Artificial Intelligence in Non-Insulin-Treated Type 2 Diabetes: From Reactive Management to Anticipatory Care

Abstract

Type 2 diabetes mellitus is a highly prevalent, heterogeneous, and progressive chronic disease. In a large proportion of patients, management is based for many years on lifestyle intervention and non-insulin glucose-lowering therapies. This long pre-insulin phase represents a crucial clinical window, in which timely recognition of metabolic deterioration, therapeutic inertia, treatment response, and individual risk trajectories may substantially influence long-term outcomes. However, routine care is still frequently based on intermittent assessments, delayed treatment adaptation, and limited integration of clinical, biochemical, behavioral, and digital data. Artificial intelligence may offer a clinically relevant opportunity to move from reactive management to anticipatory care in non-insulin-treated type 2 diabetes. Rather than replacing clinical judgment or automating treatment decisions, artificial intelligence can support clinicians by identifying hidden patterns, predicting metabolic worsening, stratifying risk, improving the interpretation of glucose data, and personalizing follow-up intensity and therapeutic timing. In this setting, its most meaningful role may be to reduce the silent interval between early deterioration and clinical action. This narrative review discusses the rationale, current applications, near-future scenarios, and implementation barriers of artificial intelligence in non-insulin-treated type 2 diabetes. Particular attention is given to advanced interpretation of glycemic data, clinical decision support, prediction of treatment failure, remote monitoring, and the potential integration of multidimensional data into more precise and timely care pathways. The review also emphasizes the need for explainable, clinically validated, equitable, and ethically governed artificial intelligence tools that can be realistically embedded into everyday diabetology practice.

1. Introduction

Type 2 diabetes mellitus (T2D) is not a static condition, but a progressive and heterogeneous disease in which metabolic control, cardiovascular risk, renal function, body weight, liver involvement, adherence, and therapeutic response may evolve along markedly different trajectories [1,2,3,4,5,6]. For many patients, the largest part of the disease course occurs before insulin initiation, during a phase managed with lifestyle intervention and non-insulin glucose-lowering agents [4,5]. This period is clinically decisive: therapeutic intensification may be delayed, deterioration may remain partially hidden, and opportunities for earlier prevention may be missed [7,8].

Despite major advances in pharmacological treatment, everyday management of non-insulin-treated T2D remains frequently episodic. Decisions are commonly based on periodic visits, isolated laboratory values, and retrospective interpretation of glycated hemoglobin. Although HbA1c remains a central marker of medium-term glycemic exposure, it cannot fully capture short-term fluctuations, post-prandial excursions, glycemic variability, treatment adherence, behavioral patterns, or the dynamic interaction between metabolic control and cardiometabolic risk [2,3,4,5,9,10]. As a result, the clinician may recognize therapeutic failure only after it has become established.

This limitation is particularly relevant in a context increasingly characterized by data abundance but interpretative fragmentation. Patients may generate information through self-monitoring of blood glucose, intermittent or professional continuous glucose monitoring (CGM), body weight tracking, physical activity records, digital platforms, laboratory tests, prescription histories, and clinical notes [3,9,10,11,12,13]. However, these data often remain disconnected, underused, or interpreted only at widely spaced time points. The gap between available information and actionable clinical insight contributes to therapeutic inertia and to a model of care that is still predominantly reactive [7,8,14,15,16].

Artificial intelligence (AI) has the potential to address this gap. In non-insulin-treated T2D, its most promising function is not to replace the diabetologist or to automate therapeutic decisions, but to enhance the clinician’s ability to recognize clinically meaningful patterns before deterioration becomes evident [14,15,16]. By integrating multidimensional data, AI may help identify patients at higher risk of diabetes-related complications [17,18] and recognize patterns associated with delayed treatment intensification [19]; the projected consequences of intervening earlier have been modelled in simulation [20], and the broader scope of these applications has been reviewed elsewhere [14,15,16].

The conceptual shift proposed in this review is therefore from reactive management to anticipatory care, a distinction developed in Section 4. It is especially important in non-insulin-treated T2D, where treatment decisions are rarely urgent in the immediate sense, but often decisive over the medium and long term.

2. Methods: Literature Search Strategy and Article Selection

This article is a narrative review. It is not a systematic review and was not designed as one: no protocol was registered, no formal screening was undertaken, and no count of records identified, screened, or excluded is reported. The aim was to build a clinically oriented framework for the use of AI in the large and still underexplored population of patients with T2D not treated with insulin, rather than to produce a systematic evidence synthesis. The manuscript was prepared with reference to SANRA, the Scale for the Assessment of Narrative Review Articles, which covers justification of the importance of the topic, statement of aims, description of the literature search, referencing, scientific reasoning, and presentation of data [21]. A pragmatic literature search was conducted using PubMed/MEDLINE, Scopus, and Google Scholar, combining terms related to the disease area and digital methodology, including: “type 2 diabetes”, “non-insulin-treated”, “artificial intelligence”, “machine learning”, “clinical decision support”, “therapeutic inertia”, “continuous glucose monitoring”, “digital health”, “remote monitoring”, “prediction”, “precision medicine”, “explainability”, and “AI governance”.

No lower date limit was applied. Searches were run up to May 2026 and were updated in July 2026 during revision, which accounts for the inclusion of literature published after initial submission. Only articles published in English were considered; this restriction may have excluded relevant work published in other languages.

Priority was given to international guidelines, consensus statements, systematic or narrative reviews, clinically relevant observational studies, randomized trials, and methodological papers addressing AI implementation in health care. Articles primarily focused on type 1 diabetes, automated insulin delivery, closed-loop algorithms, or insulin dose automation were considered only when they provided relevant conceptual background. Because selection was purposive rather than exhaustive, the review is open to selection bias, and the literature presented should be read as illustrative of the current state of the field rather than as a complete enumeration of it.

