For GODART, we developed an automated behavior monitoring system (ABMS) where participants receive phone calls to log their exercise, medication, and diet behavior each day for 24 weeks. Calls are delivered through IVR and transcribed using AI features available in Twilio software (Twilio Inc), removing the need for a human operator and allowing participants to report their daily behavior through natural speech. Rewards for completing calls are issued either at a flat rate or at an adaptive (gamified) rate. ABMS calls are supplemented by either human or automated IVR-delivered health coaching once per week. GODART’s theoretical framework, based on the SCT for behavior change, is outlined in Textbox 1, and an overview of GODART’s intervention structure is detailed in Figure 1.

This study was approved as expedited by the University of Alabama at Birmingham (UAB) Institutional Review Board under record IRB-300008752 and registered with ClinicalTrials.gov (NCT05344859). Informed consent was obtained for all participants either electronically using REDCap (Research Electronic Data Capture; Vanderbilt University) or, if needed, through telephone or a mailed physical copy. Participant data are deidentified to maintain privacy and confidentiality. Participants were compensated up to US $225.80 for their participation in study monitoring calls and attending study visits.

The GODART study uses the MOST (Multiphase Optimization Strategy) design, grounded in resource management principles and continuous improvement [36-38]. This study aligns with the preparatory phase of MOST. We are applying factorial experimentation as part of this MOST preparatory phase to assess the feasibility of two different intervention components (experimentations), with two levels in each group, yielding four experimental conditions. The Technology Experimentation component (Figure 1) focuses on comparing automated coaching through IVR with traditional human-delivered coaching. The goal here is to assess the feasibility of optimizing health coaching by leveraging accessible technology and assessing the preliminary effect when compared to human coaching. The Resource Experimentation focuses on assessing the feasibility and preliminary effect of the adaptive (gamified) rewards when compared to the fixed rewards program. Upon completion of the intervention, all participants will be invited to complete an exit interview to gauge expectations and perceived benefits and to collect suggestions for protocol improvement. Study findings will help us assess the feasibility of the protocol, acceptability of the intervention components, and their preliminary effectiveness in reducing HbA_1c, leading to a confirmatory optimization trial. Figure 2 shows a schematic diagram of the study flow.

This study aimed to recruit a total of 88 patients diagnosed with T2DM [39]. Eligible participants must be individuals aged 18 years or older, can speak English, and have HbA_1c between 6.5% and 11.5%. English language proficiency was included as an eligibility criterion, as the telephone-based intervention is delivered in English and requires a level of comprehension sufficient for participants to understand the instructions and interact meaningfully with the system. Additionally, due to resource constraints and the pilot nature of this study, we limited recruitment to English-speaking individuals to ensure feasibility and optimize initial implementation. Participants with HbA_1c within the range of 6.5% to 11.5% were included in the study to ensure that we target individuals who will benefit from improved self-management and behavior modification.

Exclusion criteria include current enrollment in other diabetes and weight management interventions, speech or hearing impairments, a major cardiac event in the 6 months prior to screening, current or imminent pregnancy, and ongoing use of insulin. Additionally, participants are excluded if they are unwilling to accept random assignment, engage in regular intervention-related phone calls, or complete study procedures. Individuals currently enrolled in other diabetes or weight management interventions are excluded to prevent confounding effects and to isolate the impact of the study interventions. As the intervention relies on telephone-based communication, individuals with hearing or speech impairments are not included. Participants with major cardiac events, ongoing insulin therapy, or current or imminent pregnancy are also excluded due to safety concerns and the need for more intensive or specialized diabetes management.

Participants were identified and recruited using the National Institute of Health–funded Integrating Biology and the Bedside (i2b2) system implemented at the UAB Health System. This self-service tool allows users to build a cohort of patients by querying clinical data and developing a recruitment database for the study. Recruitment occurred from patients across the health system, which included community clinics that serve patients with Rural-Urban Commuting Area codes of ≥4, which includes all nonmetropolitan tracts, specifically micropolitan, small towns, and isolated rural areas (ie, large towns or clusters) [40]. Using the recruitment database, opt-out letters [41] were sent to all residences of potentially eligible patients either via mail or email. The opt-out letter informed individuals about the study and allowed them to decline contact regarding the study by calling or sending an SMS text message to the phone number provided. Only those who did not opt out within 2 weeks were contacted regarding their interest in participating in the study. Additional recruitment strategies included posting study flyers in high-traffic and clinical areas among UAB clinics and collaborating with clinical staff and providers to identify and refer eligible patients. We also sent the study referral to an application that is integrated within the electronic health record system. The application nudges primary care providers about the study eligibility when a potentially eligible patient visit is scheduled.

