Managing diabetes is a tough challenge. The disease demands constant vigilance, flexible treatment plans that change with patients’ needs, and strategies to handle its emotional toll. Type 2 diabetes brings extra complications: patients must monitor blood sugar continuously, stick to their medications religiously, and cope with the stress that comes with a chronic illness. Success requires more than just fancy gadgets. We need to understand how patients think and act, educate them properly, and build strong support networks.

Blood glucose monitoring sits at the heart of the challenge. The old approach—checking blood sugar occasionally throughout the day (SMBG)—gives an incomplete picture at best. Think of it like trying to understand a movie by watching a few random scenes. Recent evidence shows that flash continuous glucose monitoring (fCGM) delivers much better results for hospitalized diabetes patients. Current guidelines suggest specific scanning frequencies: ten times daily for kids with any type of diabetes, adults with type 1, and pregnant women. Adults with type 2 need about seven scans daily. This frequency helps doctors assess how treatments, food choices, and physical activity affect glucose while preventing dangerous highs and lows.^[1]

Getting patients to actually monitor their blood sugar represents another uphill battle. Research reveals a sobering reality: just 57.6% of type 2 diabetes patients follow SMBG guidelines properly. Why such poor compliance? Several factors contribute—how long someone has had diabetes, whether they take oral medications, and crucially, their mindset. Many patients feel defeated by their condition and doubt their ability to manage it effectively.‌^[2] When patients skip monitoring, their blood sugar control deteriorates and complications become more likely. The mental health aspect compounds these problems. Interviews with Malaysian diabetes educators revealed that anxiety and depression significantly worsen adherence to both monitoring and treatment routines.^[3] We can’t separate the physical from the psychological when treating diabetes.

Treatment approaches must adapt because people respond differently to diabetes therapies. The traditional regimen—measuring glucose 3-4 times and injecting insulin 3-4 times daily—often misses the mark. Blood sugar doesn’t follow a predictable schedule; it fluctuates constantly throughout the day.^[4] Modern adaptive strategies leverage portable SMBG devices and continuous sensors to adjust insulin doses on the fly. This real-time approach has proven its worth, improving blood sugar control and reducing dangerous lows—both essential for successful diabetes management.^[5] Smart computer systems like the Diabetes Diagnostic Assistance System (DDAS) take this further by analyzing symptoms and test results to craft treatment plans tailored to each patient’s changing needs.^[6]

Healthcare providers play an irreplaceable role in diabetes care. Without proper education and support, patients struggle to manage their condition effectively. The statistics paint a concerning picture: merely 30.9% of type 2 diabetes patients achieve good blood sugar control. This gap underscores why we desperately need better diabetes education programs that empower patients to take charge of their health.^[3] COVID-19 exposed and amplified existing weaknesses in diabetes care delivery. The pandemic disrupted regular medical appointments, creating dangerous gaps in monitoring and treatment. Hospitalized COVID patients with diabetes faced a double burden—altered care protocols combined with blood sugar control problems that worsened their overall health.^[7]

This crisis demonstrated that healthcare systems need more resilience. While COVID-19 infected over 12 million people and killed more than 500,000 worldwide, we still had to maintain essential diabetes services.^[7] Remote monitoring emerged as a lifeline during this period. The IDEATel study offers compelling evidence: following 1,665 diabetes patients over five years, remote monitoring improved HbA1c, cholesterol, and blood pressure. This success provides a blueprint for tracking patients remotely and stepping in quickly when problems arise.^[8]

The pandemic transformed telemedicine from a nice-to-have option into an essential service for diabetes patients. What started as a workaround during lockdowns has proven itself to be a powerful tool for delivering care, improving health outcomes, and keeping patients connected to their healthcare teams. Let’s examine what the research tells us about telemedicine’s strengths, weaknesses, and potential future in diabetes care.

COVID-19 forced healthcare to adapt rapidly. Telemedicine allowed doctors to keep treating patients while everyone stayed safe from the virus. The numbers speak volumes: when researchers surveyed 958 type 2 diabetes patients, an impressive 91.43% said they were satisfied with telemedicine services during the pandemic. This overwhelming approval shows patients were ready to embrace virtual care.^[9] But satisfaction alone doesn’t tell the whole story. Ma and colleagues conducted a major review that proved telemedicine delivers real health benefits. After tracking patients for 12 months, they found HbA1c levels—the key marker for long-term blood sugar control—dropped by an average of 0.84 percentage points.^[10] In another trial, Sood’s team compared 199 patients using telemedicine against 83 receiving standard care. The telemedicine group saw their HbA1c fall by 1.01%, while the control group dropped just 0.68%. Though this difference wasn’t quite statistically significant (p=0.19), patient satisfaction told a different story—61.2% of telemedicine users were “very satisfied” versus only 40.8% in the standard care group.^[11] These results suggest telemedicine matches traditional care clinically while making patients happier (Fig. 1).

