This review was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines (Multimedia Appendix 1) to maximize methodological rigor, transparency, and reproducibility [29]. The review addressed the question: “How has AI been used to improve Ozempic therapy for T2D, and what conceptual framework can be proposed for its future clinical integration?” This led to the search strategy, data extraction, and synthesis of evidence. The literature search used Multisource Information Aggregator for Educational and Research Purposes (MIAGE) Scholar, an academic search interface that aggregates metadata from multiple scholarly sources. The platform provides streamlined access to peer-reviewed publications with features for citation analysis and research discovery, supported by integrations with key academic databases and identifiers. The literature search covered studies published from January 2019 to March 2025 to ensure the inclusion of the latest use of AI applications for Ozempic treatments. January 2019 was used as the starting point to identify the time when AI approaches began to be meaningfully applied to diabetes care with semaglutide, vs the time of the initial pharmacologic approval of semaglutide.

Search filters were used to include only studies that met the following criteria: English-language publications, peer-reviewed journals, and research on people. The following search terms were used in the Boolean query: “type 2 diabetes” OR “T2D” for type 2 diabetes; “Ozempic” OR “Semaglutide” OR “GLP-1 receptor agonist” for semaglutide/ozempic/GLP-1 receptor agonists; and “artificial intelligence” OR “machine learning” OR “deep learning” OR “predictive analytics” for AI.

Textbox 1 presents the Boolean query used.

Inclusion and exclusion criteria for the studies are given in Textbox 2.

Due to inherent heterogeneity in study design, AI approaches, and outcome measures, a qualitative narrative synthesis approach was used. Studies were systematically organized and synthesized in four dominant thematic clusters: (1) AI-managed patient stratification and risk prediction (6 studies), (2) imaging and body composition analysis (4 studies), (3) cardiovascular and metabolic risk evaluation (5 studies), and (4) real-world data and personalized dosing (3 studies). The 4 thematic clusters were identified through an inductive synthesis of study objectives, AI methodologies, data sources, and primary clinical end points, ensuring that each cluster reflects a dominant and recurrent application pattern observed across the reviewed literature.

These findings were integrated to develop the AI-Ozempic conceptual framework, which summarizes current applications and proposes future research directions.

A meta-analysis was not feasible, as fewer than 2 studies shared a common design, intervention definition, end point, and follow-up window required for valid statistical pooling.

PRISMA-based methods (Figure 1) identified 18 relevant articles addressing the role of AI in the management of T2D in the context of semaglutide use. The studies summarized in Table 1 span multiple designs and evidence types, providing a broad view of how AI is being applied in this therapeutic area. Among the included evidence syntheses, 1 study conducted a focused meta-analysis of placebo-controlled randomized trials assessing body composition and musculoskeletal outcomes associated with GLP-1 receptor agonist–based therapies, including semaglutide-relevant contexts [25]. Another article provided a narrative review discussing omics signatures and mechanistic considerations related to cardiovascular response in semaglutide-treated populations [14].

Several included studies were observational and real-world investigations, including retrospective cohort analyses of semaglutide effectiveness in routine clinical practice [20], as well as imaging-based work using automated clinical trial–based AI tools to quantify body composition changes after semaglutide initiation [13]. In addition, some included work applied ML approaches to large-scale real-world data in T2D populations to model cardiovascular outcomes and comparative medication-related risks in GLP-1 receptor agonist–treated groups [5]. Lastly, there were commentaries/opinion pieces providing broader perspectives on clinical, policy, and implementation implications of GLP-1 therapies in contemporary care [9]. Together, this diverse group of studies (ranging from empirical real-world analyses to conceptual perspectives) illustrates the evolving role of AI-enabled methods around semaglutide and broader GLP-1 receptor agonist use in T2D care.