Our retention efforts include a combination of strategies to ensure participants remain engaged throughout the study. Participants receive regular reminders and check-in phone calls, and study visits are conducted at LabCorp locations close to their residences to reduce the travel burden. The interventions are delivered remotely, minimizing the need for on-site treatments, and participants receive monetary compensation as part of the participation rewards. Additionally, they provide alternative contact information to ensure consistent communication. All participants receive IVR calls throughout the study as part of the engagement strategy. If any participant stops responding, the research coordinator contacts them directly to address any issues. Participants are contacted per the approved Human Subject Protocol. Up to 3 attempts are made to reach out to each participant via phone call, voicemail, or text to verify whether missed calls were due to technical issues or an intention to withdraw from the study. If a participant fails to respond to three consecutive contact attempts, no further outreach is conducted, and the participant is classified as lost to follow-up and removed from the calling list.

All patients first undergo telephone-based screening to confirm eligibility. For electronic consent, participants receive a consent form via email using REDCap, a secure web app for managing research data. If the participant is unable to complete the consent form electronically, either a phone consent is obtained or a physical copy of the consent form is sent to the participant’s mailing address. A completed, signed consent form is then uploaded to REDCap, and physical copies are stored in secure cabinets at the research site. Each participant is asked to visit a LabCorp location closest to the patients’ home addresses to provide a fasting blood sample to be analyzed for HbA_1c and lipid profile (ie, low-density lipoprotein cholesterol, high density lipoprotein cholesterol, very low density lipoprotein, triglycerides, and total cholesterol) and biometrics (ie, blood pressure, height, weight, and waist circumference). Measurements are recorded to the nearest 0.1 inches (in) for height and waist circumference and to the nearest 0.1 pounds (lb) for weight.

Once participant eligibility is confirmed from screening, the blood values used in the in-person screening visit described above are used as baseline values. In addition, a baseline survey packet consisting of sociodemographic questions and five validated survey instruments is sent electronically to the participants. The survey instruments are as follows: The Spoken Knowledge in Low Literacy in Diabetes Scale (SKILLD) assesses patients’ understanding of diabetes management [42]; the Patient Health Questionnaire-4 screens for symptoms of anxiety and depression [43]; The Diabetes Distress Scale measures the emotional burden associated with diabetes management [44]; The Diabetes Empowerment Scale evaluates patients’ self-efficacy in controlling diabetes [45], and the EQ-5D [46] survey assesses health-related quality of life. All study data are collected and securely stored in the REDCap database. If participants are unfamiliar or uncomfortable with completing the survey packet electronically or if they lack an email address, the survey packets are mailed to their preferred address. Included in the packet is a prepaid stamped envelope to return the survey packet. Upon receipt, survey responses are uploaded to REDCap, and the physical copies are stored in a secure cabinet at the research site.

Before the intervention began, participants across all arms of the study were shipped a pedometer, a glucometer with strips for 6 months, and a T2DM educational packet. The education content is highly pictorial and is based on the Diabetes Literacy and Numeracy Education Toolkit [47], a packet designed with low-literacy principles and visual aids, for T2DM education and management in patients with low literacy and numeracy skills. In addition, educational patient handouts from the Centers for Disease Control and Prevention (the National Diabetes Education Program and Diabetes Prevention Program) corresponding to all content areas measured by the SKILLD were included in the packet [48]. Topics covered in these materials included special considerations for exercise, understanding HbA_1c, eye health, foot exams for individuals with T2DM, and visual aids for making healthy food replacements to manage blood sugar. A guide for estimating portion sizes was also included to facilitate accuracy with daily diet tracking. Sending the educational packet ensures that all participants have access to the same basic principles of diabetes management before starting the intervention.

Computer-generated randomization assignment was done for all consented participants. We implemented a permuted block randomization with varying block sizes, blinded to the assessment team. Participants were randomized to four experimental conditions based on a 2 by 2 factorial design with equal probability allocation. The randomization list is concealed within the REDCap randomization module. We have programmed this such that the participant can be randomized and allocated only to an experimental condition after it is confirmed that baseline data have been collected and entered. Once the primary outcome measures are entered in REDCap, the study staff members can assign the participants to their respective experimental conditions. The statistician and the Principal Investigators are blinded, but the participants and the health coach are not blinded. All participants enrolled in the study were randomized to one of the four arms:

All participants received an initial orientation session. Study team members served as health coaches, including a clinical registered dietitian and a certified health coach. During this session, the team reviewed the SKILLD survey completed as part of the baseline survey to identify gaps in the participants’ knowledge of T2DM management. These gaps were then addressed, and an introduction to the intervention was provided. The Diabetes Literacy and Numeracy Education Toolkit educational packet was used by health coaches during the orientation process to ensure all participants have equivalent baseline knowledge regardless of their literacy level. For participants with limited literacy skills, the health coaches used this call to walk participants through the educational materials, focusing on the visual components and providing detailed verbal explanations of all key information. This individualized approach ensured that participants who cannot read are still able to access, understand, and engage with the educational content, thereby establishing a consistent foundation of T2DM management knowledge across the study population.

