In demonstration project 1, we will use descriptive statistics to evaluate feasibility metrics related to screening, recruitment, and enrollment processes, including the total number of patients screened for tobacco use, the proportion of patients who complete the smoking question on the screener, the proportion of current smokers provided automated advice to quit, and the proportion of current smokers who agree to enroll in a smoking cessation treatment program. mFLi’s recruitment and initial screening capability will be deemed feasible via the demonstration project if at least 90% of patients approached complete the prescreener, the proportion of patients identified as current smokers meets or exceeds the historical (prior 2 years) self-reported proportions recorded in the EHRs at each of the 3 partnering institutions, and at least 30% of eligible patients agree to enroll in the smoking cessation RCT (demonstration project 2).

Following receipt of smoking status information in demonstration project 1, potentially eligible patients who agree to enroll in treatment and who do not opt out of receiving information will then be directed to complete the main screener. Specifically, those who are recruited in person will have their information in the dashboard automatically transferred to mFLi’s REDCap and will be automatically sent a brief message and link to the main screener via SMS text message or email. Those recruited remotely will be seamlessly transitioned from the AAC prescreener to the main screener. Those who do not respond to the automated link will be contacted by research staff and asked to complete the screener. Study inclusion criteria are (1) aged 18 years or older, (2) self-report of smoking within the last 30 days and a history of at least 100 lifetime cigarettes, (3) ownership of a working smartphone, and (4) ability to speak English. Exclusion criteria are (1) failure to complete study enrollment within 14 days of completing the screener, and (2) being pregnant or breastfeeding. Patients meeting eligibility criteria will be invited to participate and provided with a detailed description of the study. All individuals who agree to enroll will complete a Health Insurance Portability and Accountability Act (HIPAA)–compliant electronic informed consent process.

Participants who consent to participate will be instructed to complete baseline measures, including demographics, subjective social status [42], health literacy [43,44], quit motivation [45], smoking cessation self-efficacy [46], perceived stress [47], the Patient Health Questionnaire depression scale [48], Cancer Patient Tobacco Use Questionnaire [49], and nicotine dependence [50].

After completing the baseline measures, participants (N=90) will be randomized 1:1 within mFLi to either a standard quitline treatment (QT; n=45) or a prototype version of a fully automated, digitally delivered smoking cessation treatment, Digital Motivation and Problem Solving (D-MAPS; n=45), using simple randomization sequences generated by the study statistician. Due to the nature of the protocol study and trial design, blinding of participants and the study statistician to treatment assignment was not feasible or necessary.

Participants randomized to the QT group will receive a 10-week supply of combination nicotine replacement therapy (patches and lozenges) and have their information sent to Tobacco Free Florida (a HIPAA-covered entity). Quitline staff will then proactively contact these individuals and present treatment options (eg, phone counseling, group counseling, and web-based services). Quitline counseling approaches are recommended in the Treating Tobacco Use and Dependence Clinical Practice Guideline as an evidence-based treatment option [51].

The D-MAPS treatment approach is based on the previously developed human-delivered Motivation and Problem Solving approach [52-54]. Participants randomized to the D-MAPS group will receive 10 weeks of combination nicotine replacement therapy and will be provided (via email and/or SMS text message) a link to download the app. The app-based intervention will comprise 3 distinct and personalized sessions (session 1—“Getting started,” session 2—“Continuing your journey,” and session 3—“Maintaining your progress”). In each session, users will complete brief assessments and, regardless of readiness to quit smoking, be given the option to create a wellness plan (eg, stress management, alcohol use, nutrition, physical activity, sleep, and finances). The intervention is designed for participants to repeat sessions, and participants’ responses and selections in each session will guide the delivery of dynamically tailored content (animated videos, audio, and text) to ensure a personalized UX responsive to individual needs and real-time changes in motivation or behavior. This intervention will integrate principles drawn from motivational interviewing, social cognitive theory, and cognitive behavioral therapy [55-57]. Participants will be proactively sent prompts via push notifications based on their scheduling preferences, and they can also access smoking cessation and wellness plan content on demand. Participants will be able to use D-MAPS over a 6-month intervention period.

Additionally, participants in both conditions will be provided with commercially available wearable wrist sensors (eg, Fitbit devices; Google LLC) and instructions for setup. These devices automatically and unobtrusively collect a range of health-related data, including physical activity data (accelerometer, gyroscope, and active minutes), as well as health and wellness data (heart rate, blood oxygen saturation, and skin temperature variation). Participants will be instructed to wear the sensors daily throughout the 6-month study. Specifically, participants will be asked to wear the sensors for their waking hours, or up to 16 hours per day, for the duration of the study. They will only be advised to remove the device while charging, during any extended water exposure (eg, showering or swimming), and in the event of irritation or skin reaction. At the completion of the study, participants will be permitted to keep the provided device. Fitbit data will be integrated with all app-based and assessment data.