To synthesize the findings, each of the 18 studies was grouped into 1 of 4 thematic clusters based on the primary clinical domain or challenge addressed by the AI application. The first cluster, AI-managed patient stratification and risk prediction (6 studies), focused on the use of supervised ML models and risk engines to identify patients most likely to benefit from semaglutide, stratify T2D risk profiles, and predict therapeutic responsiveness [20,27,33-36]. The second cluster, imaging and body composition analysis (4 studies), used AI or advanced modeling techniques to quantify changes in lean and fat mass, along with other physiological metrics associated with semaglutide therapy [9,13,28,37]. The third cluster, cardiovascular and metabolic risk evaluation (5 studies), examined how AI can enhance the assessment of cardiovascular risk and other metabolic complications in patients with T2D [5,14,22,25,31]. The final cluster, real-world data and personalized dosing algorithms (3 studies), investigated AI-powered dosing strategies and real-time monitoring for individualized treatment optimization [3,24,32].

While some thematic overlap exists—such as studies addressing both risk prediction and cardiovascular outcomes—this classification reflects the dominant focus of each investigation. Collectively, these studies demonstrate AI’s multifaceted potential in optimizing Ozempic therapy by improving patient selection, quantifying body composition changes, refining cardiometabolic risk evaluation, and guiding personalized dosing. Reported outcomes include enhanced HbA_1c reduction, increased weight loss, improved cardiovascular markers, greater dosage precision, reduced side effects, better patient stratification, and improved cost-effectiveness when compared to conventional use of semaglutide.

Although not all studies directly compare AI-driven dosing with standard Ozempic dosing head-to-head, several quantifiable clinical end points highlight the benefits of AI integration. Table 2 summarizes these findings, and Figure 2 [5,13,20,24,25,27,32,33,36] provides a visual representation of average performance improvements when AI is used. In the case of HbA_1c reduction, standard fixed dosage typically achieves an average reduction of approximately 1.2%, while AI-titrated or ML-controlled dosage shows an extra reduction of around 1.3%, as shown in the studies of [20,32,36]. For weight loss, traditional semaglutide treatment results in an average reduction of 5-7 kg, while AI-based models and imaging analyses [13,24] indicate a step-by-step loss of 2-3 kg, partly due to previous identification of high responders. In the case of cardiovascular risk, AI-enabled tools-inclined risk engines [36], atherosclerotic cardiovascular disease prevention models [33], and prescription algorithms [25] show additive benefits over standard care. When it comes to side effects, especially gastrointestinal side effects that occur in 20%-25% of patients with standard dosage, predictive AI models have been shown to reduce these incidents by about 20%-30% [20,27]. Finally, cost-effectiveness is improved by AI guidance, as suboptimal or delayed Ozempic use—often seen in standard practice—can be reduced through more precise resource allocation and personalized treatment strategies [24,32,36]. AI enhances glycemic control, weight management, cardiovascular system protection, economic efficiency, and the overall side effects, demonstrating advantages over one-size-fits-all dosing strategies (Figure 2). In summary, the evidence indicates that AI integration leads to improved HbA_1c reduction, extra weight loss, improved cardiovascular risk outcomes, and fewer adverse effects compared with standard regimens. Studies varied from conceptual frameworks (policy-, cost-, or biomarker-based applications) to robust ML models in observation data in the real world. Despite heterogeneity in design and AI sophistication, the collective evidence supports the notion that personalized, computer-driven Ozempic therapy can exceed standard protocols.

To clarify the algorithm’s focus of each study, Table 3 categorizes the types of AI methods used in the literature. The most prevalent techniques fall under supervised ML, encompassing commonly used models, such as logistic regression [38,39], random forest [39,40], gradient boosting [41], and support vector machines [42], which are prominently featured in studies [20,33,36]. DL approaches are also represented, with applications, such as convolutional neural networks–based [43] imaging analysis [13] and deep neural networks–based risk prediction models [5]. The d-Nav system described in Warren et al [32] exemplifies the use of reinforcement learning, which enables real-time, adaptive titration of semaglutide dosage. Studies used unsupervised learning and clustering techniques, such as k-means [39,44] or uniform manifold approximation and projection [45], aimed at identifying patient subgroups and stratification patterns [27]. Additional methods included risk engines and simulation models, such as the building, relating, assessing, and validating outcomes model and the logic learning machine [46], which were used to simulate treatment outcomes or support algorithmic recommendations [24,36]. Finally, a subset of studies offered conceptual or editorial contributions, discussing AI integration frameworks without presenting empirical implementations or new algorithmic models [9,22,31,35].