3. Why Non-Insulin-Treated Type 2 Diabetes Is a Key Scenario for Artificial Intelligence

The application of AI to diabetes has historically been dominated by insulin-treated disease, particularly type 1 diabetes, where automated insulin delivery, closed-loop systems, sensor-augmented pumps, and algorithm-driven dose modulation represent visible examples of digital innovation [3,14,16]. This focus is clinically understandable because insulin therapy requires frequent short-term decisions and carries an immediate risk of hypoglycemia. However, it has also contributed to an underestimation of the potential role of AI in the much larger population of patients with T2D who are not treated with insulin [1,16].

Non-insulin-treated T2D represents a different but equally important challenge. In this setting, the main problem is usually not minute-by-minute dose adjustment, but timely recognition of progressive loss of control, insufficient response to therapy, increasing cardiometabolic risk, and the need for treatment intensification [4,5,6,7,8]. These processes are gradual, multifactorial, and difficult to capture through isolated clinical encounters. AI may therefore be particularly useful not as an automation engine, but as a longitudinal interpretative tool [14,15,16,17].

Several features make this population especially suitable for AI applications. First, non-insulin-treated T2D is highly heterogeneous. Patients differ widely in age, disease duration, obesity phenotype, insulin resistance, beta-cell reserve, renal function, liver involvement, cardiovascular risk, socioeconomic context, adherence, and therapeutic exposure [5,6]. Second, treatment pathways are increasingly complex, with multiple drug classes that differ not only in glucose-lowering efficacy but also in effects on body weight, cardiovascular outcomes, renal protection, tolerability, persistence, and patient preferences [4,5]. Third, the timing of intensification is often uncertain, and delayed treatment adaptation remains one of the most persistent problems in routine care [7,8].

The potential value of AI lies precisely in the fact that the clinical problem is not fully solved by existing care models. Non-insulin-treated T2D is common, longitudinal, data-rich, and exposed to inertia. It is therefore an ideal setting for tools designed to improve timing, prioritization, and personalization of care.

It should be said plainly that this rationale is at present stronger than the evidence supporting it. Heterogeneity, longitudinal data density, and exposure to therapeutic inertia are structural features of this population; they establish that it is a plausible setting for AI, not that AI has been shown to help in it. As detailed in Section 5, direct evidence of clinical benefit from AI in non-insulin-treated T2D is currently limited, and the studies that come closest are retrospective, internally validated, and centered on recognizing inertia that has already occurred rather than on anticipating deterioration that has not. The framework proposed in this review should therefore be read as a research agenda rather than as a description of an established capability: an argument for where investigative effort is warranted, and a set of claims against which the field can subsequently be assessed.

4. From Reactive Management to Anticipatory Care

Reactive management remains deeply embedded in routine T2D care. In this model, patients are reassessed at predefined intervals, therapeutic decisions are made after laboratory deterioration is documented, and treatment intensification often occurs only when glycemic failure has become evident [7,8]. This approach is clinically familiar and operationally simple, but it may be poorly aligned with the progressive and dynamic nature of T2D.

In non-insulin-treated disease, deterioration often develops silently. A patient may remain apparently acceptable for months while fasting glucose rises, post-prandial excursions increase, weight changes, adherence weakens, or renal and cardiometabolic risk profiles worsen [5,6,9,10]. By the time HbA1c clearly exceeds target, the underlying trajectory may already have shifted. The consequence is not only delayed glycemic correction, but also a missed opportunity to protect the broader metabolic and vascular future of the patient.

Anticipatory care proposes a different logic. Rather than waiting for failure to become visible, it aims to identify early signals of unfavorable trajectory and to adapt care before the patient crosses a clinically meaningful threshold. This does not mean overtreatment or indiscriminate intensification. It means using available data more intelligently to distinguish stable patients from those who are beginning to drift, and to match the timing of intervention to the individual trajectory [14,15,16].

AI may support this transition by improving the continuity of interpretation. Traditional care often collects data in fragments; AI can help connect them over time. It may identify recurrent glycemic patterns, detect deviations from previous stability, combine biochemical and behavioral markers, and generate risk estimates that are updated as new information becomes available [14,15,16,18]. In this framework, the clinical question shifts from “What is the patient’s current HbA1c?” to “Where is this patient going, and how soon do we need to act?”

This framing overlaps with existing precision-medicine frameworks, and the relationship should be stated explicitly rather than left implicit. Precision diabetes, as set out in the consensus report of the American Diabetes Association and the European Association for the Study of Diabetes, is principally concerned with matching the right intervention to the right patient: it asks which treatment, on the basis of aetiology, phenotype, genotype, and predicted response, and its most developed applications concern subtype classification and monogenic forms of diabetes [6]. The broader P4 framework—predictive, preventive, personalized, and participatory—has likewise developed its predictive component largely around the onset of disease in people who are not yet ill.

Anticipatory care, as used in this review, asks a different question. It is concerned less with which treatment than with when to act, in patients whose diagnosis and treatment class are already established. Its unit of analysis is not the phenotype at a point in time but the trajectory over time, and its target is not the selection of therapy but the interval between the onset of deterioration and the clinical response to it. The two are complementary rather than competing: precision medicine addresses what should be done, anticipatory care addresses when the question should be asked. The contribution claimed here is correspondingly narrow. It is that this timing dimension has been comparatively neglected in the non-insulin-treated population, that it is in principle measurable, and that it is the aspect of care in which AI is most plausibly useful—a claim that, as Section 5 sets out, remains to be demonstrated.

This conceptual transition is summarized in Figure 1.

5. Current Applications of Artificial Intelligence in Non-Insulin-Treated Type 2 Diabetes

Glycemic assessment in non-insulin-treated T2D is still largely centered on HbA1c, which remains an essential marker for evaluating medium-term glycemic exposure and guiding therapeutic targets [2,4,5]. However, HbA1c provides an averaged estimate and does not fully reflect short-term glycemic dynamics, post-prandial excursions, intra-day variability, nocturnal patterns, or the temporal relationship between glucose profiles, meals, physical activity, adherence, and pharmacological exposure [9,10].

AI may help overcome part of this limitation by improving the interpretation of glucose data generated through SMBG, intermittent CGM, professional CGM, or connected digital devices [3,9,10,11,12,13]. Its value does not lie simply in collecting more data, but in transforming fragmented measurements into clinically interpretable patterns [14,15,16,17]. Algorithms may identify recurring post-prandial peaks, early morning dysglycemia, excessive variability, progressive upward drift, or discordance between HbA1c and daily glucose profiles [10,14,17].

