Prediabetes is a significant global health concern, with a worldwide prevalence of 9.1% (464 million people) based on impaired glucose tolerance in 2021 projected to rise to 10% (638 million people) by 2045 [1]. Similarly, impaired fasting glucose affected 5.8% of the global population (298 million people) in 2021, with an anticipated increase to 6.5% (414 million people) by 2045 [2]. In Spain, the age-adjusted prevalence of impaired glucose tolerance was 6.2% in 2021, affecting 2,600,800 adults, whereas impaired fasting glucose prevalence was 2.8%, impacting 1,231,400 individuals [3].

Lifestyle modifications have proven effective in preventing diabetes [4], particularly among individuals with risk factors such as overweight, a family history of the disease, or physical inactivity [5]. One of the most successful programs, the year-long Diabetes Prevention Program, reduced the incidence of type 2 diabetes mellitus (T2DM) by 58% over 3 years in its randomized controlled trial (RCT) [6]. However, barriers such as time constraints and low commitment often hinder patients with prediabetes from engaging in intensive interventions [7]. Digitally delivered, lower-intensity interventions offer a promising alternative by addressing these barriers [8]. Advances in digital health technologies such as wearables, mobile health (mHealth) apps, and data analytics enable personalized health education and real-time monitoring [9]. These innovations have the potential to enhance chronic disease management, including diabetes and cardiovascular conditions.

While some studies indicate that mHealth interventions may reduce the incidence of T2DM, improve glucose tolerance, promote weight loss, and lower glycated hemoglobin (HbA_1c) levels [10-12], recent systematic reviews highlight that the field is still in its early stages and the evidence remains inconclusive [13,14]. Moreover, the potential to enhance the effectiveness of mHealth interventions by combining them with educational cointerventions for health care professionals has not yet been explored.

To address this gap, we developed PREDIABETEXT (Prediabetes Text Message Digital Intervention for the Prevention of Type 2 Diabetes Mellitus), a novel, theory-driven, multifaceted digital intervention that contributes to existing research by integrating health care professional training with patient-centered interventions. By combining personalized SMS text messaging for patients with an online training program for health care professionals, our approach aimed to enhance diabetes prevention by supporting patient lifestyle changes while equipping health care professionals with the skills to effectively manage prediabetes in primary care.

Therefore, this study aimed to evaluate the effects of PREDIABETEXT on HbA_1c levels (primary outcome) and additional clinical, behavioral, and psychological outcomes through a 3-arm, 6-month cluster RCT.

The trial protocol is available elsewhere [15]. This trial was prospectively registered at ClinicalTrials.gov (NCT05110625) on November 8, 2021, before participant recruitment. The reporting of this study is based on the CONSORT (Consolidated Standards of Reporting Trials) 2010 statement extension to cluster RCTs [16] (Multimedia Appendix 1).

This study consisted of a pragmatic, 6-month, 3-arm cluster RCT conducted in primary care centers in Mallorca (Balearic Islands, Spain). Clusters were defined at the level of health care professionals working in these centers to account for variability in individual practice styles and patient interactions. Among the 58 health care professionals included, 16 centers were represented, with a median of 3 (IQR 1-6) health care professionals per center. To minimize contamination, access to the online training (intervention group B) was restricted through password protection. Randomization at the health care professional level ensured that practice-level contamination was unlikely as control health care professionals did not receive the training. This design helped maintain intervention fidelity at the health care professionals level while limiting the spread of intervention content to control health care professionals even within the same practice.