A comprehensive review of clinical trials painted a consistent picture. Telemedicine consistently lowered HbA1c levels across different timeframes: 0.57% reduction within 3 months, 0.28% at 4-12 months, and 0.26% beyond 12 months. These improvements might seem modest, but they matter. Telemedicine doesn’t just help with monitoring—it gets patients more involved in their own care, which proves crucial for managing any chronic disease.^[12]

Yet telemedicine faces real obstacles. Researchers identified 33 different barriers blocking widespread adoption. The top three? Staff struggling with technology (11%), resistance to change (8%), and worries about costs (8%). Additional roadblocks include insurance coverage, patient age, and education levels. These challenges hit hardest in poorer communities where internet access and tech skills remain limited.^[13] Still, success stories offer hope. During the pandemic, Fundación Valle del Lili’s telemedicine program served 663 diabetes patients, proving that virtual care can bridge healthcare gaps for people with chronic conditions who need ongoing support.^[14]

Adding wearable devices to telemedicine creates even better results. When researchers combined wearables with patient education, they saw remarkable improvements. Blood glucose dropped significantly (mean difference =–32.39; p<0.0001), and patients lost over 2% of their body weight in just three months. Real-time health tracking through wearables helps patients stick to their treatment plans.^[15] The data confirms what many suspected—wearables boost telemedicine effectiveness by keeping patients engaged between virtual visits.^[15]

Money matters too. Early diagnosis and treatment of type 2 diabetes could save billions. In the US alone, diabetes costs ballooned from 188 billion in 2012 to 237 billion by 2017. Telemedicine’s cost-effectiveness becomes especially important for chronic diseases requiring lifelong care.^[16] A study examining wearables for diabetes management found encouraging results after three months: over 2% weight loss, lower blood glucose, and reduced HbA1c. Interestingly, patient adherence didn’t change significantly, suggesting the benefits come from better monitoring rather than increased compliance.^[16]

AI is revolutionizing how we detect, predict, and treat diabetes. By enhancing screening accuracy and enabling earlier intervention, AI helps doctors provide targeted treatments that significantly reduce the severe health problems that come from years of high blood sugar.^[17] The technology extends beyond diagnosis—AI algorithms in insulin pumps can smooth out glucose swings, directly improving patients’ daily lives. With diabetes cases projected to hit 643 million by 2030 and 783 million by 2045, AI tools for predicting and diagnosing complications like eye and nerve damage have become indispensable.^[17]

Deep learning shines brightest in detecting diabetic retinopathy (DR), the leading cause of blindness in diabetes patients. When researchers tested AI against human eye doctors using modified Davis grading, the AI system achieved a PABAK score of 0.64 with 81% accuracy. This meant AI could actually outperform traditional human grading when analyzing retinal images.^[18] A review of 40 research studies confirmed these findings—deep learning consistently achieved excellent results across various diabetes tasks, particularly in spotting eye disease.^[19] The technology’s ability to process complex image data enables faster, more accurate screening than ever before.

AI also personalizes diabetes care through mobile health apps. These apps analyze real-time data—blood sugar readings, meal information, exercise patterns—to provide custom recommendations for the 451 million people living with diabetes worldwide.^[20] Beyond giving advice, these apps enable doctors to monitor patients remotely and intervene promptly when needed. This proves especially valuable for chronic conditions requiring constant oversight.^[21]AI excels at finding hidden patterns in vast datasets that human providers might miss, leading to more precise, individualized treatment plans.^[22]

The FDA’s approval of autonomous AI for eye screening marked a watershed moment. This system detects serious diabetic retinopathy and macular edema with 87.2% sensitivity and 90.7% specificity—all without human involvement.^{[23, 24]} Commercial systems like EyeArt perform even better, achieving 91.3% sensitivity and 91.1% specificity, making them invaluable for catching problems early.^{[25, 26]} The impact has been dramatic. Autonomous AI screening achieved 97.5% diagnostic accuracy and boosted detection of serious retinopathy to 85.7%. Most impressively, it increased screening compliance from 49% to 95% among young diabetes patients, especially helping communities with limited access to eye specialists.^{[27, 28]} This technology proves particularly vital for high-risk populations like American Indians and Alaska Natives, who face a 14.7% diabetes rate—higher than other US groups—and death rates 3.2 times the national average.^[29]