A second current application concerns therapeutic personalization and treatment intensification. Modern treatment of T2D increasingly requires decisions that go beyond glucose lowering alone, including body weight, cardiovascular risk, kidney function, liver disease, tolerability, adherence, patient preferences, drug availability, and cost [4,5,6]. AI may support this process by integrating multidimensional data and identifying patient profiles associated with different risks, treatment responses, or probabilities of therapeutic failure [6,17,18,19,20].

A practical example is the patient with progressive HbA1c worsening on metformin, increasing body weight, early albuminuria, and declining eGFR. In such a case, an AI-based decision-support system could help identify a trajectory compatible with secondary treatment failure and highlight the potential relevance of earlier intensification with agents providing cardiometabolic and renal protection, such as SGLT2 inhibitors or GLP-1 receptor agonist/GIP-based therapies. Importantly, such a system should not select treatment autonomously, but should support the clinician in recognizing the therapeutic window before stable organ damage or prolonged metabolic deterioration occurs [4,5,6,19,20].

A particularly important application is the reduction of therapeutic inertia. In routine care, intensification is often delayed even when glycemic targets are not met [7,8]. AI may help make risk more visible by identifying patients whose data suggest progressive deterioration or insufficient response before failure becomes obvious [18,19,20]. The clinical usefulness of these tools will depend on their ability to provide interpretable and actionable information rather than opaque risk scores [22,23,24,25,26].

Remote monitoring and digital follow-up represent a third domain. Patients may generate data on glucose, weight, blood pressure, physical activity, sleep, medication use, and self-reported symptoms [3,11,12,13,14,15,16]. Without intelligent filtering, these data may create overload and alert fatigue [22,23,24,25]. AI can potentially prioritize information, detect deviations from previous stability, and stratify patients according to the need for earlier review [14,15,16,17].

Finally, prediction of metabolic deterioration is the conceptual core of anticipatory care. Potential predictive targets include worsening HbA1c, loss of glycemic control, need for treatment intensification, weight gain, declining adherence, reduced persistence, progression of kidney disease, increasing cardiovascular risk, and transition toward more complex therapeutic regimens [17,18,19,20].

Current Strength of the Evidence Base

The applications described above rest on evidence that is uneven in both directness and methodological maturity. Table 1 therefore states the maturity of each application explicitly, and Table 2 summarizes the studies cited in this review that report quantitative performance for AI applied to type 2 diabetes, together with their population, endpoint, validation strategy, and limitations. Three observations follow, and they qualify the claims made in this section.

Table 1. Current and near-future applications of AI in non-insulin-treated T2D. The table summarizes major clinical areas, the current maturity of the evidence supporting each, data sources, AI functions, potential benefits, and implementation risks.

Clinical Area	Evidence Maturity	Data Sources	AI Function	Potential Clinical Benefit	Main Limitations/Implementation Risks
Glycemic interpretation	Routine practice for standardised CGM metric interpretation [9,10]; AI-driven pattern detection remains emerging	HbA1c, SMBG, CGM, meal/activity records	Pattern recognition, variability analysis, discordance detection	Better identification of hidden instability and early drift	Poor data quality, missing values, uncertain thresholds, limited access to CGM, digital visibility bias
Therapeutic personalization	Emerging evidence; retrospective observational support only [19,20]	Clinical history, labs, medications, comorbidities, adherence	Risk phenotyping, response prediction, scenario support	Earlier and more individualized intensification	Confounding by indication, limited external validation, black-box outputs requiring XAI approaches such as SHAP/LIME for clinical trust, unclear action thresholds
Remote follow-up	Emerging evidence for CGM-based follow-up; AI-based alert prioritisation investigational	Glucose, weight, BP, activity, symptoms, prescription data	Trend detection, alert prioritization, adaptive monitoring	Earlier contact for patients who are deteriorating	Alert fatigue, workflow burden, unclear responsibility, lack of interoperability, unequal digital access
Progression prediction	Investigational; no prospective validation in this population	Longitudinal clinical, biochemical, renal and behavioral data	Dynamic risk models	Prediction of loss of control, treatment failure, and complications	Bias, MNAR data, digital visibility bias, model drift, poor transportability, uncertain clinical actionability
Generative AI support	Investigational; no clinical outcome data	Clinical notes, education material, patient-reported information	Summarization, communication support, visit preparation	Reduced administrative burden and improved education	Hallucinations, plausible but incorrect outputs, need for human supervision, medico-legal uncertainty, risk of unsupervised therapeutic advice

BP, blood pressure; CGM, continuous glucose monitoring; HbA1c, glycated haemoglobin; LIME, Local Interpretable Model-agnostic Explanations, a method that approximates a complex model locally in order to explain a single prediction; MNAR, missing not at random, i.e., missingness that is itself related to the unobserved value; SHAP, SHapley Additive exPlanations, a method that attributes a prediction to the contribution of each input variable; SMBG, self-monitoring of blood glucose; T2D, type 2 diabetes; XAI, explainable artificial intelligence.

Table 1. Current and near-future applications of AI in non-insulin-treated T2D. The table summarizes major clinical areas, the current maturity of the evidence supporting each, data sources, AI functions, potential benefits, and implementation risks.