The sample size was estimated based on the primary outcome variable (HbA_1c levels). We planned to recruit 42 primary health care professionals and 420 patients (approximately 10 patients per health care professionals). On the basis of an expected SD of 0.8%, an intraclass correlation coefficient of 0.04 [22], a design effect of 1.36, and an anticipated dropout rate of 10%, a total of 420 participants (140 per study arm) was deemed sufficient to achieve 80% power to detect a between-group difference of at least 0.3% in HbA_1c levels. This effect size was selected based on previous evidence suggesting that even modest reductions in HbA_1c levels are associated with clinically meaningful improvements in diabetes risk and related complications [23]. A meta-analysis of 47 RCTs including 7677 patients demonstrated that self-management interventions resulted in a mean HbA_1c level reduction of 0.36% (95% CI 0.21-0.51) [24]. We designed our clinical trial to detect a 0.3% decrease in HbA_1c levels, which is a lower threshold than 0.5% as the minimum for clinical importance [25], to make sure that it was sensitive enough to identify even these slight changes.

Recruited physicians and nurses were assigned to 1 of 3 study arms using computer-generated random numbers. Their patients were invited to participate in the study and received the intervention according to the allocation of their health care professionals. Allocation concealment was performed at the level of health care professionals (clusters). Once baseline data collection for both health care professionals and patients was completed, the project manager (SM-M) assigned health care professionals (clusters) to intervention groups based on the pregenerated randomization sequence. To maintain allocation concealment, a centralized randomization process was implemented. Health care professionals were blinded to the allocation sequence before patient recruitment, ensuring unbiased assignment of participants. These measures ensured appropriate cluster formation and rigorous randomization, thereby maintaining the integrity of the study design. Notably, the randomization process was not stratified by any specific factors. Details regarding the randomization procedures are provided in Multimedia Appendix 4.

Due to the nature of the intervention, health care professionals and patients could not be blinded to group allocation. However, outcome assessors (ie, questionnaire administrators and laboratory staff) and statisticians were blinded to group assignment.

Descriptive statistics were used to examine the baseline characteristics of the participants. The effect of the intervention was evaluated by analyzing differences in primary and secondary outcomes between groups at the 6-month time point. All participants received the intervention as randomized, with no protocol deviations. To ensure robustness and avoid attrition bias, we used the intention-to-treat (ITT) approach, which includes all participants as originally allocated regardless of adherence or intervention completion. The missing values, outliers, and invalid data were handled using multiple imputations by chained equations. This process was repeated for 10 iterations, generating multiple imputed datasets. Final imputed values were obtained by averaging those from each iteration. Trace plots were used to assess stability and convergence across iterations. Standardized effect coefficients and their corresponding 95% CIs were computed. Between-group differences in HbA_1c levels and other outcomes were analyzed using generalized estimating equations (GEE), which accounted for the correlation between repeated measures within individuals. GEE assumes between-subject independence and within-subject dependence, modeling the within-subject correlation matrix accordingly. This method allowed for the evaluation of time and group effects on intervention outcomes while preserving the ITT framework. The clustering effect was addressed using an appropriate working correlation structure. All statistical analyses were conducted using SPSS (version 25; IBM Corp) and Stata (version 16; StataCorp), with statistical significance set at 5%.

After completing the intervention, we conducted individual interviews with the participants to explore their experiences with the PREDIABETEXT intervention. We contacted them via phone call, with up to 3 call attempts if they were unanswered. Patient interviews took place after the 6-month intervention period and no later than 1 month from the completion of the intervention.

Health care professionals who received the cointervention were interviewed within a month of completing their training. They were purposefully sampled for diversity in gender, age, and professional category (nurse or physician).

Two dynamic topic guides for patients and professionals (Multimedia Appendix 6) were used, encompassing a global evaluation of the intervention’s utility and satisfaction alongside key themes relevant to the development of digital solutions for chronic disease management. The guide was iteratively refined as data collection progressed. Supportive prompts were used only when necessary to elicit responses aligned with the study objectives. All interviews were conducted via telephone by 2 trained members of the research team (RZ-C and SM-M), both of whom were independent from the participants’ clinical care and had no previous relationship with them. The interviews, lasting approximately 15 to 40 minutes, were audio recorded, transcribed verbatim, and anonymized. Data were collected until thematic saturation was reached, meaning that no new pertinent themes emerged from participants who gave their consent. Data analysis was managed and conducted using the NVivo software (version 12; QSR International), which facilitated coding and theme development in line with the approach by Clarke and Braun [35]. Throughout this study, 2 researchers (SM-M and RZ-C) worked together to analyze the interviews. The researchers critically assessed their own role and the study’s objective to avoid influencing participants’ responses.