This call provided key information about the intervention, including expectations for the telephone-based ABMS and health coaching calls, participants’ randomization to fixed or adaptive rewards, and recommended use of a glucometer and pedometer. Participants also listened to a sample ABMS call to learn best practices for voice recognition when reporting step counts, medication use, and blood glucose testing.

The call included instructions to improve the completeness and accuracy of diet records, with reminders about commonly forgotten items (eg, beverages, condiments, and sweeteners) and guidance to avoid brand names for better matching in Nutritionix. Participants also received introductory nutrition education focused on making healthy replacements to support glycemic control, specifically limiting added sugars, refined carbohydrates, and saturated fats. These behaviors are tracked through daily ABMS calls and discussed during weekly health coaching sessions to enhance both the accuracy and efficiency of self-monitoring.

Following the orientation call, all participants were activated in the ABMS and began receiving calls starting the following Monday.

At month 6, similar to the baseline visit, each participant is asked to visit their nearest LabCorp location to provide a blood sample for HbA_1c, lipid profile, and biometrics. Participants are also sent a 6-month survey packet either electronically via email using REDCap or via physical mail, depending on the participant’s preference. This packet includes all of the surveys in the baseline packet in addition to: Speech User Interface Service Quality, which evaluates the quality, perceived ease of use, and user satisfaction of speech-based interfaces [54]; System Usability Scale, which measures system usability [55]; and the theoretical framework of acceptability survey [56] which assesses the acceptability of the intervention.

Upon completion of the intervention at month 6, participants are invited to complete a semistructured exit interview conducted over a phone call and recorded, to gauge expectations, perceived benefits, and suggestions for protocol improvement (see Multimedia Appendix 2 for interview guide).

Monitoring fidelity is essential to ensuring the intervention’s internal and external validity. In that context, we have developed a fidelity-monitoring plan (Table 1) to address the five areas of intervention fidelity outlined by the National Institute of Health Behavior Change Consortium: design of study, provider training, delivery of treatment, receipt of treatment, and enactment of treatment. Monitoring fidelity is done by the study coordinator, along with the support of the programmer, who provides periodic reports.

We are collecting feasibility, process, clinical, and patient-reported outcomes (PROs) during the intervention, with feasibility outcomes serving as the primary end points. As outlined in Tables 2 and 3, our primary outcomes include key feasibility indicators: recruitment, retention, and adherence to the intervention. Secondary outcomes include other feasibility measures such as process measures (eg, time to recruit, percentage of missing data, and enrollment rate), and selected patient-centered and clinical outcomes. The choice of clinical and patient-reported assessments is based on guidelines from the American Diabetes Association [57]. Specifically, diabetes distress is a key PRO, while HbA_1c and weight are the key secondary clinical outcomes. Exploratory outcomes include additional PROs (eg, quality of life, depression, self-efficacy, and acceptability of the intervention) and clinical indicators such as blood pressure and lipid profile. PROs are collected using validated instruments such as the Diabetes Distress Scale, EQ-5D, Patient Health Questionnaire-4, and the Diabetes Empowerment Scale. Intervention acceptability is assessed using the Speech User Interface Service Quality, System Usability Scale, and theoretical framework of acceptability scales, in addition to qualitative exit interviews. PA is measured using step count data. Adherence data, including daily self-monitoring of blood glucose and medication adherence, are collected through the ABMS IVR system.

In addition to recruitment, retention, and adherence-related measures, other feasibility measures include process, resource, and management measures such as attrition rate, coach time, time to recruit participants, and technology delivery-related issues (Table 3).

Data are being collected and managed in a custom-designed REDCap database. Data will be exported to the statistical analysis software, using relational logic checks to identify internal inconsistencies and out-of-range values. Systematic cleaning and verification of all collected data will be conducted to examine ranges, provenance, and missing values. Any inaccurate data points, outliers, and influential cases will be identified and corrected within REDCap. Dashboards and reports will be created and periodically updated to assist the team with recruitment and retention efforts, and track and minimize missing data.