Clinical Area	Evidence Maturity	Data Sources	AI Function	Potential Clinical Benefit	Main Limitations/Implementation Risks
Glycemic interpretation	Routine practice for standardised CGM metric interpretation [9,10]; AI-driven pattern detection remains emerging	HbA1c, SMBG, CGM, meal/activity records	Pattern recognition, variability analysis, discordance detection	Better identification of hidden instability and early drift	Poor data quality, missing values, uncertain thresholds, limited access to CGM, digital visibility bias
Therapeutic personalization	Emerging evidence; retrospective observational support only [19,20]	Clinical history, labs, medications, comorbidities, adherence	Risk phenotyping, response prediction, scenario support	Earlier and more individualized intensification	Confounding by indication, limited external validation, black-box outputs requiring XAI approaches such as SHAP/LIME for clinical trust, unclear action thresholds
Remote follow-up	Emerging evidence for CGM-based follow-up; AI-based alert prioritisation investigational	Glucose, weight, BP, activity, symptoms, prescription data	Trend detection, alert prioritization, adaptive monitoring	Earlier contact for patients who are deteriorating	Alert fatigue, workflow burden, unclear responsibility, lack of interoperability, unequal digital access
Progression prediction	Investigational; no prospective validation in this population	Longitudinal clinical, biochemical, renal and behavioral data	Dynamic risk models	Prediction of loss of control, treatment failure, and complications	Bias, MNAR data, digital visibility bias, model drift, poor transportability, uncertain clinical actionability
Generative AI support	Investigational; no clinical outcome data	Clinical notes, education material, patient-reported information	Summarization, communication support, visit preparation	Reduced administrative burden and improved education	Hallucinations, plausible but incorrect outputs, need for human supervision, medico-legal uncertainty, risk of unsupervised therapeutic advice

BP, blood pressure; CGM, continuous glucose monitoring; HbA1c, glycated haemoglobin; LIME, Local Interpretable Model-agnostic Explanations, a method that approximates a complex model locally in order to explain a single prediction; MNAR, missing not at random, i.e., missingness that is itself related to the unobserved value; SHAP, SHapley Additive exPlanations, a method that attributes a prediction to the contribution of each input variable; SMBG, self-monitoring of blood glucose; T2D, type 2 diabetes; XAI, explainable artificial intelligence.

Table 2. Representative studies of artificial intelligence applied to type 2 diabetes cited in this review. Studies are ordered by directness of relevance to the non-insulin-treated population. Performance metrics are reported as published; where calibration or external validation was not reported, this is stated explicitly rather than left blank.

Study	Population, Setting and Sample Size	AI Approach	Clinical Endpoint	Validation Strategy and Reported Performance	Direct Relevance to Non-Insulin-Treated T2D	Main Limitations
Musacchio et al., 2024 Int J Med Inform [19]	Adults with T2D on metformin monotherapy and two consecutive mean HbA1c > 7.0% (>53 mmol/mol). AMD Annals database, 271 Italian diabetes clinics, 2005–2019, drawn from ~1.5 million records. Inertia-NO n = 20,067; inertia-YES n = 13,029.	Logic Learning Machine (Rulex), a rule-based “clear box” model generating explicit if-then rules and variable thresholds without a post hoc explainability layer.	Presence versus absence of therapeutic inertia (failure to intensify despite persistent above-target HbA1c).	Internal split (70% learning/30% test). No external validation. ROC-AUC 0.81; accuracy 0.71; precision 0.80; recall 0.71; F1 0.75. Calibration not reported. Identified two distinct HbA1c patterns: a modest rise with high variability (SD > +0.57%) was more strongly associated with inertia than a large rise.	Direct. The only study identified that is restricted to non-insulin-treated T2D on oral monotherapy.	Internal validation only; calibration not reported. Classifies inertia that has already occurred rather than predicting future deterioration. Italian specialist clinics only, limiting transportability to primary care. Data window (2005–2019) precedes widespread SGLT2i and GLP-1RA use.
Ravizza et al., 2019 Nat Med [18]	Adults with diabetes and no CKD at baseline. Two US real-world EHR sources: IBM Explorys (n = 417,912, feature selection) and Indiana Network for Patient Care (n = 82,912, model development).	Random forest using seven data-driven selected routine variables (age, BMI, eGFR, creatinine, glucose, albumin, HbA1c), compared with logistic regression.	Incident chronic kidney disease.	Model derived and tested across two independent real-world databases; no prospective validation. AUC 0.833 (random forest) versus 0.827 (logistic regression). Reported to outperform published trial-derived algorithms on case-by-case comparison. Calibration not reported.	Indirect. Mixed diabetes population, not restricted by treatment modality. Renal endpoint, not therapeutic failure.	Real-world data completeness and coding heterogeneity. Calibration not reported. Marginal gain over logistic regression (delta-AUC 0.006). Predicts complication onset rather than treatment response or timing of intensification.
Musacchio et al., 2024 Mach Learn Knowl Extr [20]	Adults with T2D on dual or triple therapy with HbA1c above threshold at two consecutive visits. AMD Annals (1,186,247 patients, 2005–2019); analysed cohort n = 85,239 (20,015 modelling; 65,224 simulation).	Logic Learning Machine combined with a counterfactual “what-if” scenario simulation of timely insulin initiation.	Attainment of HbA1c target at 12 months following (simulated) timely insulin initiation.	Internal split. No external validation. Simulated outcomes are model-projected, not observed. AUC 0.704, described by the authors as modest; accuracy 0.64; specificity 0.65; recall 0.61. Simulation projected a 17% relative increase in patients reaching target.	Marginal. The modelled decision is insulin initiation, i.e., the exit point from the non-insulin-treated population.	Modest discrimination, acknowledged by the authors as reflecting reliance on baseline variables alone. Counterfactual projections are not prospectively verified. Industry-funded. Data window precedes current pharmacological options.
Fan et al., 2021 Front Pharmacol [17]	165 adults with medication-nonadherent T2D (of 800 screened), single-centre inpatient records, Sichuan, China, 2010–2015. 129 (78%) had HbA1c ≥ 7%.	Seven algorithm families used to build 18 prediction models; the best-performing model was selected for each outcome.	Onset of diabetic nephropathy, peripheral neuropathy, angiopathy and eye disease; glycaemic control status.	Internal test set only. No external validation. Best test-set AUC: nephropathy 0.902 +/− 0.040; angiopathy 0.889 +/− 0.059; neuropathy 0.859 +/− 0.050; eye disease 0.832 +/− 0.086; HbA1c 0.825 +/− 0.092. Calibration not reported.	Indirect. Nonadherent T2D overlaps conceptually, but the setting is single-centre and inpatient.	Sample size (n = 165) is small relative to the number of models fitted and to the high reported AUCs, making overfitting likely. Selecting the best-performing model per outcome further inflates apparent performance. No external validation; no calibration; single centre.
Kopitar et al., 2020 Sci Rep [27]	3723 adults with no prior diagnosis of diabetes attending preventive health examinations, 10 Slovenian primary care centres, 2014–2017.	Glmnet, random forest, XGBoost and LightGBM compared against a regression baseline. Reported against TRIPOD.	Fasting plasma glucose level and detection of impaired fasting glucose/undiagnosed T2D.	100 bootstrap iterations (internal). Calibration explicitly assessed. Best AUC 0.859 (Glmnet) versus 0.854 (regression baseline). R2 0.26–0.36 across models, indicating weak calibration. The authors conclude that machine learning conferred no clinically relevant advantage over regression.	None. Screening for undiagnosed diabetes in a population without diabetes; outside the target population of this review. Cited for methodological context only (Section 7).	Single database; high proportion of missing values. Working-age screening population underrepresents older adults, in whom T2D is most prevalent. Findings concern case-finding, not longitudinal management.