Participants in both treatment groups will be prompted to complete ecological momentary assessments (EMAs) once per day during the initial 8 weeks, and once per week for the remainder of the study (4 months) via push notification. The method by which participants will complete EMAs will differ by treatment group, with QT participants completing EMAs online via REDCap and D-MAPS participants completing EMAs via the app. Using EMAs will allow us to examine the dynamic impacts of multiple psychological, social, and environmental factors during a quit attempt. Specifically, participants will be asked questions relating to their current smoking status, perceived stress level, smoking self-efficacy, motivation to quit smoking, and nicotine replacement therapy adherence [58].

The primary outcome measures for demonstration project 2 will be the feasibility of the D-MAPS app, assessed via usability and acceptability. Specifically, usability will be assessed with the 10-item System Usability Scale [59], as well as through a single item assessing the percentage of participants who rate the app as “easy to use.” It will be deemed usable if at least 80% of participants score >70 (range 0-100) on the System Usability Scale and at least 80% of participants respond “yes” (dichotomous yes or no) to the app being “easy to use.” The acceptability of and satisfaction with the D-MAPS app will be assessed via the Client Satisfaction Questionnaire [60]. It will be deemed acceptable and satisfactory if (1) at least 80% of participants score >3 (range 1-4) on the Client Satisfaction Questionnaire, (2) at least 80% of participants report being mostly or very satisfied with the intervention, (3) at least 80% of participants respond “yes” to recommending the intervention to a friend to quit smoking, and (4) at least 80% of participants rate the quality of the intervention as good or excellent. Finally, various app engagement metrics will be measured (eg, video playback duration and frequency of app access).

This study is not designed or powered to test significant differences between our 2 conditions in demonstration project 2. However, we will describe cessation outcomes in both treatment conditions using 7-day point prevalence abstinence at each follow-up and calculate CIs. These data, combined with feasibility metrics, will be useful when designing a future, fully powered trial.

Machine learning and deep learning, subfields of artificial intelligence, offer promising tools for identifying complex patterns in health-related data via computer algorithms in multiple domains, including cancer [61]. These methods are particularly well suited for analyzing data from diverse sources (eg, self-report, clinical assessments, wearable devices, and GPS) to generate predictive models of behavior. Despite this potential, machine and deep learning remain underutilized in mHealth interventions [62]. Therefore, the primary goal of demonstration project 3 is to evaluate the capacity of the mFLi platform to support the application of machine and deep learning methods for predicting key behavioral health outcomes (eg, smoking cessation and relapse). By integrating and analyzing data from the participants’ EHRs (eg, on diagnosis and treatment) and data collected from the 90 patients enrolled in demonstration projects 1 and 2 (split randomly; 70% of participants assigned to the training phase and 30% assigned to the validation and testing phases), this project aims to assess the feasibility of applying advanced analytic techniques to mHealth-collected data, thereby highlighting mFLi’s potential to shift traditional paradigms in behavioral research and support next-generation personalized interventions.

The application of machine and deep learning algorithms to mHealth can be enabled via two approaches: (1) cloud-based client server and (2) on-device app.

In cloud-based client server setup, the machine learning and deep learning algorithms are hosted in the cloud (server), with access to powerful computing resources (graphics processing units), and the data are sent from the mobile device (client) over the internet to the cloud. The main advantage in this case is that conventional and complex machine learning algorithms can be directly used. In such a scenario, the analysis can begin with conventional ensemble machine learning techniques, including random forests, which are robust to heterogeneous data types and have demonstrated strong performance in prior related studies [63-66]. The importance of different factors will be ranked using methods based on the Gini index (eg, permutation testing), and associated statistical metrics will be used to assess their significance. To address limitations inherent in traditional machine learning approaches, we will also explore deep learning models, which integrate feature extraction directly into the predictive modeling process [67]. These models are particularly well suited for large, high-dimensional datasets. To enhance model interpretability, we will use techniques such as Shapley Additive Explanations values and interpretability maps, allowing us to rank order key predictors contributing to model outcomes [68-70]. If sample size limits the performance of deep learning models, we will use hybrid models that combine both conventional machine learning and deep learning techniques [71]. The main disadvantages in this case are latency and privacy concerns.

In on-device app setup, the machine and deep learning algorithms can run directly on the mobile device. Typically, the algorithms are developed offline as described earlier (training mode), then deployed on the device (inference mode). The inference mode is less computationally demanding and can benefit from existing frameworks such as Core ML (Apple Inc) on Apple iOS and TensorFlow Lite (Google LLC) on Android systems. This approach provides advantages such as real-time performance, offline functionality, and improved privacy by maintaining the data secured on the device. This approach will be favored in our design.

We will make every effort to keep missing data at a minimum. We will compare the characteristics of responders and participants who dropped out of the study. Additionally, several imputation strategies (eg, last value carried forward, regression approaches, and missing equals smoking) will be used as sensitivity tests.