AI doesn’t stop at diagnosis—it optimizes treatment too. Algorithms predict complications like kidney and nerve damage with an AUROC of 0.85, helping doctors adjust treatments proactively.^{[17, 22]} Using data from 27,904 type 2 diabetes patients, a new Treatment Pathway Graph system lets AI customize medication plans by considering individual patient factors and predicting which treatments will work best.^[30] This personalization proves essential since patients respond so differently to diabetes medications.

Several hurdles remain. Data comes in different formats from various sources, making integration difficult. AI devices must prove their reliability. Patients worry about privacy, and some lack motivation to use new technology.^{[20, 31]} Li and colleagues emphasized that standardizing data formats and combining different information sources will be key to making AI more effective in diabetes care.^[32] Trust matters too. While patients appreciated AI’s efficiency for analyzing eye images via telemedicine, many still wanted human doctors involved in the process.^[33] Building confidence through education and transparency will be essential for widespread adoption.

Looking ahead, AI’s ability to analyze continuous glucose monitor data—which records readings every 1-5 minutes—promises even more personalized care.^{[34, 35]} As these technologies mature, they’ll not only improve medical outcomes but enhance daily life for millions living with diabetes.

Diabetes care relies on two main types of AI: machine learning (ML) and deep learning (DL). Each brings unique strengths to different aspects of patient care.

Machine learning encompasses techniques like support vector machines (SVM), decision trees, and random forests. A major review covering 450 clinical studies revealed how extensively these tools are used to identify high-risk diabetes patients.^[36–38] SVMs, for instance, classify patients by complication risk with accuracy rates between 8% and 45% for type 2 diabetes, helping doctors focus on those who need help most.^[38] In Western China, researchers built an XGBoost model that achieved an impressive AUC of 0.9122 for predicting diabetes risk, demonstrating ML’s power for early intervention.^[39] Khalate and Sukeshkumar’s comprehensive review tested multiple algorithms—logistic regression, decision trees, random forests—and found consistently high accuracy, sensitivity, and specificity. Their work confirms ML’s potential to transform diagnosis and patient care.^[40]

ML also excels at preventing complications. Decision support systems using ML can predict dangerous low blood sugar episodes in type 1 diabetes with remarkable accuracy. This early warning system has dramatically reduced hypoglycemic events, making life safer and better for patients.‌^[41] In a practical example, combining continuous glucose monitoring with an ML-powered insulin advisory system reduced glucose variability significantly—coefficient of variation dropped from 0.36 to 0.33 (p=0.045). This smoother glucose control translates directly to better patient outcomes.^[42]

Deep learning takes a different approach, using neural networks to decode complex information like retinal images or continuous glucose data. An ensemble CNN model combining VGGNet and residual neural network architectures achieved stunning results: AUC of 0.973, sensitivity of 92.25%, and specificity of 89.04% for detecting serious diabetic eye disease.^{[43, 44]} The model learned from 76,370 retinal images from over 13,000 diabetes patients, enabling earlier diagnosis and treatment of vision-threatening complications.^[43]

Natural language processing (NLP) represents an emerging frontier. By understanding clinical notes and patient narratives, NLP helps create more personalized care. A review of 1,849 articles showed that adding NLP to diabetes self-management tools increases patient engagement and treatment adherence.^[37] Digital health technologies powered by AI promise even greater advances—helping prevent diabetes in high-risk groups and supporting patients who can’t make it to in-person appointments.^[45]

Predicting who will develop diabetes has become increasingly sophisticated thanks to machine learning. Researchers now use various algorithms—Logistic Regression, Decision Trees, Random Forest, XGBoost, and SVM—each offering different advantages for identifying future type 2 diabetes cases.