AMD, Associazione Medici Diabetologi; AUC, area under the receiver-operating characteristic curve; BMI, body mass index; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; EHR, electronic health record; GLP-1RA, glucagon-like peptide-1 receptor agonist; SGLT2i, sodium-glucose cotransporter-2 inhibitor; T2D, type 2 diabetes; TRIPOD, Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis.

Table 2. Representative studies of artificial intelligence applied to type 2 diabetes cited in this review. Studies are ordered by directness of relevance to the non-insulin-treated population. Performance metrics are reported as published; where calibration or external validation was not reported, this is stated explicitly rather than left blank.

Study	Population, Setting and Sample Size	AI Approach	Clinical Endpoint	Validation Strategy and Reported Performance	Direct Relevance to Non-Insulin-Treated T2D	Main Limitations
Musacchio et al., 2024 Int J Med Inform [19]	Adults with T2D on metformin monotherapy and two consecutive mean HbA1c > 7.0% (>53 mmol/mol). AMD Annals database, 271 Italian diabetes clinics, 2005–2019, drawn from ~1.5 million records. Inertia-NO n = 20,067; inertia-YES n = 13,029.	Logic Learning Machine (Rulex), a rule-based “clear box” model generating explicit if-then rules and variable thresholds without a post hoc explainability layer.	Presence versus absence of therapeutic inertia (failure to intensify despite persistent above-target HbA1c).	Internal split (70% learning/30% test). No external validation. ROC-AUC 0.81; accuracy 0.71; precision 0.80; recall 0.71; F1 0.75. Calibration not reported. Identified two distinct HbA1c patterns: a modest rise with high variability (SD > +0.57%) was more strongly associated with inertia than a large rise.	Direct. The only study identified that is restricted to non-insulin-treated T2D on oral monotherapy.	Internal validation only; calibration not reported. Classifies inertia that has already occurred rather than predicting future deterioration. Italian specialist clinics only, limiting transportability to primary care. Data window (2005–2019) precedes widespread SGLT2i and GLP-1RA use.
Ravizza et al., 2019 Nat Med [18]	Adults with diabetes and no CKD at baseline. Two US real-world EHR sources: IBM Explorys (n = 417,912, feature selection) and Indiana Network for Patient Care (n = 82,912, model development).	Random forest using seven data-driven selected routine variables (age, BMI, eGFR, creatinine, glucose, albumin, HbA1c), compared with logistic regression.	Incident chronic kidney disease.	Model derived and tested across two independent real-world databases; no prospective validation. AUC 0.833 (random forest) versus 0.827 (logistic regression). Reported to outperform published trial-derived algorithms on case-by-case comparison. Calibration not reported.	Indirect. Mixed diabetes population, not restricted by treatment modality. Renal endpoint, not therapeutic failure.	Real-world data completeness and coding heterogeneity. Calibration not reported. Marginal gain over logistic regression (delta-AUC 0.006). Predicts complication onset rather than treatment response or timing of intensification.
Musacchio et al., 2024 Mach Learn Knowl Extr [20]	Adults with T2D on dual or triple therapy with HbA1c above threshold at two consecutive visits. AMD Annals (1,186,247 patients, 2005–2019); analysed cohort n = 85,239 (20,015 modelling; 65,224 simulation).	Logic Learning Machine combined with a counterfactual “what-if” scenario simulation of timely insulin initiation.	Attainment of HbA1c target at 12 months following (simulated) timely insulin initiation.	Internal split. No external validation. Simulated outcomes are model-projected, not observed. AUC 0.704, described by the authors as modest; accuracy 0.64; specificity 0.65; recall 0.61. Simulation projected a 17% relative increase in patients reaching target.	Marginal. The modelled decision is insulin initiation, i.e., the exit point from the non-insulin-treated population.	Modest discrimination, acknowledged by the authors as reflecting reliance on baseline variables alone. Counterfactual projections are not prospectively verified. Industry-funded. Data window precedes current pharmacological options.
Fan et al., 2021 Front Pharmacol [17]	165 adults with medication-nonadherent T2D (of 800 screened), single-centre inpatient records, Sichuan, China, 2010–2015. 129 (78%) had HbA1c ≥ 7%.	Seven algorithm families used to build 18 prediction models; the best-performing model was selected for each outcome.	Onset of diabetic nephropathy, peripheral neuropathy, angiopathy and eye disease; glycaemic control status.	Internal test set only. No external validation. Best test-set AUC: nephropathy 0.902 +/− 0.040; angiopathy 0.889 +/− 0.059; neuropathy 0.859 +/− 0.050; eye disease 0.832 +/− 0.086; HbA1c 0.825 +/− 0.092. Calibration not reported.	Indirect. Nonadherent T2D overlaps conceptually, but the setting is single-centre and inpatient.	Sample size (n = 165) is small relative to the number of models fitted and to the high reported AUCs, making overfitting likely. Selecting the best-performing model per outcome further inflates apparent performance. No external validation; no calibration; single centre.
Kopitar et al., 2020 Sci Rep [27]	3723 adults with no prior diagnosis of diabetes attending preventive health examinations, 10 Slovenian primary care centres, 2014–2017.	Glmnet, random forest, XGBoost and LightGBM compared against a regression baseline. Reported against TRIPOD.	Fasting plasma glucose level and detection of impaired fasting glucose/undiagnosed T2D.	100 bootstrap iterations (internal). Calibration explicitly assessed. Best AUC 0.859 (Glmnet) versus 0.854 (regression baseline). R2 0.26–0.36 across models, indicating weak calibration. The authors conclude that machine learning conferred no clinically relevant advantage over regression.	None. Screening for undiagnosed diabetes in a population without diabetes; outside the target population of this review. Cited for methodological context only (Section 7).	Single database; high proportion of missing values. Working-age screening population underrepresents older adults, in whom T2D is most prevalent. Findings concern case-finding, not longitudinal management.