All procedures were approved in June 2021 by the Primary Care Research Committee and the Balearic Islands Research Ethics Committee (reference IB4495/21PI). This study complied with the ethical principles outlined in the Declaration of Helsinki.

To ensure data confidentiality and integrity, all participant information was anonymized before analysis, securely stored on encrypted servers at the Health Research Institute of the Balearic Islands, and accessed only by authorized members of the research team under strict confidentiality agreements. Data protection measures adhered to national and European regulations, including the General Data Protection Regulation.

Consent from the participants was obtained, and they were informed of their right to withdraw at any time. No financial compensation was provided; participation was entirely voluntary. All information collected was treated confidentially and anonymized, and ethical oversight was maintained throughout the study.

The participants’ flow through the study is illustrated in the CONSORT flowchart in Figure 1. Recruitment of health care professionals began on October 20, 2021, and concluded on June 1, 2022. Of the 172 professionals invited, 58 (33.7%; n=30, 52% physicians and n=28, 48% nurses) from 16 centers consented to participate and completed the baseline telephone interview, yielding a recruitment rate of 43.6% (58/133). Of these 58 professionals, 18 (31%) were allocated to intervention group A, 20 (34%) were allocated to intervention group B, and 20 (34%) were allocated to the control group. Notably, there were no dropouts among health care professionals during the 6-month intervention period.

Patient recruitment took place between November 17, 2021, and August 2, 2022. A total of 365 patients were enrolled. Randomization occurred at the health care professional level, with each health care professional (cluster) assigned to 1 of 3 groups: intervention group A (n=106, 29% of the patients), intervention group B (n=140, 38.4% of the patients), or the control group (n=119, 32.6% of the patients) according to the health care professional’s allocation. Postintervention data were collected at the 6-month time point from 78.4% (286/365) of the participants, resulting in an attrition rate of 21.6% (79/365). The distribution of participants who completed the 6-month period was as follows: 28.7% (82/286) from intervention group A, 39.2% (112/286) from intervention group B, and 32.2% (92/286) from the control group. The primary reasons for attrition included inability to reach participants via phone or failure to attend in-person appointments. Missing values for various outcomes were primarily due to these reasons or because some participants completed the questionnaires but did not attend para-clinical assessments.

The health care professionals had a mean age of 49.69 (SD 10.15) years, with most being female (48/58, 83%). On average, participants reported 23.59 (SD 9.22) years of professional experience in health care.

Table 1 presents baseline patient demographics and clinical characteristics summarized at the level of 58 health care professionals (clusters) allocated to the control group (n=20, 34% professionals; 119/365, 32.6% patients), intervention group A (SMS text messaging only; n=18, 31% professionals; 106/365, 29% patients), and intervention group B (SMS text messaging+training; n=20, 34% professionals; 140/365, 38.4% patients). The average percentage of female patients per professional was 42.02% (SD 22.89%) in the control group, 49.06% (SD 23.72%) in intervention group A, and 45.71% (SD 18.05%) in intervention group B. The mean patient age per professional was 60.75 (SD 3.40) years in the control group, 58.66 (SD 2.78) years in intervention group A, and 59.84 (SD 3.83) years in intervention group B, with an overall mean of 59.79 (SD 3.50) years. These results indicate a broadly comparable demographic profile across clusters at baseline, supporting the validity of between-group comparisons in subsequent outcome analyses.

Patient-level baseline characteristics are summarized in Multimedia Appendix 7. The mean age of the patients was 59.79 (SD 9.75) years, and 54.5% (199/365) were female. Full details of the baseline characteristics have been reported elsewhere [36].

