to the patients' EHRs [12,18,23]. This data feeds the algorithm and facilitates personalized treatment decisions. However, the use of such technology places patients at risk due to privacy and security threats. The communication between various healthcare providers, technology suppliers, and patients is not discussed in the literature, which makes the implementation of data privacy more difficult because of the lack of regulatory framework that would ensure effective AI-powered decisions [24,26]. Even where such measures exist, their implementation is still an issue in the context of continuously learning algorithms. Nevertheless, the studies highlight a single important consideration, which suggests that the success of clinical AI application relies on proper data protection policies and training of health professionals.

3.1.10. Ethical Foundations and Bioethics Governance in AI-Enabled Diabetic Care

The development of an ethical framework for the implementation of AI tools/technologies to manage diabetes will require a consideration of a wide range of factors beyond simply technical efficiency (see Figure 8). With respect to ethics and practice, diabetes patients should be viewed as a unique individual rather than simply considered via algorithm-driven data/results [19]. Many older adults with diabetes may be cognitively impaired or economically disadvantaged and may have difficulty understanding informed consent, navigating technology-assisted monitoring, and accessing technology-assisted medical care. As a result, it is essential to ethically develop human-centric AI technologies per the principles of beneficence, non-maleficence, justice, and respect for autonomy [12,23,26].

Figure 8. Conceptual framework for the responsible and ethical use of artificial intelligence in diabetic care

Figure 8. The seven steps of the model include: validating data; consenting to use that data; secure storage; contributing to AI; controlling access to the data; governing the use of AI; and implementing clinical systems where AI are viewed only as assisting in a diagnostic process not replacing normal clinical judgement

.3.1.11. Artificial Intelligence, Clinical Judgment, and the Patient–Clinician Relationship

The use of AI technology poses several difficulties in terms of the relationship between patients and clinicians because of its role in clinical decision-making and the automation of predictive analysis. Even if the use of such algorithms helps to ensure the accuracy and safety of decision-making, AI systems cannot substitute clinical decisions because their work needs to be continually monitored and adjusted. In the case of glucose prediction and insulin dosing, the multidisciplinary team remains responsible for the correct interpretation of results and subsequent action to avoid acute and chronic complications [20,25,28]. The same principle applies to other areas of work within the team since AI serves as a supplement to clinical reasoning in each specialty [12,23,26].

3.1.12. Impact of AI on the Management of Type 1 and Type 2 Diabetes

Whereas it was expected that AI would have a uniform effect, it turns out that AI influences T1D and T2D management in a different, complementary way (see Table 3) [10,11,12]. With respect to T1D, AI serves as a monitoring, follow-up, and automation tool, facilitating patient independence [13,14,20,22,28,32]. Improvement rates were found to be between 0.2% and 0.6% reduction in HbA1c (usually achieving below 7%), TIR improvements of 5–15 percent, and less clinically relevant hypoglycemia events when using AI-assisted tools [12,14,31].

In T2D, the application of AI is geared towards predicting and preventing risk, which entails stratifying metabolic and cardiovascular risk at an early stage using longitudinal analysis of medical records in addition to educational and digital tools for making changes to lifestyle [21,23,26,27]. The degree of clinical utility relies heavily on patient compliance and the effectiveness of multidisciplinary interventions [25,29]. The application of AI technology relies not just on its sophisticated nature but also on its proper integration within supervised medical practices and digital literacy among patients, their caregivers, and healthcare providers. There is evidence that using automation for preventive care can aid in maintaining lower hospital admissions and critical incidents with type 1 diabetes (T1D)), and that predicting and preventing complications in Type 2 Diabetes (T2D) cases is essential. However, the methodology utilized in previous studies could be improved (such as, through longitudinal studies demonstrating the results noted in adult populations) [26,30].