AMD, Associazione Medici Diabetologi; AUC, area under the receiver-operating characteristic curve; BMI, body mass index; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; EHR, electronic health record; GLP-1RA, glucagon-like peptide-1 receptor agonist; SGLT2i, sodium-glucose cotransporter-2 inhibitor; T2D, type 2 diabetes; TRIPOD, Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis.

First, direct evidence in non-insulin-treated T2D is limited. Of the studies summarized, only the analysis of therapeutic inertia in metformin monotherapy by Musacchio et al. is restricted to this population [19]; it reported a ROC-AUC of 0.81 for discriminating patients in whom intensification was delayed, and identified a pattern that is not intuitive in routine practice, in which a modest HbA1c rise accompanied by high variability was more strongly associated with inertia than a large rise. The remaining studies address mixed diabetes populations [18], nonadherent inpatients [17], the transition to insulin therapy [20], or case-finding in people without diagnosed diabetes [27]. The literature cited elsewhere in this review that is specific to the non-insulin-treated population concerns glucose monitoring rather than AI [11,12,13]. This asymmetry is not incidental. It reflects precisely the neglect that motivates this review, and it should be stated rather than obscured.

Second, reported performance is more modest than the volume of publication suggests, and validation is consistently incomplete. Calibration was reported in only one of the studies summarized [27], and none reported prospective or independent external validation. Where machine learning was compared directly against conventional regression, the advantage was small or absent: Ravizza et al. reported an AUC of 0.833 for a random forest against 0.827 for logistic regression across approximately half a million records [18], while Kopitar et al., assessing discrimination, calibration, and variable-selection stability together, concluded that more sophisticated models conferred no clinically relevant improvement over regression [27]. High reported discrimination must also be read against sample size: the AUC values above 0.85 reported by Fan et al. derive from 165 patients across 18 candidate models, with the best-performing model selected per outcome [17], a configuration in which optimistic performance is expected rather than surprising.

Third, and consequently, the applications discussed in this section differ in maturity and should not be read as a single body of established practice. Advanced interpretation of glucose data is supported by international consensus and is already in routine use [9,10]. Retrospective identification of therapeutic inertia has observational support in the relevant population [19,20]. Prospective prediction of individual metabolic deterioration, and clinical action taken on the basis of that prediction, remains investigational. These three levels are recorded for each application in Table 1.

6. Near-Future Applications and Plausible Clinical Scenarios

The near future of AI in non-insulin-treated T2D will probably not be defined by fully autonomous therapeutic systems. More realistically, its clinical impact will depend on tools capable of improving the precision, timing, and continuity of routine decision-making [14,15,16,22,23,24,25]. In this setting, the most credible applications are those that can be embedded into existing care pathways, reduce information fragmentation, and support clinicians in identifying when and how to intervene.

A major limitation of current diabetes care is that relevant information is dispersed across electronic health records, laboratory databases, glucose meters, CGM platforms, prescription records, body weight measurements, blood pressure values, wearable devices, and patient-reported outcomes. AI could help integrate these heterogeneous data streams into a coherent representation of the patient’s trajectory [14,15,16]. Instead of presenting separate fragments, future systems may summarize longitudinal patterns, detect discordant trends, and highlight clinically meaningful changes.

One plausible near-future application is personalization of follow-up timing. Current schedules are often based on fixed intervals, local organization, or clinician availability. AI could support a more dynamic model by estimating the probability of near-term deterioration and suggesting follow-up intensity accordingly [18,19,20]. Patients with stable multidimensional profiles could continue routine monitoring, whereas patients showing early unfavorable trends could be prioritized for earlier review, laboratory reassessment, educational reinforcement, or therapeutic reconsideration.

Future systems may also expand risk stratification beyond glycemia alone. T2D is a cardiometabolic disease, and the patient’s future trajectory is shaped by the interaction among glycemia, body weight, blood pressure, lipids, kidney function, liver involvement, frailty, adherence, and social determinants of health [5,6]. AI may enable more granular stratification by combining metabolic, renal, hepatic, cardiovascular, and behavioral variables. However, this broader stratification increases the need for transparency and clinical plausibility [22,23,24,25,26].

Generative and conversational AI tools may support patient education, structured communication, administrative work, visit preparation, and preliminary triage [22,23,24,25,26,28]. In non-insulin-treated T2D, their most realistic near-term role is not autonomous counselling or treatment modification, but clinician-governed support: summarizing clinical histories, identifying missing information before the visit, drafting patient-friendly explanations of agreed therapeutic plans, reinforcing adherence messages, and collecting patient-reported information.

These uses are not equivalent, and treating them as a single category obscures a steep gradient of both clinical risk and evidentiary support. Five levels can usefully be distinguished. Administrative assistance, such as scheduling, coding, and correspondence, carries the lowest clinical risk and does not touch clinical content. Documentation support, such as summarizing histories or identifying missing information before a visit, remains low-risk provided the output is verified by the clinician who signs it. Patient education, such as explaining an agreed therapeutic plan in accessible language, introduces moderate risk, because errors reach the patient directly and therefore require validated source content and explicit escalation rules. Clinical reasoning, such as interpreting a trajectory or weighing competing risks, carries substantially greater risk. Therapeutic recommendation, such as proposing a change in medication, carries the greatest risk of all.

