The prevalence of prediabetes is approximately 10% in the UK population [1], with higher rates in adults aged 40 years or older and those of South Asian or African Caribbean ethnicities [2]. Considering the rising prevalence of type 2 diabetes, which accounts for disproportionate and increasing costs to the individual, society, and health care systems globally [3,4], primary prevention is a current public health priority.

Landmark randomized controlled trials (RCTs) have repeatedly shown that intensive face-to-face diabetes prevention programs (DPP) are effective in reducing the risk of type 2 diabetes by approximately 50% [5-7]. The implementation of these landmark studies is expensive, time consuming for the patients and for health systems, and the uptake is often by those with the lowest risk for type 2 diabetes [8]. One solution is adapting the traditional DPP into a web-based DPP using wearable technologies and web-based programs. These are increasingly available from commercial providers at a low cost to the patient or health commissioners, but the evidence of their effectiveness in reducing the risk of type 2 diabetes is limited [9,10]. For example, in a recent RCT of 2062 people in India and the United Kingdom with impaired glucose tolerance, those who received 2 to 3 weekly SMS texts providing lifestyle advice did not have a significant reduction in diabetes conversion (defined by international criteria for fasting plasma glucose or hemoglobin [HbA_1c] levels) compared with controls who received standard lifestyle advice at baseline only [11]. Similarly, a pilot pre-post noncontrolled study of a 16-week web-based DPP was associated with reduced weight and lower HbA_1c levels, improvements that persisted at the 2-year follow-up [12]. An RCT of a fully automated and algorithm-driven email, web and mobile DPP found significant improvements in biomedical outcomes, weight, and diabetes risk at 6-months compared with the wait list control group [13].

Components of behavior change techniques considered to be most effective in improving diet and physical activity (PA) are based on self-regulatory behaviors, such as goal setting, self-monitoring, giving feedback, utilizing social support, and motivational interviewing (MI) [14,15]. Interventions based on a psychological theory, for example, the theory of planned behavior, are considered to be more effective and have better outcomes in high-risk populations [16], although the minimum threshold of intensity, such as the number of sessions or messages or duration of the intervention, is not known [14,17,18].

This RCT aimed to test the feasibility of a web-based DPP, consisting of a wearable technology that records PA, integrated with SMS texts based on MI techniques, and lifestyle education delivered via a smartphone app, over 12 months in participants with prediabetes. The primary aims were to assess (1) the potential size of the study population; (2) the proportion of those who consented to be screened for eligibility; (3) the proportion of those who were screened and who were eligible, consented, and randomized; (4) the proportion of those who were randomized and who completed the intervention; and (5) the proportion of those who completed the 6-month and 12-month follow-ups.

Our secondary aims were to measure the change in biomedical outcomes, including reducing weight and increasing PA, to inform the possible range of effect sizes and obtain outcome variance estimates required for sample size calculations in a full-scale trial.

This was a two-arm, parallel, single-blind RCT conducted over 12 months. The trial has been reviewed and given favorable opinion by the London City and East Research Ethics Committee (16/LO/1505).

We recruited patients from 3 clinical commissioning groups in London (Lambeth, Southwark, and Lewisham), which comprise a population of 912,687 residents, with one of the highest prevalence rates of type 2 diabetes [19] in England and with broad socioeconomic and ethnic diversity [20]. Patients were recruited from participating primary care surgeries with list sizes greater than 6000 patients.

Patients with an HbA_1c level of 39 to 47 mmol/mol were defined as being in a prediabetes state according to the current American Diabetes Association criteria [21]. They were identified using a 2-stage process. First, patients at high risk of developing type 2 diabetes were identified by the general practitioner (GP) who conducted a search on the Egton Medical Information Systems (EMIS) web (clinical software where GPs can access patient health records) using HbA_1c results recorded in the previous 12 months. Additional data extracted from EMIS included anonymous ID, gender, date of birth, postcode, ethnic origin, QDiabetes score [22], and BMI. Potentially eligible patients (HbA_1c level of 39 to 47 mmol/mol) were invited to undergo screening for eligibility. All GPs used the same postal invitation. Second, the inclusion criteria at screening were as follows: adults aged between 18 and 65 years; BMI≥25 kg/m² (≥23 kg/m² if of Asian ethnicity) [19]; permanent residents in Lambeth, Southwark, or Lewisham; owning a smartphone (iPhone or Android models only)—defined as logging on at least once per day to the internet; being fluent in conversational English; and being ambulatory (eg, capable of walking without mobility support equipment).

The exclusion criteria included diabetes (not including past history of gestational diabetes); pregnancy; planning a pregnancy or lactating during the duration of the study; severe mental illness (severe depression with suicidal ideation, psychosis, bipolar affective disorder, dementia, learning difficulties, substance problem use, or dependence); severe physical disability (eg, that would prevent any increased uptake of physical exercise); advanced active disease, such as cancer or heart failure; any other condition that requires glucose-altering drugs; BMI≥50 kg/m²; and current participation in a weight loss program or DPP. When in doubt, we sought GP confirmation of eligibility.

