AI-Driven Lifestyle Interventions for Reducing Diabetes Risk: Systematic Review

Introduction

Prediabetes is a growing global health burden. In 2024, the estimated global prevalence was 12.0% for impaired glucose tolerance and 9.2% for impaired fasting glucose, with both projected to increase through 2050 [1]. Diagnostic thresholds differ in sensitivity and clinical usage across glycemic markers [2]. Prediabetes affects approximately one-third of US adults and is associated with increased risks of type 2 diabetes, cardiovascular disease, and stroke [3,4]. Annual progression to type 2 diabetes is estimated at 5% to 10%, and lifetime progression may reach 70% without intervention [5,6]. Lifestyle modification, including dietary change, physical activity, and weight loss, can delay or prevent progression [7].

The Diabetes Prevention Program (DPP) is a structured, evidence-based lifestyle intervention combining dietary change, physical activity, and behavioral coaching to prevent or delay progression to type 2 diabetes among high-risk adults [8]. The original DPP trial and its translations established that this approach can substantially reduce progression, and this evidence underpins the Centers for Disease Control and Prevention’s National DPP and similar programs now delivered internationally [9,10]. However, in an analysis of 44 low- and middle-income countries, fewer than half of high-risk individuals reported receiving fundamental prevention activities [11]. Structured prevention programs remain concentrated in a limited number of countries [12]. In the United States, access is uneven, particularly in rural and underserved communities, and participation may be constrained by program intensity, limited time, scheduling conflicts, and coach-participant cultural mismatch [13-15]. A recent scoping review confirmed that DPP reach remains lowest among the populations at greatest risk, reinforcing that the evidence-practice gap is one of access and sustained engagement rather than intervention efficacy [16]. These barriers highlight the need for scalable approaches that can extend prevention services without compromising effectiveness and equity.

Advances in artificial intelligence (AI), including adaptive algorithms, machine learning–based personalization, predictive analytics, and conversational agents, have enabled digital interventions that tailor recommendations, coaching intensity, and engagement strategies to individual users [17,18]. Integration of continuous glucose monitoring (CGM) with AI may also support personalized risk assessment and behavioral feedback [19], but the clinical effectiveness of these approaches remains uncertain. Several reviews have examined digital diabetes prevention interventions, but each leaves a distinct gap relevant here. A systematic review of digital health technologies for type 2 diabetes prevention found that just over half of included randomized trials achieved clinically significant outcomes, but AI-enabled and non-AI digital tools were evaluated together rather than distinguished by mechanism [20]. A broader systematic review of digital health interventions for type 2 diabetes prevention, structured around health system outcomes such as population health, patient experience, and cost, similarly pooled digital tools, including AI-based interventions, without isolating AI-specific effects or appraising evidence by AI type [21]. A 2025 scoping review by Lalani et al [17] mapped AI and machine learning apps across the prediabetes care continuum. That review did not systematically appraise study quality, evaluate comparative effectiveness against standard DPPs, or organize evidence by AI intervention type and outcome domain. Across this literature, no review has yet combined systematic quality appraisal with stratification of comparative evidence by AI mechanism and outcome domain, which is necessary to determine whether specific categories of AI intervention, rather than digital delivery in general, are associated with clinical benefit.

This review had three objectives: (1) to evaluate the effects of AI-driven lifestyle interventions on clinical, behavioral, and engagement outcomes; (2) to assess whether findings differed by AI intervention type; and (3) to assess whether findings differed by population or delivery setting.

Methods

Eligibility Criteria

This systematic review was reported in accordance with PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 [18], including the PRISMA 2020 for Abstracts checklist (Multimedia Appendix 1) and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) [22]. Eligibility criteria were established in advance based on the research question. Studies were eligible for inclusion if they (1) enrolled children, adolescents, or adults with prediabetes or elevated risk for type 2 diabetes, or healthy participants enrolled in feasibility or pilot studies; (2) evaluated a lifestyle intervention delivered through a mobile, web, or SMS text messaging platform with an AI-driven component; (3) included a comparator such as usual care, minimal or no intervention, a waitlist, a non-AI digital intervention, or a standard lifestyle program when the study design was comparative; (4) reported at least 1 clinical, behavioral, or engagement outcome; (5) used a randomized, nonrandomized comparative, observational, single-arm, feasibility, or pilot design; and (6) published a peer-reviewed English-language report from January 1, 2015, through August 3, 2026. Elevated diabetes risk was defined by impaired fasting glucose, impaired glucose tolerance, hemoglobin A_1c (HbA_1c) of 5.7% to 6.4%, overweight or obesity (BMI at least 25 kg/m²), or a validated diabetes risk score.