The evidence base does not extend evenly across this gradient, and at its upper end it is largely absent. To our knowledge, no prospective or randomized study has evaluated LLM-based decision support in the management of type 2 diabetes; the available evidence is confined to retrospective benchmarking against expert judgment. In the most directly relevant assessment, GPT-4 was asked to extract and critically evaluate diabetologist-developed treatment plans from the clinical notes of patients with complex type 2 diabetes. It discerned personalized treatment goals accurately, but performed suboptimally when asked to identify plans that could be further improved and personalized [29]. The pattern is instructive and maps onto the gradient above: the model was competent at comprehension and extraction, and weakest precisely at the higher-order judgment that therapeutic decision support would require. Claims at that upper level should therefore be treated as untested rather than merely unproven, and the burden of demonstration rests with those proposing such systems.

These applications require particular caution. Large language models may generate plausible but incorrect information, omit clinically relevant nuance, or produce recommendations that are not aligned with guidelines, local regulatory constraints, or the individual patient’s comorbidity profile [22,23,24,25,26,28,30]. This risk is especially important when outputs concern medication changes, risk communication, or interpretation of symptoms. For this reason, generative AI should remain supervised, auditable, and clearly separated from autonomous therapeutic decision-making.

A clinically acceptable use of generative AI should therefore be bounded by predefined tasks, validated content sources, escalation rules, and human review. In this framework, conversational systems may improve continuity and health literacy without transferring medical responsibility away from the diabetologist. Current and near-future applications are summarized in Table 1.

7. Limitations, Risks, and Barriers to Implementation

The clinical translation of AI in non-insulin-treated T2D remains limited by scientific, technical, organizational, ethical, regulatory, and cultural barriers [22,23,24,25,26,28,30,31,32]. These limitations are not marginal details: they determine whether an algorithm becomes a useful clinical instrument or another source of complexity in already overloaded diabetes care pathways.

AI depends on data quality. In T2D, relevant information is often incomplete, inconsistently coded, or distributed across disconnected systems. Missing data are not random: they often reflect social vulnerability, limited access to care, reduced engagement, or organizational barriers. Algorithms trained on selected populations, highly structured datasets, or technologically advanced settings may not generalize to older adults, patients with multimorbidity, socially disadvantaged individuals, or routine clinics [22,23,24,25,26].

From a methodological perspective, this problem should be explicitly understood as a missing-not-at-random (MNAR) issue rather than a simple technical inconvenience. In real-world diabetes care, missing values may reflect reduced access to care, lower digital literacy, socioeconomic vulnerability, fragmented follow-up, or lower patient engagement. Standard imputation strategies may therefore obscure clinically meaningful patterns in vulnerable subgroups and inadvertently reinforce selection bias. For AI tools intended to support anticipatory care, missingness itself may need to be treated as an informative signal, not merely as a defect to be statistically corrected [22,23,24,25,26].

AI may also amplify existing inequalities if it is developed and deployed without attention to equity. Patients who generate more digital data, use connected devices, attend visits regularly, and have better health literacy may be more visible to algorithmic systems. Conversely, those with limited digital access, language barriers, lower socioeconomic status, or fragmented care may be underrepresented in training datasets and less likely to benefit from digital tools [26,28,30].

Explainability is another central requirement. In diabetes care, treatment decisions require clinical reasoning, patient preferences, safety considerations, and shared responsibility. A black-box prediction that identifies a patient as high risk without explaining the underlying drivers may be difficult to trust and difficult to translate into action [22,23,24,25,26]. Clinicians need to understand whether risk is driven by rising glucose values, weight gain, worsening renal function, reduced adherence, previous treatment failure, or other factors.

In this context, explainable artificial intelligence (XAI) approaches such as LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) may be useful because they estimate the relative contribution of individual variables to a model output [33,34]. For example, a predicted high risk of deterioration may be driven mainly by rising fasting glucose, worsening renal function, increasing glycemic variability, weight gain, or declining treatment persistence. This level of transparency may improve clinical trust and facilitate actionable interpretation, provided that explanations are presented in a format that is understandable within routine clinical workflow.

Even accurate tools may fail if poorly integrated into workflow. Busy diabetes clinics and primary care settings cannot accommodate additional platforms, excessive alerts, or outputs that require complex interpretation. AI outputs must be concise, timely, clinically meaningful, and connected to clear actions. Systems should define who receives the alert, who is responsible for acting on it, and how the response is documented [24,25].

Finally, AI in diabetes care raises regulatory and medico-legal questions. Predictive models and decision-support systems may evolve over time, interact with clinical workflows, and influence treatment decisions. Ethical governance is essential when models process sensitive health data or generate individualized risk predictions. Reporting standards such as CONSORT-AI and SPIRIT-AI, together with broader governance frameworks, are relevant to the design, evaluation, and implementation of these tools [28,30,31,32].

AI-based outputs should therefore be regarded as non-binding consultative support rather than autonomous clinical directives. Even when a decision-support system is technically validated, the final responsibility for diagnosis, therapeutic intensification, follow-up timing, and risk communication must remain with the clinician, who is required to interpret the algorithmic output in the context of the individual patient. This distinction is particularly important in the case of false negatives, false positives, or discordant clinical judgment, where algorithmic silence should not be interpreted as permission for therapeutic inertia [24,25,26].

Practical Barriers to Implementation

Beyond data quality and explainability, a set of practical barriers determines whether an algorithm reaches the clinic at all. These are routinely underestimated in the AI literature, and they are the reason why promising models rarely become clinical tools.

Interoperability is the first. Predictive models require longitudinal data drawn from electronic health records, laboratory systems, glucose platforms, and prescription databases that were designed independently and rarely share coding standards. In many European health systems, diabetes care is distributed across primary care, hospital specialist clinics, and regional registries, each with its own data model. Without shared terminologies and interfaces, a model validated in one setting cannot ingest the data of another, and building the data pipeline typically costs more effort than building the model.