Overall, the distribution highlights the dominance of supervised ML techniques in predicting Ozempic-related outcomes. However, more advanced AI technologies—such as reinforcement learning, explainable AI [47], and multiomics integration—are emerging in a small but growing number of studies [14], signaling a gradual shift in the field toward more sophisticated and adaptive modeling approaches.

A translational matrix (Table 4) maps each study’s AI method, input data type, clinical objective, and real-world readiness level. For example, studies using logistic regression [20] or gradient boosting [33] rely on EHR data to forecast patient responsiveness to Ozempic and are categorized as having moderate translational readiness for clinical use. In contrast, DL-based imaging models [13] are at a pilot testing stage—effective at quantifying changes in body composition, but not yet deployed in routine clinical workflows. Conceptual works that suggest AI integration with omics data to enhance cardiovascular risk modeling [14] remain theoretical, lacking empirical implementation.

The matrix serves as a practical tool for clinicians and researchers to assess how close each AI approach is to real-world application. (1) Clinically validated techniques; for example, reinforcement learning-based insulin system (eg, d-Nav [32]), which can be adapted for semaglutide use. (2) Partially validated methods, including risk engines [36], which have undergone external validation but lack prospective studies specific to Ozempic dosage. (3) Theoretical or early-stage innovations; for example, clustering of multiomics, which is promising but has not yet been tested in clinical workflows.

Overall, this matrix acts as a practical guide to identify which AI methods are ready for integration into Ozempic-based care, highlighting those already in clinical use and those still in development.

A structured term frequency analysis was conducted on the titles of the 18 included studies. Conceptually equivalent terms (eg, GLP-1 RA and GLP-1 receptor agonists, machine learning, artificial intelligence, and deep neural networks) were consolidated under unified labels. As shown in Table 5, GLP-1 receptor agonists (n=7, 38.9%) and diabetes-related terminology (n=6, 33.3%) were the dominant clinical anchors. Notably, machine learning/artificial intelligence and risk or prediction terms each appeared in 22.2% (n=4) of titles, indicating a substantial methodological emphasis on computational modeling and data-driven stratification. Semaglutide and obesity-related terms were present in 16.7% (n=3) of studies, further reflecting the clinical focus on outcome optimization. Overall, the distribution of terms reflects the emerging integration of AI-driven prediction within GLP-1–based therapeutic strategies, suggesting an expanding methodological role in personalized diabetes care.

Personalized dosing for Ozempic (semaglutide) is common and has been done by physicians for many years in real-world practice. The framework we have proposed using AI is not to replace physician-led personalization but to provide the same type of information as the physician in a consistent and more scalable manner for all patients. Therefore, the review’s goal is to enhance the transparency, reproducibility, and consistency of how each physician adjusts treatments individually within their own patient population. Based on the findings from the systematic review above, the proposed AI-Ozempic framework was developed. Each of the four main evidence categories was mapped to specific components of the framework: (1) AI-assisted patient classification and response prediction [20,33,36], (2) body composition analysis through use of medical images [13,28], (3) evaluation of cardiovascular and metabolic risk [5,22,25], and (4) development of personalized dosing strategies based on real-world clinical experience [3,24,32]. Therefore, the goal of the AI-Ozempic framework is to be an operational, tangible tool that provides a way to apply the various AI-based applications of these types of evidence-based practices; therefore, it will be a step beyond developing only a theoretical model. As demonstrated in the results section and supported by key findings, this framework establishes a structured ecosystem for continuous monitoring and adaptive optimization of Ozempic therapy in T2D through real-time, AI-driven decision-making. The overarching objective is to leverage patient engagement for ongoing data collection, predictive modeling, clinical decision support, and personalized, proactive management of T2D. The proposed framework directly addresses key limitations in existing Ozempic treatment models by enabling real-time, data-driven decision-making. Traditional protocols are often dependent on fixed dosage regulations and retrospective evaluations, while this framework facilitates dynamic dosage through continuous patient monitoring. Predictive algorithms predict glycemic variability or side effects before they occur, thus changing the paradigm from reactive to proactive care. Furthermore, continuous feedback loops support iterative model forwarding and promote treatment of treatment over time. Integration into EHRs ensures that AI-generated recommendations are contextually actionable, thereby aligning clinical workflows with predictive insights.