For this review, an AI-driven intervention was defined as an intervention that used algorithmic personalization, adaptive goal setting, predictive modeling, machine learning–based tailoring, or a conversational system with individualized adaptation. Scripted, nonadaptive rule-based chatbots and fixed decision tree interventions without personalization, learning, or adaptation were excluded. Additional exclusions were qualitative-only studies, interventions delivered solely by human coaches without algorithmic personalization, interventions focused exclusively on treatment of established diabetes without a prevention or risk reduction objective, hardware-only technologies without a behavioral intervention, conference abstracts without full reports, protocols without results, editorials, commentaries, and reviews.

Protocol and Registration

No formal review protocol was prepared, and the review was not prospectively registered. This reflects an exploratory synthesis rather than a prespecified confirmatory review.

Information Sources

PubMed, Embase, CINAHL, Web of Science, and IEEE Xplore, and ClinicalTrials.gov were searched. Embase was searched via Elsevier platform, CINAHL via EBSCOhost, and Web of Science via the Clarivate Web of Science interface; PubMed, IEEE Xplore, and ClinicalTrials.gov were searched directly through their native web interfaces. Reference lists of included studies were examined for potentially eligible reports (backward citation searching), and forward citation searching was conducted in Google Scholar. No dissertation, preprint, or other gray literature databases were searched beyond ClinicalTrials.gov. Study authors, experts, and manufacturers were not contacted to identify additional studies or obtain unpublished data.

Search Strategy

The searches were conducted on August 3, 2026, and included English-language reports published from January 1, 2015, through the search date (reflecting the increasing adoption of AI apps). The strategy combined controlled vocabulary and text words for 3 concepts: prediabetes or elevated diabetes risk, AI-driven digital health, and lifestyle behaviors or engagement relevant to diabetes prevention. The strategy was developed de novo rather than adapted or reused from a previously published review, and no validated study design search filter was applied, as none exists for this literature. The strategy was constructed and internally reviewed by the review team but was not peer reviewed by an independent librarian. The search was run once, on August 3, 2026, across all sources, and was not subsequently updated or rerun closer to submission. Records were exported to Covidence (Veritas Health Innovation), which was used for duplicate removal and screening [23]. The complete database strategies, platform-specific syntax, and source-specific record counts are provided in Multimedia Appendix 2, and the full PRISMA-S checklist [24], with “not applicable” items explained, is provided in Multimedia Appendix 3.

Selection Process

One reviewer (OIE) screened all titles and abstracts and assessed all retrieved full-text reports against the eligibility criteria. A second reviewer (TN) independently verified all exclusion decisions. Disagreements were resolved through discussion, with adjudication by a third reviewer (RAF) when consensus could not be reached. No automation tool was used to determine eligibility. Detailed reason-level coding was retained for full-text exclusions but not for title and abstract exclusions.

Data Collection Process

One reviewer (OIE) extracted data from all included studies using a standardized, piloted extraction form (Multimedia Appendix 4); a second reviewer (TN) independently verified all extracted data, with disagreements resolved by a third reviewer (RAF).

Data Items

The extraction form (Multimedia Appendix 4) captured study design, country and setting, participant characteristics, sample size, intervention duration and delivery platform, AI functionality and origin (commercial, research-developed, or nonprofit or public), comparator, funding source, author-industry relationships when reported, and study findings. Outcomes were grouped as clinical (weight, BMI, HbA_1c, fasting glucose, and CGM-derived metrics), behavioral (diet, physical activity, and sleep), or engagement (initiation, retention, use, session completion, adherence, and perceived usefulness). All reported outcome measures and time points relevant to these domains were sought. Unclear or unreported information was recorded as not reported; no values were imputed.

Study Quality Assessment

Study quality was assessed using the National Heart, Lung, and Blood Institute (NHLBI) Study Quality Assessment Tools appropriate to each study design: controlled intervention studies, observational cohort and cross-sectional studies, and pre-post studies without a control group (Multimedia Appendix 5 [25-35]) [36]. The nonrandomized comparative study was evaluated with the controlled intervention studies tool, since its design included a concurrent, non–randomized control arm; items assessing random sequence generation and allocation concealment were scored “not applicable.” One reviewer (TN) completed each assessment, and a second reviewer (OIE) independently verified the ratings. Individual items were coded as “Yes,” “No,” “Cannot determine,” “Not reported,” or “Not applicable.” Because the NHLBI tools do not prescribe numeric cutoffs, overall ratings of “Good,” “Fair,” or “Poor” reflected qualitative judgment of the number and seriousness of design-specific limitations.