The primary objective of this feasibility study is not to serve as a definitive, well-powered trial, but to conduct a pilot study that is sufficiently large to estimate design parameters, as well as assess feasibility parameters, such as attrition rate, adherence, and pre-post correlation [58-60]. Accounting for the bias and precision of the estimated pooled SD can help us conduct rigorous power and sample size calculations for a subsequent confirmatory study [61]. We are randomizing 88 participants across four conditions (22 per experimental condition). This is almost twice the number of participants per condition recommended when undertaking pilot studies, when there is no prior information on which to base a sample size [62]. The goal of our pilot study is also to estimate precise effect sizes and design parameter estimates, which requires sufficiently large pilot studies. Based on simulations published by Teare et al [61], an external pilot study should have at least 70 participants when estimating the pooled SD for a continuous outcome. Because we are using a factorial experiment, for each of the main effects, we will have 88 participants, with 44 participants in each of the two comparators. Hence, we have a sufficiently large pilot study to detect precise estimates of effect sizes for the two main effects and other design parameter estimates, such as pre-post correlation.

We also present power calculations of our experiment based on our sample size of 88. Assuming a type 1 error rate of 0.05, a conservative pre-post correlation of 0.5 in ANCOVA, and an attrition rate of 15%, we will have 80% power to detect an effect size (Cohen d) of 0.572 in our primary outcome. This would imply a main effect difference of 0.86% units of HbA_1c, assuming an SD of 1.5% [63]. Even if we assume the pre-post correlation is 0.3, which is very conservative, we will have 80% power to detect an effect size of 0.63 for the main effect. Assuming a more realistic pre-post correlation of 0.7, we can detect an effect size of 0.472. Missing outcomes data will be imputed using maximum likelihood estimation or a multiple imputation procedure, and primary analyses will not be restricted to only completers. Hence, our proposed study sample size of 88 participants is powered to detect an even smaller effect size than suggested here since the effective sample size, even assuming an attrition of 15%, would be between 74 and 88. Thus, our pilot study is well-powered to detect moderate to medium effect sizes. These power calculations were done using the MOST package in R (R Core Team), developed by Dziak et al [64].

This paper sought to outline the study protocol used in a pilot study grounded in SCT, aiming to assess the feasibility of an AI-assisted, low-cost, lifestyle modification intervention for glycemic control in the Deep South. The intervention is designed to be accessible through a standard telephone service and leverages MOST, an efficient and rigorous resource-management and continuous-improvement framework. The MOST design allows us to explore (1) a fixed versus adaptive (gamified) rewards program, and (2) automated versus human-delivered weekly health coaching for four total experimental conditions to see which is most effective for T2DM management. Findings from this pilot and feasibility study will inform the resources needed for a well-powered optimization trial. We will estimate key design parameters—such as effect sizes and pre-post correlations—to guide future sample size calculations. The study also helps identify protocol modifications to improve scalability, recruitment, and participant-centeredness.

The rationale for this study comes from the rising prevalence of T2DM in the Deep South and the lack of self-management in rural, underserved, or low digital literacy populations [69]. Given this escalating prevalence of T2DM, challenges in access to primary care and time available to assist patients in self-management of T2DM, coupled with disparities in its management, there is a need for an optimized, low-cost, scalable intervention that complements primary care for T2DM management to ensure health equity. Synchronous human coaching has been shown to be effective in glycemic control for patients with T2DM [70] and has been explored in rural-specific settings as well [71,72]. A nurse-led telecoaching study for rural patients was successful and cost-effective in T2DM management [73-80]. However, providing 10-15-minute live telephone calls to each patient at frequent intervals presents a significant barrier for scalability. This led to the development of automated health coaching delivered via an automated phone call by an IVR system. Participants use the keypad of any phone (landline, mobile, or smartphone) to interact with the IVR system, where the human coach is replaced by high-quality recorded interactive scripts [81]. IVR-delivered coaching is thus far more scalable than traditional delivery. IVR-based behavior monitoring has been shown to be effective in driving behavior changes [81]. Hence, the proposed intervention could contribute to an interprofessional primary care system, alleviating access issues. Our proposed resource optimization experimentation arms of fixed and adaptive gamified rewards are relevant in the context of existing wellness programs offered by employers and payors to incentivize behavior change. The intent is to promote behavior tracking that can facilitate behavior change where needed, leading to improved health outcomes and improved employee productivity, and less health care expenditure, such as emergency department visits or avoidable hospitalizations. The cost-effectiveness sustainability was not part of our pilot study; this is an important but different aspect of the Multiphase Optimization Strategy Preparatory phase and will be addressed in our future work using approaches such as microsimulations [82]. The GODART project aims to take a step toward building such a system. An overview of the innovative aspects and limitations is discussed below.