Logistic Regression remains popular because doctors can easily understand and explain its results. One study found that models using four blood tests plus body measurements achieved an auROC of 0.87, substantially outperforming traditional risk scores. The Finnish Diabetes Risk Score managed only 0.66, while the German version reached 0.73. This superior performance comes from considering multiple factors like age, BMI, and family history together.^{[46, 47]} Decision Trees and Random Forests handle more complex patterns in the data. These methods have been tested extensively on datasets like the PIMA Indians database, showing varying but generally strong accuracy in predicting diabetes risk.^{[48, 49]} XGBoost stands out for exceptional performance—one ensemble model achieved an AUC of 0.884, proving the power of advanced machine learning.^{[50, 51]}

Hidden Markov Models represent another breakthrough, predicting 8-year diabetes risk with 86.9% AROC. This beats the Framingham Risk Score’s 78.6% by a significant margin, showing ML’s superior ability to distinguish future diabetes cases.^{[52, 54]} However, accuracy isn’t everything. One model achieved 78.26% prediction accuracy with 21.74% cross-validation error. Since missing a diabetes diagnosis (false negative) can have serious consequences, models must balance catching all cases against avoiding false alarms.^{[47, 54]}

Quality datasets drive quality predictions. The Pima Indians Diabetes Database, containing records from 768 female patients, serves as a standard benchmark for testing algorithms like K-nearest neighbors.^[55] The Framingham Heart Study provides another crucial resource, helping identify risk factors through its long-term patient tracking. Their risk model estimates 8-year diabetes probability based on current health status.^[52] Using diverse datasets ensures models work across different populations.

Deep learning opens new possibilities, like predicting heart disease risk from eye photos. One remarkable study showed AI could determine age (within 3.26 years), sex (97% accuracy), smoking status (71% accuracy), and blood pressure (within 11.23 mmHg) just from retinal images. This matches human expert performance while revealing connections between diabetes and broader health risks.^[56] Researchers have also used deep learning to detect diabetic nerve damage. Testing the SqueezeNet architecture on 1,561 diabetes patients, they achieved an AUC of 0.8013 for identifying neuropathy from retinal images.^[57] Since nerve damage often goes undetected until severe, early identification proves crucial.

Gulshan’s team set the gold standard for AI eye screening. Their deep learning model detected serious diabetic retinopathy with 90.3% sensitivity and 98.1% specificity on one dataset, and 87.0% sensitivity with 98.5% specificity on another. This reliability makes AI screening practical for widespread use.^[58] The technology saves healthcare systems money while catching problems earlier.

Beyond images, AI predicts diabetes from electronic health records too. Ravaut’s team built a gradient boosting model that achieved an AUC of 80.26 using administrative data. Remarkably, the top 5% highest-risk patients identified by the model accounted for 26% of Ontario’s total diabetes costs. This enables smarter resource allocation and targeted prevention.^[59] Choi’s group took a different angle, predicting new diabetes cases in heart disease patients. Their model achieved an AUC of 0.78 over 5 years, revealing how cardiovascular and diabetes risks intertwine.^[60]

Retinal imaging’s accessibility makes it perfect for widespread screening. A model detecting early kidney problems from eye photos achieved an AUC of 0.81, enabling quick specialist referrals when needed.^[61] Bora’s team created algorithms that predict future diabetic retinopathy risk with AUCs of 0.79 (internal) and 0.70 (external validation), helping prevent vision loss through timely intervention.^[62] Different imaging techniques expand screening options. Liu’s group used OCT images with a lesion-aware attention network, achieving 91.68% accuracy for detecting diabetic kidney disease.^[63] Sridevi showed how combining CNNs with telemedicine platforms improves remote diagnosis accuracy, particularly benefiting underserved areas.^[64] Perhaps most impressively, Sabanayagam’s algorithm detects chronic kidney disease from retinal photos with an AUC of 0.911, demonstrating how one image can reveal multiple health risks.^[65]

NLP unlocks valuable insights hidden in medical records, clinical notes, and other text sources. This technology promises to transform diabetes prediction by making sense of unstructured information that traditional analysis misses.

The sheer volume of text in healthcare requires automated processing. NLP converts messy clinical notes into organized data that machine learning can analyze. One review found this shift from manual to automated methods has accelerated dramatically, reflecting NLP’s growing importance in chronic disease research.^[66] By extracting key concepts from discharge summaries and nursing notes, NLP reveals patient histories and risk factors that might otherwise go unnoticed.

Real-world applications prove NLP’s worth. Researchers built an alcohol misuse classifier that achieved 0.78 AUC with 56% sensitivity from electronic records. Since lifestyle factors like drinking significantly impact diabetes risk, this demonstrates how NLP can predict health outcomes from clinical narratives.^[67] Even more impressive, Schwartz’s team created an NLP algorithm that identifies prediabetes discussions with 98% precision and recall. Catching prediabetes early allows intervention before full diabetes develops.^[68]

NLP also helps predict serious outcomes. Ye’s group combined clinical notes with machine learning to predict mortality in critically ill diabetes patients, achieving an outstanding 0.97 AUC. This shows NLP’s potential for handling complex medical scenarios where multiple factors interact.^[69] By understanding not just what doctors write but what patients say, NLP provides a more complete picture of diabetes management challenges.