Effect Measures

For the randomized trial, effect measures included between-group risk differences with 1-sided 95% CIs for the composite primary outcome and the weight loss threshold; glycemic, engagement, and physical activity outcomes were reported as between-group percentages or medians as presented by the investigators. Because program initiation and completion were not part of the trial’s prespecified noninferiority framework, 95% CIs for these risk differences were calculated by the review team using the normal approximation method for the difference of 2 independent proportions. For the nonrandomized comparative study, between-group mean differences, adjusted odds ratios, CIs, and P values were reported. For observational and uncontrolled pre-post studies, effect measures included within-group percentage or absolute change in weight, percentage-point changes in CGM metrics, changes in HbA_1c or fasting glucose, and percentage- or rate-based behavioral and engagement measures, as reported by each study.

Synthesis Methods

Results were organized by study design and by 3 outcome domains: clinical, behavioral, and engagement. Findings were then stratified by investigator-defined AI intervention type: AI lifestyle coaching systems (adaptive goal setting and tailored feedback), predictive AI platforms (real-time or longitudinal data generating individualized predictions or recommendations), and conversational AI systems (chatbot interfaces emphasizing engagement). Variation by population and setting was explored descriptively across age, baseline metabolic risk, geographic location, and socioeconomic status. The randomized trial was prioritized in interpretation, followed by the nonrandomized comparative study; observational and uncontrolled pre-post findings were treated as descriptive and hypothesis-generating.

No meta-analysis was performed because the included studies evaluated fundamentally different designs and estimands and reported incompatible outcomes, including threshold-based risk differences, continuous percentage change, absolute change, CGM-derived metrics, and engagement measures, so a common standardized effect metric could not be calculated. Pooling randomized, nonrandomized comparative, observational, and uncontrolled pre-post evidence would have produced a summary estimate with unclear causal and clinical interpretation. Potential overlap among reports from the same commercial platforms further limited aggregation. No meta-analysis was performed; results were summarized narratively using the original metrics.

Weight and glycemic estimates are tabulated by study design in Multimedia Appendix 6 [25-35]. Ranges and medians were reported only when at least 2 studies within a design category used sufficiently comparable metrics. No statistical test of heterogeneity, meta-regression, or sensitivity analysis was conducted because no pooled estimate was calculated. The synthesis was reported using the Synthesis Without Meta-Analysis guideline (Multimedia Appendix 7).

Reporting Bias Assessment

ClinicalTrials.gov was searched on August 3, 2026, using terms aligned with the database search concepts to identify completed, ongoing, or potentially unpublished eligible studies. The registry search did not add any included study. A formal statistical assessment of reporting bias was not performed because the number of studies was small, outcomes were heterogeneous, and many commercial evaluations were not prospectively registered.

Certainty Assessment

A formal certainty-of-evidence framework, such as GRADE (Grading of Recommendations Assessment, Development, and Evaluation), was not prespecified or applied. The evidence could not be organized into sufficiently comparable intervention-comparator-outcome groupings for a formal domain-level certainty rating. Risk-of-bias assessments and study design were instead used to calibrate interpretation, and the absence of a formal certainty assessment is acknowledged as a limitation.

Results

Study Selection

The searches identified 8032 records: PubMed (n=1756), Embase (n=5172), CINAHL (n=493), Web of Science (n=300), IEEE Xplore (n=308), and backward or forward citation searching (n=3) (Figure 1). After 1732 duplicates were removed, 6300 records were screened by title and abstract, and 6272 records were excluded because they did not meet eligibility criteria. A total of 28 reports were sought and retrieved for full-text assessment. Overall, 17 were excluded: no trial results, trial registration, or protocol only (n=3); commentary (n=4); wrong intervention (n=2); wrong population (n=3); conference proceeding (n=4); and wrong study design (n=1). A total of 11 studies were included.