Our study has multiple strengths and innovative methodologies. From a technological standpoint, we combine NLP with standard telephone-delivered monitoring and coaching. By incorporating natural language understanding (AI) techniques, we aim to minimize the user burden in self-monitoring, which requires minimal time and low health literacy levels [83,84]. The study uses AI to allow participants to “speak” their daily updates into the phone for recording, significantly easing the data collection process. Additionally, we are implementing a gamified rewards scheme based on data collection done via phone to improve self-efficacy and self-monitoring behavior [31]. Gamification is widely recognized for its potential in managing chronic diseases, and our study leverages this approach through a simple reward system [85].

From the study design perspective, an additional strength is the adoption of the MOST framework, which emphasizes efficient resource management and continuous improvement [36]. Our study objectives align with MOST’s preparatory phase, and it is also structured to serve as the preparatory phase for a future large-scale MOST optimization phase. MOST’s preparatory and optimization phases are more efficient for intervention development than a traditional randomized controlled trial (RCT) since it would take substantially greater resources (time and sample size per RCT) to meet this goal [37]. These MOST phases in our study will provide more information than a typical feasibility and well-conducted pilot study using an RCT approach. Further, in MOST, intervention optimization could be an iterative and continuous process compared to a traditional RCT.

From the data collection and engagement perspective, our study leverages simple telehealth technology based on IVR to collect periodic data from participants and keep them engaged throughout the study period. Collecting these data as part of the diabetes self-management activities instead of apps or logs is meant to reduce the participant burden and keep the data collection process simple and accessible, regardless of patients’ health literacy or technological access. Further, analyses of these data can help develop learning models that can enable further optimizations through the development of just-in-time adaptive interventions [86,87]. Finally, blood draws are collected at the LabCorp location nearest to each patient’s home. This strategy minimizes potential delays by reducing the need for patients to travel to specific clinic locations.

There were several limitations to this study. First, the study design required participants to engage daily in a phone-based intervention, which may limit generalizability, as individuals who enrolled were likely more motivated at baseline than the broader population of adults with type 2 diabetes. As a result, the findings may be most applicable to patients already inclined to engage in self-management, highlighting the need for future research to develop strategies that also reach individuals with lower baseline motivation.

Second, although our system uses AI to process information collected through IVR calls, the coaching recommendations themselves are not generated by AI. Instead, they are developed by a human coach and then delivered asynchronously via IVR. Thus, the automated call arm still depends on human input and is not entirely independent of human resources. Since the design of this study, there has been a major shift in AI technology with the availability and mainstreaming of large language models. Future iterations of this project will be able to leverage LLM-based approaches to further enhance AI-based ABMS and AI-based health coaching.

We also asked participants to report whether they were checking their blood sugar, but did not collect actual glucose values. This was due to the lack of a clinical remote monitoring team available to respond to hypo or hyperglycemia alerts. Future studies may incorporate continuous glucose monitoring or partner with remote monitoring teams, such as registered nurses within a health care system, to respond to critical alerts and provide timely feedback and support.

Additional limitations include the exclusion of non-English speakers, which may reduce the generalizability of findings to linguistically diverse populations. All behavioral data was self-reported, which introduces the potential for recall bias—particularly underreporting—since more detailed entries (eg, food logs) led to longer IVR calls and may have discouraged full reporting. Additionally, the study does not include a scale to assess social desirability bias in the self-reported PROs, which we will incorporate in our future study. Coaching fidelity was not comprehensively evaluated: the automated coaching was still guided by a human coach, and the degree of standardization or comparability between the human and automated arms was not assessed. Given the applications of LLMs, we anticipate this issue to be addressed by the adoption of LLM in our GODART technology platform.

The rising prevalence of T2DM, specifically in the Deep South and rural areas, calls for interventions that can be delivered remotely yet with minimal digital complexity. These could serve as scalable digital care management plans offered in primary care settings to address the current primary care limitations due to limited access and time. This study seeks to test the feasibility of a telehealth and AI-driven intervention designed to be easily accessible, even for individuals with minimal digital literacy. By gathering data from this study, we intend to lay the groundwork for a future large-scale optimization trial as a step toward building an interdisciplinary primary care system that could integrate health technology effectively, ensuring that patients receive care at the right time and place.