Artificial intelligence (AI) has transitioned from a novel supplementary tool to a working component of the transformation of the diabetes care delivery model; as presented herein through the review of 22 quality studies, definitive changes from diagnosis through treatment, monitoring, and education for people with diabetes will be achieved through innovative technology over time [10,32]. For example, machine-learning classifiers using electronic health records (EHRs) have demonstrated success in identifying individuals at high risk of developing type 2 diabetes up to five years before they attain clinical diagnostic criteria, with AUROC scores approaching 0.90 [10]. Unsupervised learning has uncovered hidden subtypes within the apparent dichotomy between T1D and T2D, each having a unique genetic signature and different susceptibility to complications and treatment strategies [17]. Through the application of NLP, risk-factor information embedded within unstructured data can be efficiently extracted [18]. Predictive capabilities range throughout the entire clinical timeline. On the short-term scale, LSTM and transformer algorithms trained on CGM data accurately predict glucose fluctuations well enough to allow for prevention either through patient action, clinical intervention, or even automated insulin delivery [13,22,30]. On the intermediate scale, ML algorithms predict the likelihood of 4-hour hypoglycemia in hospital patients with AUROC > 0.90, allowing nursing staff and pharmacists to act before complications arise [25]. On the long-term scale, ensemble algorithms stratify the risk of developing CKD, cardiovascular complications, and death with accuracy that far outperforms standard clinical scoring tools, paving the way for individualized complication monitoring and prophylactic treatment plans [16,21]. The revolution in treatment can be most easily witnessed in the realm of insulin therapy. Hybrid closed-loop devices that use model predictive control technology algorithms, which have now received FDA clearance and CE marking, consistently improve TIR by 10%–15% in comparison to conventional pump therapy in T1D, providing clinically relevant improvement in terms of reducing hypoglycemia risk and variability in blood sugar levels [12,14,31]. The fully automatic systems remove the necessity for carbohydrate counting and boluses, completely changing the life of T1D patients on a day-to-day basis [14]. RL-driven algorithms trained in physically plausible simulation environments begin to personalize the dose of insulin delivered in T2D, with preliminary studies showing clear clinical benefits in terms of increasing TIR and decreasing hypoglycemia risks [20]. ML-based bolus advisory mealtime support systems are an intermediate solution for patients who choose not to abandon human input altogether [32]. The improvement to the processes of identifying complex entities through the use of artificial intelligence (AI) technology occurs through convolutional neural network (CNN) algorithms, are constructed from large volumes of data on fundus photographs used to diagnose diabetic retinopathy at all stages of severity, achieve better overall accuracy in terms of both sensitivity and specificity than current human experts in this area (i.e., ophthalmologists), and are currently being utilized in community based outreach programs [15]. Using photographs of diabetic foot ulcers captured via smartphones in resource-poor areas, computer vision (CV) algorithms provide similar diagnostic accuracy as trained medical professionals, showing that AI potentially can replicate the abilities of trained medical professionals even where they do not exist [29]. Long short-term memory (LSTM) neural networks, trained on continuous glucose monitoring (CGM) datasets collected over time, can accurately predict who will develop increased cardiovascular risks many years prior to the development of abnormal conventional cardiovascular risk markers in type 1 diabetes (T1D) patients [28]. The innovative technologies mentioned are also supporting the establishment of an integrated digital self-management ecosystem. In individuals with T2DM, conversational AI platforms, behavioural coaching engines, and gamified self-management tools have been demonstrated to lead to a decrease in HbA1c between 0.7% and 1.1%. The magnitude of these decreases in HbA1c is comparable to the addition of another medication to the treatment regimen [23,27]. Telehealth platforms have also maintained dry glycemia during healthcare interruptions, and have been successful in overcoming barriers to access for patients from rural and underserved populations [24]. The combination of these disparate technologies into a unified digital health ecosystem (integrating CGMs, insulin administration records, medication information, behavioural metrics, and patient reported data into a single longitudinal data collection stream) represents the foundation for precision-medicine approach for the management of diabetes in the future. 4.1. LIMITATIONS OF THE EVIDENCE This systematic review has several methodological limitations that must be borne in mind when interpreting the findings of this review. For example, while the studies included in the review used the PRISMA standards for high quality and reporting of studies, the studies vary widely in terms of their methodologies (i.e., randomized controlled trial vs observational cohort vs systematic review with statistical analysis vs computational modeling research vs pilot feasibility study) and as such, they had a high degree of heterogeneity. This variety of methodologies hindered conducting meta-analysis, making it challenging to draw robust conclusions regarding the effectiveness of interventions in different settings and populations. Second, AI-based interventions have been studied on cohorts varying in size from 20 to more than 60,000 individuals and follow-up time from 4 weeks to 10 years. While the majority of interventional studies relied on relatively small samples with 6 months of follow-up at most, limiting the opportunity to assess the long-term efficacy of interventions in terms of glycemic control and potential complications, some predictive models could only be validated in the derivation cohort and artificial scenarios. Third, the bias toward evidence originating in wealthy regions such as North America and East Asia poses a problem when it comes to generalizing findings to LMICs where the prevalence of diabetes is highest and increasing fastest. Even in high-income countries, vulnerable groups such as racial and ethnic minorities, low-literacy populations, and those lacking access to the internet are not represented among the validation datasets of the majority of the studied AI tools, which could compromise the ability of those tools to function properly in the very patients who would most benefit from them. Fourth, cost-effectiveness, business cases, and approval mechanisms of the studied applications were scarcely addressed, despite their critical importance for transitioning novel technologies into healthcare practice. Lastly, very few of the included articles addressed the topic of ethical frameworks for ensuring data privacy and algorithmic accountability of AI-based tools used in clinical decision making. 4.2. Future Directions and Clinical Readiness The future path of AI in diabetes treatment includes several major milestones. One such milestone would include digital twins, where AI models are trained using individual patients' genome data, metabolic data, physiological data, and behavioral data, and then predict individual disease progression and suggest preventive measures before the onset of any complications [11,16]. Milestone implementation also relates to the use of federated learning wherein several clinical datasets can be used to create AI learning groups across geographies and will prevent the need for a centralised patient database for AI training with cases across multiple ethnicities/Nationalities (26). The ability to develop multimodal artificial intelligent models that integrate information from various sources (i.e., continuous glucose monitors (CGM), diet, physical activity, medication adherence, psychosocial) all at the same time could change the current treatment paradigm for diabetes. There is a spectrum of maturity that currently exists relative to the clinical readiness of applications of AI to relevant diabetes-related technologies. The highest level of clinically ready AI applications would include closed-loop insulin delivery systems that incorporate CGM hypoglycaemia forecasting software used for the management of T1D which have both been regulatory approved and supported by real-world use evidence (12,14,31). Deep learning AI-based diabetic retinopathy detection via fundus imaging is being integrated into population-based screening initiatives and is currently under regulatory evaluation in several countries (15). Conversational AI self-care applications for patients with T2D have been put into commercial use with promising early results (23,27). Reinforcement learning for T2D insulin dosage optimisation is in development/testing, long-term glucose forecasting using transformer architectures is currently being studied, and digitally engineered "twin" AI models are being developed for risk profiling studies (20,30). In order to maximise the potential of Artificial Intelligence (AI) for people with diabetes worldwide, there are numerous different activities that need to take place simultaneously. The validity and sustainability of current data must be validated through multicenter, multi-ethnic clinical trials with prolonged follow-up periods. Creation of a flexible, yet rigorous regulatory framework is necessary for the efficient delivery of effective innovations to patients, while maintaining the highest level of safety. Modification of financial systems is also required to ensure accessibility to advancements in technology for all persons with diabetes. Education should include components regarding digital literacy and AI understanding. Finally, persons with diabetes should be actively involved in both the development and evaluation of these innovations.