‎

Study Characteristics

A total of 11 studies published from 2017 through 2025 evaluated AI-driven lifestyle interventions for reducing diabetes risk (Table 1). Ten studies were conducted in the United States, and 1 was conducted in rural India [25]. Populations were predominantly adults with prediabetes or elevated risk; 2 studies enrolled adolescents [26,27], and 1 study enrolled adults with no diabetes [25]. Study sample sizes ranged from 13 to 16,327 participants and summed to 30,9991 participant records, although overlap among related commercial platform reports precluded estimation of the number of unique participants. Follow-up ranged from 33 days to 12 months. The evidence base comprised 1 randomized clinical trial [28], 1 quasi-experimental nonrandomized comparative study [25], 4 observational studies [29-32], and 5 uncontrolled pre-post studies [26,27,33-35]. Seven studies evaluated AI lifestyle coaching systems [25,28-31,33,34], 2 evaluated predictive AI platforms [32,35], and 2 evaluated conversational AI chatbots [26,27]. Eight interventions were delivered primarily through mobile apps [28-35], 3 used SMS text messaging platforms [25-27], and 2 integrated CGM [32,35]. Ten studies evaluated commercially developed interventions; the Arogya World intervention was nonprofit-led [25]. No purely research-developed or publicly developed platforms were identified.

Quality Assessment of Included Studies

Out of 11 studies, 5 were rated “Good,” 3 “Fair,” and 3 “Poor” (Table 2 and Multimedia Appendix 5). The randomized trial was rated “Good” [28], and the nonrandomized comparative study was rated “Fair” [25]. Among the 4 observational studies, Graham et al [30], Graham et al [31], and Veluvali et al [32] were rated “Good,” and Auster-Gussman et al [29] was rated “Fair.” Among the 5 uncontrolled pre-post studies, Everett et al [33] was rated “Good,” Zahedani et al [35] was rated “Fair,” and Sapre et al [26], Stein and Brooks [34], and Stephens et al [27] were rated “Poor.” Common limitations included lack of randomization or comparison groups, selection of highly engaged digital platform users, incomplete adjustment for confounding, attrition or outcome availability among subsets of enrolled participants, and short follow-up.

Results of Individual Studies

The randomized trial by Mathioudakis et al [28] provided the strongest comparative evidence (Table 3 and Multimedia Appendix 6). For the combined target of at least 4% weight loss and at least 150 minutes per week of physical activity, 12.6% (23/183) of AI-DPP participants and 12.4% (23/185) of human coach–led DPP participants met the outcome (risk difference 0.1%; 1-sided 95% CI –5.5%), supporting noninferiority. For at least 5% weight loss, the corresponding proportions were 16.9% (31/183) and 20.0% (37/185) (risk difference –3.1%; 1-sided 95% CI –9.7%), with the lower bound remaining within the prespecified –15 percentage point noninferiority margin. Among participants with prediabetes-range HbA_1c at baseline, in each arm, 26.9% (35/130) achieved at least 0.2% point reduction. Program initiation (171/183, 93.4% vs 152/185, 82.7%; risk difference 10.7%, 95% CI 4.2%-17.3%) and completion (117/183, 63.9% vs 93/185, 50.3%; risk difference 13.7%, 95% CI 3.6%-23.7%) were higher in the AI-DPP arm, but these engagement differences did not translate into superior clinical outcomes.

In the quasi-experimental study by Chadwick et al [25], AI-personalized messaging did not improve BMI compared with static messaging (between-group change 0.0 kg/m², 95% CI –0.6 to 0.5; P=.95) and did not increase the odds of achieving at least 30 minutes per day of physical activity (adjusted odds ratio 1.0, 95% CI 0.7-1.3; P=.74). Overall program completion was 96.9% (1048/1082). Because allocation was not randomized, residual confounding and selection bias remain possible.

The 4 observational studies reported uncontrolled weight changes. Across the 2 reports using percentage change, the median was 4.53% (range 4.1%-4.96%) [30,31]. Across the 2 reports using absolute change, the median was 2.8 kg (range 1.5-4.1 kg) [29,32]. Behavioral and engagement outcomes were not reported for these studies.

The 5 uncontrolled pre-post studies reported within-group changes without a concurrent comparison group. Two short-duration CGM-integrated studies (1 observational, Veluvali et al [32], and 1 uncontrolled pre-post, Zahedani et al [35]) reported increases in time in range and reductions in hyperglycemic events. These glycemic findings are hypothesis-generating because comparison groups were absent and the observed changes cannot be separated from regression to the mean, monitoring-related behavior change, or other uncontrolled factors. In the uncontrolled pre-post study by Everett et al [33], participants had a mean weight change of –1.6 kg, a median HbA_1c change of –0.1 percentage point, and a mean physical activity change of 2.8 metabolic equivalent of task hours per week. Across the 2 reports using percentage weight change, the median was 2.34% (range 2.3%-2.38%) [34,35]. Stein and Brooks [34] also reported changes in meal patterns. The 2 adolescent chatbot studies (Sapre et al [26] and Stephens et al [27]) primarily reported feasibility, responsiveness, perceived usefulness, or goal progress rather than clinical effectiveness. These nonrandomized findings do not permit causal inference.