We collected sociodemographic data, including age, gender, postcode of residence, employment status, educational level, and self-reported ethnicity. On the basis of the participant’s postcode, we determined their indices of multiple deprivation (IMD) 2015 rank, which indicates the relative level of socioeconomic deprivation in their area [23].

The intervention was based on the theory of planned behavior, which states that to change behavior, people need to form an intention [16]. Intention formation is influenced by 3 constructs: expected value or positive attitude (people see the value in making the change), subjective norm (significant others and peers also value the change), and self-efficacy (people believe they are capable of making the change).

For feasibility parameters, the primary outcomes were proportion recruited and randomized and proportion followed up. The primary clinical outcomes were change in weight (kg) and PA (mean steps per day) from baseline to 12 months, with an interim measure at 6 months. Follow-up PA was the mean step count of the 7 days of wear leading up to and including the day of the follow-up appointment.

The secondary outcomes were a change in HbA_1c levels and BP at 6 and 12 months and waist circumference, waist-to-hip ratio, and lipid levels at 12 months. The HbA_1c level was analyzed as a continuous and categorical variable, with the following categories: normal (<42 mmol/mol), prediabetes (39-47 mmol/mol), and diabetes (>47 mmol/mol).

We aimed for 100 participants per arm, as this was a large enough sample to inform the practicalities of delivering the intervention, recruitment, uptake, and attrition to inform a full-scale trial rather than measures of intervention effects.

Before randomization, participants wore the Buddi wristband for 1 week to familiarize with the technology and to collect baseline activity data. All patients were offered a brief educational session on the use of the Buddi device, and an instruction manual was provided. Randomization of participants was conducted by the data manager from an independent clinical trials unit using computer-generated randomization blocks of random sizes and stratified by surgery in a 1:1 ratio. Allocation concealment of the randomization list was held in a password-locked computer. This was an open-label study, but outcome assessors, laboratory technicians, and researchers entering and scoring the data were blinded to patients’ allocations.

The full statistical analysis plan for this study is provided in Multimedia Appendix 2. Analysis and reporting were in line with the consolidated standards of reporting trials guidelines [29], including its extensions for pilot and feasibility trials; primary analyses were conducted on an intention-to-treat basis and using a two-sided significance level of .05. Statistical analyses were mainly descriptive, aiming to provide estimates of key feasibility parameters and to inform power calculations for a future definitive trial. Descriptive subanalyses were used to explore participation rates among participants based on ethnicity, education level, IMD 2015 score, BMI, depressive symptoms, readiness to change, and self-efficacy. The proportion of missing data for individual items and measures was examined to determine the suitability of instruments and the level of burden for a future full-scale trial. We compared baseline characteristics of (1) those who were eligible and who did and did not consent and (2) those who did and did not provide 12-month follow-up data for the primary clinical outcomes. The differences in treatment effect for the primary and secondary outcomes between the arms at 6- and 12-month follow-ups were analyzed using analysis of covariance–based, linear mixed effects models with prerandomization values as a covariate [30]; STATA’s mixed command was used for estimation.

Figure 1 presents participant flow through the trial. Over a 6-month period, we sent postal invitations to 142 surgeries, whose patient list sizes were over 6000. A total 21 of 142 surgeries (14.8%) agreed to participate, from which 194,892 patients had the first-stage screening performed electronically by their GP. Out of these, 5124 were potentially eligible and invited for further screening by their GP by letter and telephone. The number of people who attended and consented for second-stage screening was approximately 11 patients per week, requiring 28 hours per week of a full-time equivalent research worker. Approximately half of those who attended the screening were eligible and consented to be randomized.

The baseline characteristics of the participants are presented in Table 1. The two arms were broadly similar (except for ethnicity) and representative of the local catchment in terms of high levels of deprivation, with over half being of nonwhite ethnicity.

Data were available for 192 participants (92 in the intervention arm). Of these 192, a total of 80 (80.0%) and 60 (65.2%) participants were adherent to the intervention in the intervention and control arms, respectively (X²₁=4.6; P=.03). Participants used the smartphone app on average 41.5% (SD 27.5) and 33.2% (SD 27.5) of days in the control and intervention arms, respectively.

Table 2 presents a descriptive summary of the clinical outcomes at each time point. We observed that the distributions for each primary outcome were positively skewed; however, log-transforming the data did not improve the distribution or alter the results of the analyses below. The pooled SD at 12 months for weight was 15.43 kg and for PA was 3588.76 steps.

When categorizing participants’ metabolic status based on their HbA_1c values, 3 (3.5%) and 5 (7.3%) in intervention group and 2 (2.2%) and 6 (6.9%) in the control group met the cut-off for diabetes (>47 mmol/mol) at 6 and 12 months, respectively. Participants with HbA_1c >47 mmol/mol were referred to the GP. Participants were informed of their results via telephone and told to contact their GP. A total of 4 (4.7%) and 9 (10.1%) participants returned to the normal range (<39 mmol/mol) in the intervention and control arms, respectively, at 6 months. At 12 months, 0 and 4 (4.6%) participants were in the normal range in the intervention and control arms, respectively. There was no difference in the proportion of participants who developed type 2 diabetes or who returned to normal HbA_1c levels between the 2 groups at 6 months (X²₂=2.0; P=.36) or 12 months (X²₂=3.2; P=.20).