Regulatory status is the second, and it is unsettled. Software that computes a risk estimate intended to guide a clinical decision may meet the definition of a medical device, with the conformity assessment, clinical evaluation, and post-market surveillance obligations that follow. The European Artificial Intelligence Act adds a further layer, classifying certain health applications as high-risk and imposing requirements on data governance, transparency, human oversight, and logging [30]. Models that continue to learn after deployment sit awkwardly within frameworks designed for static devices, since the version evaluated may not be the version in use.

Reimbursement is the third, and possibly the most consequential. In most health systems no established payment mechanism exists for predictive or decision-support software in diabetes care, and cost-effectiveness has not been demonstrated for any of the applications discussed in this review. Without a reimbursement pathway, adoption depends on local initiative or research funding, which constrains both scale and reproducibility.

Clinician acceptance is the fourth. Adoption depends less on measured accuracy than on whether the output is trusted, arrives while a decision is being made, and reduces rather than adds work. A prediction delivered outside the consultation, or one that cannot be reconciled with the clinician’s own assessment, will be disregarded irrespective of its discrimination [24].

Computational infrastructure and data governance are the fifth and sixth, and they interact. Running longitudinal models across large populations requires storage, computing capacity, and technical staff that most diabetes services do not have. Outsourcing these to external providers raises questions about where sensitive health data are processed, under whose responsibility, and with what audit trail [28,30]. Governance arrangements that are workable within one national system are not necessarily transferable to another, which limits the portability of both models and the evidence supporting them.

Importantly, these barriers should not be considered in isolation. Data fragmentation, limited interoperability, missing-not-at-random data, unequal digital visibility, algorithmic opacity, and uncertain responsibility interact with each other and may reduce real-world effectiveness even when internal predictive performance appears satisfactory [24,26]. Missingness and digital visibility bias are closely linked: patients who use connected devices, attend visits regularly, and generate structured data may become disproportionately visible to AI systems, whereas patients with fragmented follow-up, lower digital literacy, or social vulnerability may be underrepresented or mischaracterized. Without specific governance, this interaction may create a two-speed model of care, in which AI improves management mainly for patients who are already better monitored. This is why clinical validation should assess not only discrimination or calibration, but also actionability, workflow impact, equity, safety, and the ability to reduce therapeutic inertia in routine care across different patient groups.

This concern is not hypothetical. In a direct comparison of machine learning and regression-based models for detecting dysglycemia in primary care, Kopitar et al. found no clinically relevant gain in discrimination from the more complex models, weak calibration across all of them, and substantial variability in which variables were selected over time; they concluded that calibration, interpretability, and stability should govern model choice rather than predictive performance alone [27].

These barriers are easier to locate when mapped onto the pathway they interrupt. Figure 2 follows the same information from patient-generated data to clinical decision, and marks at each transition the obstacle most likely to stop it.

8. Practical Implications for Diabetologists

For diabetologists, the most useful way to interpret AI is not as a competing intelligence, but as a possible extension of clinical observation. The challenge in non-insulin-treated T2D is not the absence of therapeutic options, but the difficulty of using the right information at the right time to protect the patient’s future trajectory [4,5,6,7,8,14,15,16].

Its practical value therefore lies in improving timing, prioritization, and precision [14,15,16,19].

However, AI should not flatten clinical reasoning into algorithmic obedience. A model may indicate that a patient is at high risk of worsening control, but the clinician must understand whether this reflects disease progression, reduced adherence, socioeconomic difficulty, adverse effects, or an inappropriate therapeutic strategy [22,23,24,25,26].

In practical terms, the most relevant applications for routine diabetology may be simple and concrete: identifying who should be recalled earlier, who is likely to fail current therapy, who may benefit from structured education, who requires more intensive monitoring, and who may need therapeutic intensification before prolonged deterioration occurs. These are not abstract technological goals; they are everyday clinical problems.

9. Conclusions

AI is unlikely to transform non-insulin-treated T2D by replacing clinical judgment or automating therapeutic decisions. Its most credible contribution is narrower and more useful: helping clinicians recognize earlier when a patient’s trajectory is becoming unfavorable.

The management of non-insulin-treated T2D remains vulnerable to intermittent assessment, therapeutic inertia, fragmented data, and delayed recognition of metabolic deterioration. HbA1c, periodic visits, and conventional follow-up schedules remain essential, but they may not always capture the dynamic complexity of the disease [2,4,9,10]. AI may help bridge this gap by integrating multidimensional data, detecting hidden patterns, predicting deterioration, and supporting more individualized follow-up and treatment timing [14,15,16,18,19,20].

The strength of this proposition should not be overstated. Prospective randomized evidence for AI-supported decision-making in this population is currently absent rather than merely limited; the supporting studies are retrospective and observational. Real-world implementation studies are scarce, and the few available models have not been evaluated prospectively or validated externally. Cost-effectiveness has not been established for any of the applications discussed. Where machine learning has been compared directly with conventional regression in diabetes prediction, the incremental gain has been small [18,27], which suggests that the value of these tools will lie in integration, timing, and interpretability rather than in raw predictive accuracy. The transition from reactive management to anticipatory care is therefore best understood as a direction of travel and a research agenda, not as an available capability. It does not imply overtreatment, technological autonomy, or replacement of the diabetologist. It means using data more intelligently to see earlier, decide more precisely, and intervene before failure becomes established—and it requires evidence that does not yet exist.

For that transition to become possible, AI tools must be clinically validated, explainable, equitable, interoperable, and embedded into realistic care pathways [22,23,24,25,26,28,30,31,32]. Their success should not be judged only by predictive accuracy, but by their ability to reduce therapeutic inertia, improve patient trajectories, support safer decisions, and enhance the quality of clinical reasoning. Generating that evidence is the more useful task. It requires prospective studies in non-insulin-treated T2D specifically, rather than extrapolation from mixed diabetes cohorts or from type 1 disease; external validation across health systems with differing data completeness; reporting against established standards such as TRIPOD, CONSORT-AI, and SPIRIT-AI [31,32]; and endpoints that measure clinical action and patient trajectory rather than discrimination alone. Until then, AI in non-insulin-treated T2D should be understood less as a new therapeutic actor and more as a proposed new way of reading clinical time—one whose value remains to be demonstrated.