To effectively implement this conceptual model, it must be anchored to well-defined and measurable clinical key performance indicators (KPIs). By aligning each AI-driven function—such as dosing adjustments or risk alerts—with specific KPIs, health care teams can systematically monitor, evaluate, and refine the performance of the AI-Ozempic system in clinical practice.

Each AI-driven component in the framework is explicitly linked to one or more clinical KPIs through operational pathways. For instance, predictive models for hypoglycemia generate alerts that prompt dose review or patient outreach, leading to measurable reductions in hypoglycemic events. Similarly, adherence monitoring systems detect missed doses via digital health inputs, triggering automated reminders that improve medication persistence rates. Glucose forecasting models anticipate elevated HbA_1c levels, prompting earlier intervention and influencing long-term glycemic control KPIs. This structured logic—AI prediction → clinical action → measurable KPI outcome—ensures traceability and real-world impact. By linking AI functionality directly to these KPIs, clinicians can objectively assess whether the model achieves clinically meaningful gains (eg, lower HbA_1c by 0.5% above standard care). Ongoing KPI monitoring also facilitates iterative improvements in the underlying AI algorithms, fostering a cycle of evidence-based practice.

The reviewed studies used different methods of AI with different clinical objectives. No single AI method is considered the optimal Ozempic therapy. Instead, effectiveness depends on the clinical task. The supervised ML model that uses EHR data has provided the most evidence regarding patient stratification and predicting treatment response. The DL model that uses images provides the most evidence related to assessing and tracking changes in body composition due to Ozempic therapy. However, they are still mostly in the pilot phase. The adaptive and reinforcement learning models have been shown to have the greatest potential for real-world clinical application when providing for personalized dosing of Ozempic therapy; however, there is currently very limited evidence available for these models, and much of the evidence that exists is not specific to semaglutide. The risk engines and simulation models have also demonstrated the ability to assess long-term cardiovascular and cost-effectiveness of Ozempic therapy, and while they do indirectly assist with making routine dosing decisions, they are generally not sufficient to make these decisions.

In order to transform AI-driven Ozempic therapy from a proof-of-concept to standard clinical practice, more strategic directions must be followed. First, large-scale multicenter randomized controlled trials (RCTs) involving different patient populations are critical to validating AI-guided dosing algorithms and benchmarking them against standard-of-care regimens. In addition, extended follow-up assessments beyond 24 weeks are necessary to evaluate long-term outcomes such as glycemic durability, cardiovascular protection, and patient adherence. The development of interoperable and standardized data ecosystems, using models such as the Observational Medical Outcomes Partnership common data model [48] and compliance with findable, accessible, interoperable, and reusable (FAIR) principles [49], will be crucial to integrating different data streams, including EHRs, laptops, imaging, and omics. Transparency and explainability in AI models must be prioritized to promote clinician trust, support regulatory paths, and ensure that algorithms can be evaluated rigorously for fairness and reliability. Equally important is the mitigation of algorithmic bias and compliance with regulatory frameworks such as the Food and Drug Administration’s (FDA) Good ML Practice [50], General Data Protection Regulation [51], and the Health Insurance Portability and Accountability Act [52], which ensure fair use across varied T2D populations. Finally, an extensive cost-use analysis is required to determine whether the personalized approach can justify the high cost of semaglutide, thus informing reimbursement strategies and decisions by decision-makers. Collaborating on future innovation with these imperatives will accelerate the integration of AI into Ozempic-based care, improve glycemic control, cardiometabolic health, and general patient quality.