Social media represents an untapped data source. A review of 87 studies showed NLP can extract real-world diabetes insights from online patient discussions, complementing traditional clinical data.^[70] However, accuracy remains paramount. Juhn and Liu stressed that NLP must reliably extract information from health records to build trustworthy predictive models.^[71] For remote care, NLP shows particular promise. Tahayori demonstrated 83% accuracy in predicting patient outcomes with 0.88 AUC, suggesting NLP could enhance telemedicine and enable timely interventions for diabetes patients.^[72]

Combining deep learning and NLP in clinical decision support systems represents healthcare’s cutting edge. These technologies work together to predict diabetes onset, guide patient care, and improve outcomes through intelligent data analysis.

Deep learning excels at finding patterns in complex medical records. Rajkomar’s team showed this powerfully—their neural networks predicted hospital events including mortality with 0.93-0.94 AUROC. Applied to diabetes, such systems could alert doctors to risks and suggest interventions based on each patient’s unique data.‌^[73] Meanwhile, NLP extracts meaning from clinical notes that would otherwise remain buried. A major review found machine learning methods now dominate NLP research, with these systems successfully converting unstructured text into actionable insights. For diabetes care, this means understanding complete patient stories—not just lab values—to guide better treatment decisions.^[66] When personalized properly, these approaches dramatically improve how engaged patients feel and how well they follow treatment plans.

The combination of telemedicine, AI, and predictive modeling is fundamentally changing diabetes care. These technologies create new possibilities for helping patients, encouraging self-care, and achieving better health outcomes. Yet alongside clear benefits come real challenges around implementation and acceptance that we must address.

Telemedicine has proven itself to be an essential tool for remote diabetes management. Cheng and Kao’s research provides compelling evidence—adding telehealth consultations slashed post-meal glucose swings from 169 mg/dL to 111 mg/dL in newly diagnosed patients (p<0.001).‌^[74] When COVID-19 struck, telemedicine became a lifeline. Sotomayor’s review of 317 studies found it successfully maintained diabetes care during lockdowns, though results varied by implementation.^[75] Patients embraced virtual care enthusiastically—91.43% reported satisfaction, with most finding it convenient and cost-effective.^[9]

AI brings complementary strengths. Smart algorithms analyze patient patterns to predict health trajectories and personalize treatments. Evidence shows AI-enhanced telemedicine reduces HbA1c by 0.37% to 0.71% compared to standard care.^[76] Predictive models identify at-risk patients before complications develop, enabling preventive intervention rather than reactive treatment.

However, significant obstacles remain. Technology access varies wildly across communities. Rural and low-income patients may still need face-to-face visits for comprehensive care, even as telemedicine helps with routine monitoring.^[77] This digital divide threatens to worsen health disparities unless we ensure equitable access to virtual care tools.

Healthcare providers show mixed feelings too. While doctors (4.06±0.69) and nurses (4.02±0.61) generally support telemedicine, concerns persist about integrating it with other duties and its practical usefulness.^[78] Shifting policies and payment structures add complexity to adoption. Education proves vital—when implemented well, telemedicine measurably improves patient self-care skills, with DSMQ scores rising from 6.79 to 7.32 (p=0.0015).^[79] Yet without proper training for patients and providers alike, we can’t realize telemedicine’s full potential.

AI technologies raise additional concerns. As healthcare becomes increasingly data-driven, patients worry about privacy and security of their sensitive information. These fears could slow AI adoption unless we establish strong safeguards and clear ethical guidelines that protect patient welfare while enabling innovation.

Moving forward, success requires weaving together telemedicine, AI, and predictive modeling into a coherent system. Healthcare professionals, policymakers, and technology developers must collaborate to build frameworks that maximize benefits while addressing limitations. Only through sustained research, development, and patient-centered design can we fully harness these transformative tools for diabetes care. The technologies exist—now we must ensure they reach everyone who needs them, work reliably in real-world settings, and genuinely improve life for the millions living with diabetes worldwide.

The author is the principal investigator in a project of the Medical University of Plovdiv in Bulgaria titled “Opportunities of telemedicine and generative artificial intelligence for a holistic approach in the management of type 2 diabetes”.

The authors have no funding to report.

The authors have declared that no competing interests exist.