The intervention, consisting of SMS text message delivery, was implemented between February 21, 2022, and January 13, 2023. Over the 6-month intervention period, 246 individuals with prediabetes (n=106, 43.1% in intervention group A and n=140, 56.9% in intervention group B) received 72 intervention messages (3 per week) across 7 staggered cohorts initiated as participants were recruited. In addition, 2 reminder messages—one for the baseline visit and one for the final visit—were sent to each participant, bringing the total to 74 SMS text messages per participant.

A total of 18,204 SMS text messages were sent during the trial, of which 186 (1.02%) failed to be sent correctly. A total of 13.8% (34/246) of the participants missed at least one intervention message; however, no participant received <50% of the intended messages (<37 SMS text messages). Overall, 94.3% (232/246) of participants received at least 95% of the planned messages (≥70 SMS text messages). The messaging platform’s performance was monitored throughout the study to ensure consistent and reliable delivery.

Raw data for the primary and secondary outcomes in intervention groups A and B and the control group at baseline and the 6-month time point are available in Multimedia Appendix 8. At 6 months, the mean HbA_1c level was 6.07% (SD 0.29%) in intervention group A, 6.12% (SD 0.25%) in intervention group B, and 6.18% (SD 0.79%) in the control group (Figure 2).

The ITT analysis showed no statistically or clinically significant differences in HbA_1c levels between the 2 intervention groups and the control group. Intervention group A was associated with a small, nonsignificant reduction in HbA_1c levels (β=−0.05, 95% CI −0.21 to 0.10; P=.50), whereas intervention group B also showed a nonsignificant reduction (β=−0.04, 95% CI −0.12 to 0.10; P=.56; Table 2).

The risk of progression to diabetes is shown in Multimedia Appendix 10. The GEE model for the risk of progression to diabetes showed a lower risk of progression to diabetes for both intervention groups A (adjusted odds ratio 0.51, 95% CI 0.12-2.21) and B (adjusted odds ratio 0.49, 95% CI 0.12-1.94) versus the control group, but this effect was not statistically significant.

Among the 18 health care professionals in intervention group B who completed the online training cointervention, the knowledge score immediately after completing the course was significantly higher (mean 12.17, SD 2.20) compared to baseline (mean 6.83, SD 1.33). However, after the 6 months of the intervention, the knowledge score in intervention group B (mean 8.89, SD 3.92) did not significantly differ from that in intervention group A (mean 6.33, SD 2.06) or the control group (mean 7.09, SD 1.37), neither of which received the educational intervention (full details are available in Multimedia Appendix 11).

This 6-month pragmatic RCT conducted within primary health care settings as part of routine practice evaluated a multifaceted digital lifestyle intervention grounded in multiple behavior change theories. The intervention did not significantly reduce HbA_1c levels, the primary outcome, when compared to the control group receiving usual care.

The PREDIABETEXT intervention, carefully co-designed with end users, offers several strengths. Clinically, its scalability at a low cost makes it a sustainable option for health care institutions. To address the low participation and retention rates often observed in high-intensity, in-person programs such as the Diabetes Prevention Program, we opted for a low-intensity intervention. In addition, delivering PREDIABETEXT within a real-world setting leveraging the resources of the Balearic Islands health service (eg, secure access to electronic health records, SMS text message communication, and the Moodle platform for online training) provided 2 key benefits: a more accurate assessment of the intervention’s impact in routine clinical practice and the potential for its immediate integration into the health system if proven effective. This approach is in line with that of similar studies reporting that the use of remotely administered interventions was convenient for participants with prediabetes [37].