Despite its promise, AI-enhanced semaglutide therapy must address critical ethical and legal challenges to ensure safe, equitable, and transparent deployment. Key issues include privacy (protecting patient data on a large scale), informed consent (ensuring patients understand how AI informs or overrides physician decisions), algorithmic bias (mitigating underperformance in underrepresented subgroups), and data ownership (clarification of who controls patient information). Explainable AI models are crucial for compliance with bodies such as the FDA [50] and European Medicines Agency [53], reassuring clinicians that dosing recommendations are not “black boxes.” Establishing clear management frameworks and ongoing involvement with patients, clinicians, and decision-makers will maintain public trust as AI-based Ozempic therapy transitions from experimental to mainstream practice.

Ensuring that AI models used in Ozempic therapy meet standards of trustworthiness is critical. Future studies must include bias detection protocols, explainability tools (eg, Shapley additive explanations [54] and local interpretable model-agnostic explanations [55]), and rigorous fairness audits to ensure equitable treatment across age, gender, and ethnic groups. Integration with trustworthy data preparation pipelines and alignment with the FDA’s Good Machine Learning Practice for Medical Device Development guidelines [50] and the EU (European Union) Artificial Intelligence Act [56] will be vital for regulatory approval and ethical deployment.

Although many studies emphasize AI’s promise to individualize semaglutide therapy, several restrictions limit current generalizability. Methodological heterogeneity in AI methods (random forests, gradient boosting, convolutional neural networks, k-means, etc) varies widely, complicating direct comparisons and hindering standardized best practices [35,36]. Integration barriers across data sources: merging structured EHR data with unstructured inputs (imaging, wearables, and omics) is still a challenge, and limits extensive predictive modeling [20,31]. Limited long-term evidence: most findings focus on short-term outcomes (12-24 weeks) without exploring extended cardiovascular risk reduction, hospitalization, or mortality [33,36]. Small or homogeneous samples: restrictive sample sizes and limited demographic diversity reduce external validity and increase risk of bias [14,26]. Regulatory and ethical challenges: tools such as d-Nav [32] and logic learning machine [24] face developing supervision, which requires robust explainability and data-privacy compliance. Large-scale RCT deficit: the absence of a sizable, multicenter RCT to confirm the AI-driven semaglutide dose limits direct comparisons with standard regimens. Insufficient long-term follow-up: the persistent effect of AI-defined semaglutide on weight maintenance, cardiovascular risk, and compliance is still unclear due to a lack of multiyear data. Missing cost-use analyses: few studies quantify whether AI-based personalization offsets the high cost of semaglutide, leaving overall economic feasibility unexplored. To address these limitations, especially through large studies and trials, transparent AI models and robust economic evaluations are critical to efficiently integrate AI solutions into routine semaglutide therapy and broader diabetes treatment.

AI technologies now enable semaglutide therapy to move from fixed, population-level dosing toward data-driven personalization. Evidence from 18 peer-reviewed studies (2019-2025) shows that AI-enhanced regimens deliver greater HbA_1c reduction, larger weight loss, and fewer adverse events than standard protocols. High CQAS scores across relevance and clinical-impact domains confirm the maturity of the evidence base, although transparency deficits remain for proprietary or black-box models.

The review introduces 3 advances that have not previously been combined in the semaglutide literature: (1) CQAS benchmarking—a 5-domain rubric that allows objective comparison of methodological quality and clinical use; (2) a translational matrix that links each AI technique to its input data, application domain, and real-world readiness—showing, for example, that reinforcement-learning titration engines are nearest to deployment whereas multiomics integration remains conceptual; and (3) a trustworthy AI-Ozempic framework—a closed-loop architecture that couples predictive analytics, adaptive dosing, and patient-feedback modules to measurable clinical KPIs such as time-in-range, medication persistence, and cardiovascular-event reduction.

Together, these components outline a practical roadmap for precision diabetes care and lay the foundation for scalable digital therapeutics across metabolic disorders.

Validation in multicenter, long-horizon trials; FAIR-compliant, integrative data infrastructures; bias-mitigated and explainable models; and robust cost-use evidence are the critical levers that can convert AI-guided semaglutide from experimental promise into routine care across diabetes and related cardiometabolic diseases.