Result of Syntheses by AI Type

As shown in Table 4, AI lifestyle coaching systems were the most frequently evaluated category (7 studies) and included the only randomized trial [25,28-31,33,34]. That trial supported noninferiority of 1 AI-DPP to a human coach–led DPP and found higher initiation and completion without superior clinical outcomes [28]. The remaining coaching system studies reported nonrandomized weight, behavioral, or engagement findings. Two predictive AI platforms integrated CGM and reported short-term within-group glycemic changes but neither used a concurrent comparison group, so causal attribution to the AI or CGM component is not possible [32,35]. Evidence for conversational AI chatbots was limited to 2 small feasibility or process studies focused primarily on engagement and perceived usefulness [26,27].

Result of Syntheses by Populations and Delivery Settings

Evidence was insufficient to determine whether effects differed by populations or settings (Table 4). The 2 adolescent studies reported primarily engagement outcomes and could not be compared with adult studies [26,27]. One uncontrolled study reported larger changes in time in range among participants with poorer baseline glycemic status, but no randomized subgroup comparison was performed [35]. One observational study reported similar changes in weight across rural and nonrural subgroups [31], whereas a nonrandomized study conducted in rural India found no advantage of AI-personalized messaging over static messaging in terms of BMI or physical activity [25]. No study reported adequately powered income-stratified or socioeconomic status–stratified treatment effects.

Reporting Biases

The ClinicalTrials.gov search did not identify any completed study. No formal reporting bias judgment was made because the evidence base was small, outcomes were heterogeneous, and registration was uncommon among commercial platform evaluations.

Certainty of Evidence

Certainty of evidence was not formally rated. The predominance of nonrandomized designs, limited replication, inconsistent outcome reporting, and potential commercial sponsorship bias reduced confidence in conclusions beyond the single randomized comparison.

Discussion

Principal Findings

This systematic review addressed 3 objectives: evaluating the clinical, behavioral, and engagement effects of AI-driven lifestyle interventions; determining whether effects differed by AI intervention type; and assessing whether effects differed by population or delivery setting. For the first objective, the evidence base was small and methodologically heterogeneous. The only randomized controlled trial supported noninferiority of 1 AI-DPP to a human coach–led DPP, but greater engagement did not produce superior clinical outcomes. A nonrandomized comparative study found no benefit of AI-personalized messaging over static messaging in terms of BMI or physical activity, and observational and uncontrolled pre-post studies reported weight, glycemic, behavioral, and engagement changes that do not support causal attribution. For the second objective, findings differed descriptively by AI type: coaching systems had the broadest evidence base, predictive platforms reported short-term glycemic changes only in uncontrolled settings, and conversational chatbot evidence remained limited to small feasibility studies. For the third objective, evidence was insufficient to determine whether effects differed by age, socioeconomic status, geographic setting, or baseline metabolic risk.

These findings extend the scoping review by Lalani et al [17], which described the breadth of AI apps in prediabetes but did not systematically appraise quality or organize effectiveness evidence by AI type and outcome domain. The current review’s single randomized controlled trial suggests that engagement and clinical effectiveness should, at least provisionally, be interpreted separately: higher initiation and completion in the AI arm did not confer superior clinical benefit [28]. This warrants revisiting as more evidence emerges, since a systematic review and meta-analysis of digital mental health interventions found the opposite, that greater engagement was generally associated with better outcomes [37]. Engagement’s value as a proxy for effectiveness may therefore be context-specific rather than fixed, and larger, adequately powered trials are needed before the dissociation seen here is treated as more than a feature of this 1 trial. Coaching systems have the broadest evidence base, but most supporting studies are nonrandomized. Predictive platforms have reported short-term CGM changes, but these observations cannot be distinguished from monitoring effects, regression to the mean, or concurrent behavior change [32,35]. Chatbot evidence remains limited to small feasibility studies [26,27]. AI-driven interventions may offer flexible and scalable delivery, but they also introduce barriers related to digital literacy, internet and device access, affordability, language, and cultural fit [38,39]. Health systems may reasonably evaluate these tools as adjuncts to established prevention services, particularly when implementation includes equity monitoring and human support, but the evidence does not support replacing standard DPPs or assuming that increased engagement will improve clinical outcomes.