The intervention did not demonstrate a statistically significant improvement in glycemic control. Although there was a reduction in HbA_1c levels and FPG values for intervention groups A and B compared to the control group, this effect was not statistically significant. Previous literature in this area presents heterogeneous results. A recent systematic review evaluating the efficacy of remotely administered lifestyle interventions for preventing T2DM found that 6 out of 8 interventions resulted in significant reductions in weight or glycemic biomarkers such as HbA_1c and fasting glucose [37]. A long-term study of an internet-based diabetes prevention program demonstrated a reduction in HbA_1c levels by a mean of 0.38% (SD 0.07%) after 1 year and 0.43% (SD 0.08%) after 2 years [38], supporting the notion that both multimodal and long-term interventions can yield meaningful results. Furthermore, a systematic review and meta-analysis showed that digital health interventions resulted in a 12% risk reduction in T2DM incidence (risk ratio=0.88, 95% CI 0.77-1.01; I²=0.6%; P=.06) compared to the control group [39]. Consistent with our findings, a systematic review reported a controversial effect of digital health interventions on HbA_1c and FPG values [14]. In addition, a study in Spain evaluating the effects of a low-intensity nurse-led lifestyle intervention on glycemic control in individuals with prediabetes did not find statistically significant differences in FPG at 9 months [40]. Several factors may explain the results of our intervention. First, it was low intensity and relatively short in duration (6 months), whereas previous studies suggest that more comprehensive, long-term interventions may yield greater reductions in diabetes risk. Second, participants had relatively mild metabolic alterations at baseline, and adherence to behavior change recommendations may have varied, potentially affecting outcomes. Finally, while the study was sufficiently powered to detect changes in HbA_1c levels, it may have been underpowered to detect small but meaningful reductions in diabetes incidence.

The findings of our study showed no significant beneficial effect on triglyceride, cholesterol, HDL, or LDL levels. These results align with those of a systematic review and meta-analysis that reported minimal effects on lipid parameters in participants with prediabetes using digital health interventions [14].

Regarding the anthropometric measures of study participants, similar to our study, a long-term study of an internet-based diabetes prevention program showed no significant effect on body weight after 1 or 2 years of follow-up [38]. In contrast, the findings of a systematic review by Jeem et al [14] showed a positive effect of mHealth interventions on anthropometric measures such as body weight, waist circumference, and BMI among middle-aged and older patients with prediabetes. A meta-analysis estimated the overall effect of stand-alone eHealth interventions on T2DM prevention to be a mean weight loss of −3.34% (95% CI −4% to −2.68%) from baseline to up to 15 months of follow-up across all interventions [41]. Another meta-analysis that evaluated the effect of technology-mediated diabetes prevention interventions on weight showed a pooled weight loss effect of 3.76 kg (95% CI 2.8-4.7; P<.001), as well as improvements in glycemia in patients with prediabetes [42].

Regarding cardiovascular outcomes, the ITT analysis did not reveal any significant differences between the groups in systolic and diastolic blood pressure or the REGICOR-Framingham score. Previous systematic reviews and meta-analyses have focused primarily on measuring blood pressure and have not assessed cardiovascular risk scores. In this regard, our trial provides new data. The REGICOR-Framingham score is a highly relevant outcome measure for evaluating the success of interventions aimed at preventing T2DM due to its strong association with cardiovascular disease, a major complication of T2DM. Patients with T2DM are at significantly elevated risk of cardiovascular disease, with relative risks approximately 1.8 to 3.3 times higher than those for individuals without diabetes [43]. According to a study a reduction of 1% to 2% in the 10-year cardiovascular risk is generally considered clinically significant [44]. Thus, accurately assessing and reducing cardiovascular risk, which can be achieved through the REGICOR-Framingham score [45], not only supports T2DM prevention efforts but also addresses the leading cause of morbidity and mortality in this population. Future research should explore higher-intensity, personalized digital health interventions; extend the intervention period; and test strategies to enhance patient engagement and adherence.

One of the strengths of our study is the use of validated questionnaires, which enhanced both the internal and external validity of our findings. The use of blinded outcome evaluators strengthened internal validity, whereas the representativeness of the recruited patients, reflecting the general population with prediabetes of the Balearic Islands, improved external validity.