The application of AI (artificial intelligence) technology to the field of diabetes management has undoubtedly progressed from the theoretical to the practical domain. The literature base on this subject was evaluated through 22 high-quality studies (i.e., the impact of AI technology on metabolic control, early detection of complications, personalized insulin delivery, and enhanced patient self-management in both T1D [type 1 diabetes] and T2D [type 2 diabetes]). Some specific examples of this include closed-loop insulin delivery systems have improved TIR [time-in-range] by between 11-15% compared to traditional insulin delivery regimens, comprehensive use of deep learning algorithms for screening for diabetic retinopathy with over 96% sensitivity, use of gradient-boosting algorithms to predict future complications with an AUROC (area under the receiver operating characteristic) of 0.91, and HbA1c (glycosylated hemoglobin) reductions of 0.7-1.1% in T2D patients through the application of coaching programs.

However, there are numerous critical issues that must be addressed before AI technology can be integrated into routine clinical practice: robust multicenter, multiethnic validation studies with long-term follow-up; development of comprehensive data governance and storage policies that are specific to AI applications; implementation of guidelines governing ethical and legal issues around algorithm transparency and accountability for AI applications; and systematic (in a deliberate, organized manner) inclusion of training related to digital and AI competency in medical training programs.

Achievement of all the above will enable AI to change the approach of diabetes management from an episodic, reactive mode to a more proactively focused and personalized preventive one. Diabetes patient care will be changed from crisis-oriented to an emphasis on personalized prevention, with the potential for improving patient outcomes through decreased diabetes-related risk factors, improved quality of life, and a more cost-effective manner for the health care system. The scientific basis is established; the challenge will be to provide equitable and responsible application to this population.

Supplementary Materials

Supporting Data Provided: The Supplemental Tables in Initial Table S1 contain detailed count and measurements of study quality, as per PRISMA's definitions; Tables S2 of the Supplemental Data provide a detailed list of inputs for each of the algorithm's components as well as who provided the source data for the algorithm to be trained on by whom in the training set, along with their associated clinical outcomes; Figure S1 of the Supplemental Data details the location of each study, along with the Digital Health Infrastructure Index at the local level for each of those studies.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Acknowledgments

We would like to thank all our colleagues at every participating center for their contributions to the achievement of this study. We would also like to extend our appreciation to the reviewers for their time and efforts in helping us to improve this study's methodological framework.

Conflicts of Interest

The authors declare that there is no conflict of interest. The funding sources, personal relationships, or affiliations have not impacted the writing and submission of this paper.