These findings carry direct implications for how health systems deploy AI-driven tools alongside standard DPP delivery. Because access to in-person DPP remains uneven, particularly in rural and underserved communities, and participation is often constrained by program intensity, scheduling conflicts, and coach-participant cultural mismatch [13-15], and because DPP reach remains lowest among the populations at greatest risk [16], AI-driven platforms offer a plausible route to extend reach where implementation barriers, not intervention efficacy, are the limiting factor. The randomized evidence here supports this reach argument more than a clinical superiority argument: the AI-DPP achieved meaningfully higher initiation and completion than human coaching, which matters practically for programs struggling to enroll and retain high-risk populations, although it did not outperform coaching on clinical end points. For health systems and payers, this suggests that AI-DPP tools are best positioned as a scalable engagement layer that extends the reach of the Centers for Disease Control and Prevention’s National DPP model [9,10], a way to channel more at-risk individuals into some form of structured prevention programming, rather than as a substitute expected to outperform coach-led care on weight, glycemic, or activity outcomes. This distinction matters for real-world deployment decisions: because structured prevention programs remain concentrated in a limited number of countries and settings [12], a program evaluating whether to adopt an AI-DPP tool to reach a larger population should weigh it primarily against the counterfactual of no program participation at all, where higher initiation and completion are likely to translate into more people receiving some prevention benefit, rather than against human coaching alone. At the same time, because 10 of 11 included studies evaluated commercially developed platforms with limited independent replication, health systems should treat vendor-reported engagement and weight outcomes cautiously until independent, adequately powered trials, ideally across diverse populations and settings, confirm generalizability beyond the single existing randomized comparison.

Limitations

This review has limitations at both the evidence and review process levels. Most included studies were observational or uncontrolled and were susceptible to selection bias, confounding, regression to the mean, and outcome availability among selected subsets [40,41]. Follow-up was often short, and outcomes were measured and reported inconsistently. Most interventions were commercially developed, and several studies reported direct industry funding, employment, or authorship involvement. Because the NHLBI tools do not incorporate funding source, this potential favorable outcome and selective reporting bias was considered separately in interpretation and summarized in Table 2 [42]. Participants’ overlap across reports from the same commercial platforms could not be excluded. Restriction to English-language publications may have introduced language bias and reduced representation of evidence from non-Western settings. The initial screening and data extraction were performed by 1 reviewer (OIE) and verified by a second reviewer (TN) rather than being conducted independently in duplicate. The review was not prospectively registered, no formal protocol was prepared, and no formal certainty assessment was conducted. Meta-analysis was not performed because designs, estimands, and outcome metrics were not sufficiently comparable; consequently, differences across AI types or populations were descriptive rather than statistically tested. Rapid development of AI systems also means that findings may not generalize to newer large language model–based tools [43,44]. Future studies should use adequately powered randomized designs, prespecified subgroup analyses, standardized clinical and engagement outcomes, longer follow-up, transparent reporting of AI methods and commercial relationships, and independent evaluation in low- and middle-income countries and noncommercial settings.

Conclusions

The evidence does not establish clinical superiority of AI-driven lifestyle interventions over standard diabetes prevention care. One randomized trial supports noninferiority of a specific AI-DPP at 12 months; other findings are predominantly nonrandomized, commercially developed, and hypothesis-generating. To our knowledge, this is the first systematic review to stratify AI-driven diabetes prevention evidence by AI mechanism (coaching, predictive, and conversational) and by outcome domain, extending prior scoping work that mapped the breadth of AI apps in this space without appraising study quality or comparative effectiveness against standard DPPs. This design- and mechanism-stratified synthesis makes explicit where evidence is strongest (engagement and weight) and weakest (causally interpretable glycemic, subgroup, and generalizability effects), a distinction that broader digital health reviews pooling AI and non-AI interventions together have not drawn. For health systems and payers weighing AI-DPP adoption, this mapping offers a concrete, evidence-based basis for treating these tools as adjuncts that may extend reach and engagement, rather than substitutes expected to match or exceed standard DPP outcomes, while monitoring for equity effects as platforms scale. Effectiveness, sustainability, equity, and cost-effectiveness of AI-driven lifestyle interventions require independent, adequately powered randomized trials before broader adoption is warranted.