The fixed and random effects of the mixed effects models for weight and PA are presented in Multimedia Appendix 3. There was no treatment effect on weight or PA at 6 or 12 months (Table 2), or for HbA_1c levels, waist circumference, waist-to-hip ratio, lipid levels, or BP at 12 months.

Our sensitivity analyses adjusting for baseline characteristics did not alter our conclusions of the primary outcome analysis above. Readiness to change (URICA stage) did not moderate the treatment-by-time effect for weight (F_1,2=0.10; P=.90) or PA (F_1,2=0.56; P=.57).

Our analysis of responders to the control or intervention arm indicated that improvements in weight and/or PA were associated with baseline smoking status (X²₂=11.6; P=.003) and PHQ-9 categorical score (X²₄=10.7; P=.03; Multimedia Appendix 4).

Testing the associations between IPAQ scores and mean step counts revealed a significant positive correlation between the intervention arm’s IPAQ total activity score and step counts at baseline only (r₈₅=0.22; P=.04; Multimedia Appendix 5). The series of 1-way analysis of variances showed that the IPAQ activity category (low, moderate, or high) was only associated with step counts for the intervention arm at baseline (F_2,84=4.83; P=.01).

We successfully reached our target sample size in 36 weeks, with a recruitment rate of 5 to 6 patients per week per full-time research worker. For full-scale RCT testing the intervention with an expected effect size of 0.1 and 90% power, we estimate recruitment will take about 3.5 years per research worker or just over 1 year for 3 research workers (Multimedia Appendix 6).

We piloted the statistical plan that we would anticipate using for a full-scale RCT, which appeared to be valid, although hypothesis testing was only a secondary aim. The only significant difference observed was that those in the intervention group had lower PA levels at 6 months compared with those in the control group, but they were also less likely to complete the intervention and more likely to be noncontactable or have withdrawn at follow-up.

An important difference between our study and previous RCTs [11] is that we showed that it is feasible to use a wearable technology that continually records levels of PA (ie, step count). Furthermore, our study included male and female participants compared with another study that included only male participants [31]. We also had higher rates of retention than those in other studies. For instance, in an RCT of a fully automated and algorithm-driven web-based DPP [13], 72% of intervention participants were still interacting with the program at 6 months. We achieved 87% of participants attending their 6-month follow-up.

A key strength is that this is the first feasibility study in the United Kingdom, which tests an automated web-based DPP designed to mirror the landmark face-to-face DPPs. The sample was representative of the ethnic and social diversity common in prediabetes. We noted that there was variation by GP surgery in response, and this suggests that clustering should be considered in any full-scale RCT.

An important observation is the importance of sustaining the functionality of the technology. Despite a prior road-test of the wearable technology and its connectivity to the smartphone app by the commercial provider in a handful of volunteers, there were mechanical and technical failures that may have resulted in participants randomized to the intervention arm receiving less of the required dose and withdrawing because of these issues. For example, we observed a sharp decrease in the mean step counts for participants at 6 months and 12 months compared with baseline, suggesting that the wristband may not have been accurately recording PA over time or that participants were unaware of ongoing technical issues. Other explanations could be that participants reduced their step count over time despite being in the intervention group or that participants reduced the amount of time they wore the device.

Participant adherence to the intervention could not be comprehensively documented as the technology for this did not exist; therefore, we were not able to capture how the participants used the app, particularly if they accessed the DPP education materials (and for how long) or if they read the SMS messages sent to them.

The follow-up rates for the primary outcomes were, on average, 75%, and predictors of missing data included higher BMI, presence of depressive symptoms, and current smokers (ie, those more at risk of developing type 2 diabetes and increasing the risk of underestimation).

We assessed the degree to which the wristband-derived step counts (ie, objectively measured PA) were associated with a validated self-reported measure of PA levels. We observed that the 2 measures only corresponded to a limited extent in the control arm, with step counts being weakly and positively correlated with total self-reported activity at 2 time points. The self-reported scores were not associated with step counts in the intervention arm. There are several possible reasons for this discrepancy, one being that these participants disengaged from the intervention (ie, wearing the wristband because of the technical problems or the higher intensity of messages).

The responses of the participants to our implementation processes questionnaire were generally favorable of the intervention; however, they did note several key areas for improvement for a full-scale RCT. These included improving the wristband’s clasp, offering the wristband in different styles, and improving the app’s accessibility (eg, offering the educational material in other digital formats) and formatting (eg, improving the readability and precision of the activity graphs and personalizing the content and frequency of messages).

This study found that there is a sufficiently large target population of patients for screening and a reasonably good participation rate (ie, patients are keen to receive support for diabetes prevention). Ensuring an optimum balance in the intensity of information sent and the functionality of the technology are potential key components to consider for a full-scale RCT.