However, this trial has several limitations. Information and recall biases may have been introduced by measuring physical activity and eating habits through self-reported data. To mitigate these biases, objective measures such as pedometers, 24-hour dietary recalls, or food frequency questionnaires could have been used to provide a more accurate assessment.

Another potential limitation is the choice of assessment tool for cardiovascular risk. While measures such as the American College of Cardiology/American Heart Association pooled cohort equation may provide broader insights, especially in diverse populations, we selected the REGICOR-Framingham score due to its strong association with cardiovascular disease. This equation has been specifically adapted and validated for the Spanish population, making it particularly suitable for evaluating intervention effectiveness in our setting. Future research should consider incorporating more contemporary and widely applicable risk assessment tools to enhance generalizability across different populations.

Although multiple secondary outcome comparisons were conducted without formal statistical adjustment, most did not show significant changes. Nevertheless, the findings should be interpreted with caution due to the increased risk of type I error.

The Hawthorne effect may have influenced the results, potentially leading to an underestimation of the intervention’s effect. Participants in the control group may have altered their behavior simply because they were aware of being part of a study, thereby narrowing the differences between the intervention and control groups.

Moreover, the generalizability of the study may be limited by the exclusion of non–Spanish-speaking individuals and minority groups. In addition, hard-to-reach populations were likely underrepresented as one key eligibility criterion required participants to have a recent test result. Socially vulnerable individuals, who are more likely to develop T2DM but less likely to visit their primary care physician, may not have been adequately represented in this study. Future research should explore and implement recruitment strategies that more effectively engage these underserved groups. This study was conducted in Mallorca, Spain, which might limit the generalizability of the results to other populations or settings. Furthermore, the duration of the study was 6 months, which might be too short to capture the long-term effects of the intervention. Evaluating HbA_1c levels at 6 months offers useful short-term insights; however, extending the intervention period in future research could provide a more comprehensive understanding and help mitigate the effects of regression to the mean, ensuring more reliable long-term conclusions.

Another limitation at the health care professional level was that our 16-hour training program might be too lengthy, potentially leading to time constraints in applying the training. Moreover, to enhance knowledge retention, future studies should consider incorporating reinforcement sessions at 1- to 3-month intervals as strategic reminders. In addition, we prioritized health care professionals with a higher number of patients meeting the prediabetes criteria for inclusion. Although this strategy might introduce selection bias, it is unlikely to have affected our findings.

Finally, the cost-effectiveness of the PREDIABETEXT intervention was not assessed in this study. Future research should include a comprehensive cost-effectiveness analysis to evaluate the scalability and sustainability of the intervention.

In conclusion, this multifaceted digital intervention for the prevention of T2DM in the primary care setting did not significantly improve glycemic control in individuals with prediabetes who received SMS text messages or those whose health care professionals also received training compared to the usual care control group. On the basis of these findings, future research should prioritize the development and evaluation of higher-intensity digital health interventions that include more frequent interactions and personalized assignments to enhance participant engagement. Wearable devices offer real-time monitoring and personalized feedback, which can improve patient self-management and intervention outcomes. These interventions may be more effective when combined with intensive, multimodal approaches delivered simultaneously to at-risk populations. In addition, expanding research across multiple countries and focusing on the recruitment of underserved groups will improve the generalizability of the findings. Conducting real-world evidence studies to evaluate the effectiveness of digital health interventions in diverse populations and settings using electronic health records and other real-world data sources to assess outcomes is also recommended.

Furthermore, applying implementation science frameworks to study the integration of digital health interventions into routine clinical practice is recommended. This approach will help identify facilitators of and barriers to implementation and develop strategies to enhance the adoption and scalability of these interventions.

By addressing these priority research questions and using the recommended methodologies, future studies can advance the field of digital health interventions for diabetes prevention, ultimately improving health outcomes and informing health care policy and practice.