Cigarette smoking is a serious public health problem and the leading cause of preventable disease, disability, and death in the United States [1]. While overall smoking rates have decreased, smoking rates among individuals who are socioeconomically disadvantaged remain high [2-5]. Individuals who are socioeconomically disadvantaged are affected disproportionately from smoking-related diseases [2-5]. Individuals who are socioeconomically disadvantaged may be less likely to try to quit smoking (eg, due to lower self-efficacy and maladaptive beliefs) and are less likely to use evidence-based strategies (such as quitlines).

One strategy to engage these individuals is by using computer-tailored messages delivered via texting. Computer-tailored health communication involves the use of computer programs to select the best personalized message for an individual. This form of communication has been effective in enhancing motivation to quit and supporting cessation [6-12]. In a systematic review, computer-tailored messages showed higher cessation rates than nontailored messages (relative risk 1.17, 95% CI 0.97-1.41). Computer-tailored messages work because participants are more likely to engage and carefully process personally relevant messages [13]. Thus, methods to increase the relevance of the message may increase the effectiveness of the smoking cessation intervention. Computer-tailored messages delivered via texting over phones may be particularly attractive for individuals who are socioeconomically disadvantaged. Mobile phone ownership is now nearly universal (95%) [14], and these rates are consistent across socioeconomically disadvantaged groups (less than high school education, 92% and those earning <US $30,000, 92%). The texting app is the most commonly used app on mobile phones [14].

In our National Institutes of Health–funded trial, we are testing Adapt2Quit, which uses machine learning algorithms (ie, recommender systems) to computer-tailor messages for an individual and deliver them via texting [15]. Recommender systems can be programmed to continuously learn from user feedback and improve personal relevance and customer engagement [15]. Adapt2Quit also includes proactive texting facilitation to the state quitline in which participants are asked if they want to be referred to the quitline, and those who agree are referred to the quitline. Quitlines are programs in the United States that offer evidence-based support (eg, counseling and medications) to individuals who smoke. These individuals can either directly contact quitlines using the number 1-800-QUIT-NOW or be referred by a health care provider or tobacco cessation programs. Although quitlines have been established in all 50 US states, they are vastly underused. The use of texting to promote quitline use could potentially increase quitline use, and thereby increase the effectiveness of the smoking cessation intervention for the participants.

In our preliminary studies, Adapt2Quit increased engagement and 30-day point-prevalence effectiveness over a standard messaging that randomly selected these messages [16]. In a subanalysis among those who reported lower education (n=49), the average rating of the messages was higher on more days than the average ratings of true comparison messages (77% vs 23%; P<.01). This pilot study showed the potential of the Adapt2Quit trial among individuals who are socioeconomically disadvantaged. Since the beginning of the trial, we have completed another randomized controlled trial that tested the machine learning algorithm compared to standard messaging among those who were recruited via social media [17]. In the other trial, there was no significant difference in 6-month point-prevalence smoking cessation (adjusted odds ratio 0.81, 95% CI 0.61-1.08) between the recommender system (146/412, 35.4% of participants with outcome data) and messages tailored by standard messaging based on participants’ baseline readiness to quit groups (156/389, 40.1% of participants with outcome data). However, the other trial differs from the current trial being reported in that the messages were delivered via email, it did not target individuals who are socioeconomically disadvantaged, and the system did not facilitate referral to the quitline. Furthermore, another group has evaluated a similar recommender system approach for message selection in a pilot trial for self-reported smoking cessation [18]. However, additional studies testing recommender systems are needed to uncover the potential of these systems for increasing the effectiveness of computer-tailored smoking cessation interventions.

The Adapt2Quit intervention is based on the self-determination theory (SDT). SDT-based interventions support autonomous decisions and are designed to increase intrinsic motivation and self-regulation and they have been shown to improve motivation and cessation outcomes among those who smoke [19-24]. Three innate psychological needs are theorized to promote self-motivation: autonomy, relatedness, and competence [19]. Autonomy reflects the need to engage in behaviors with a sense of choice or personal endorsement [25]. Autonomy support can be provided by acknowledging an individual’s unique perspective and offering choice [25,26]. Relatedness reflects the degree to which an individual feels connected to and understood by others. Competence is a feeling of being capable and is akin to self-efficacy, the belief that one can bring about desired outcomes. By supporting autonomy, relatedness, and competence, health behaviors are more likely to be internalized and maintained. Specific examples of how Adapt2Quit supports the SDT are provided in Textboxes 1 [27] and 2. Our aims and hypotheses align with our proposed model, which suggests that relevant personal messages from the Adapt2Quit system will enhance system engagement. This increased engagement is expected to improve internal processes (ie, the SDT constructs measured using the Perceived Competence Scale [PCS]), leading participants to adopt cessation-supporting behaviors (external processes such as quitline and nicotine replacement therapy [NRT] use), ultimately resulting in smoking cessation.

This paper describes the protocol and status of our ongoing randomized controlled trial testing the Adapt2Quit intervention among individuals from socioeconomically disadvantaged backgrounds who smoke. Our comparison participants receive the biweekly texting facilitation messages but do not receive the motivational messages recommended by the Adapt2Quit intervention. Our protocol description includes the intervention and comparison system as well as the description of the analyses of the primary and secondary hypotheses listed in the following sections. Since recruitment for our study has finished, we describe the demographics of the participants we have recruited. Our follow-up data collection is in progress.

This study is a 2-arm randomized controlled trial comparing participants who received smoking cessation text messages selected by the Adapt2Quit recommender system plus a texting facilitation approach to connect participants to the quitline with participants who received the texting facilitation but no smoking cessation text messages. Study participants were blinded to group assignments. They were recruited from health systems in diverse geographic areas. Recruitment started in April 2020 and was completed in March 2024. Our target sample size was 750 individuals from socioeconomically disadvantaged backgrounds who smoke.

The Adapt2Quit study has been approved by all participating institutions’ institutional review boards (IRBs), including Baystate (BH-20-123) and Johns Hopkins (IRB00280942), with the University of Massachusetts Chan Medical School (UMass Chan) IRB (H00018991) being the central IRB. The study is listed in ClinicalTrials.gov (NCT04720625). It is being carried out in accordance with the Helsinki Declaration, and all participants have signed an informed consent form before enrollment in the study. Participants have been informed that they have the right to withdraw their consent at any point during the study, and the reason for their withdrawal will also be documented if the participant so wishes. The participants have received compensation for completing study activities. They received US $25 for completion of baseline, US $25 for the 6-month follow-up, US $50 for completing the carbon monoxide (CO) verification assessment, and an additional US $50 for participation in the qualitative interview. Upon randomization, all patients have been assigned an individual ID number so their data remain anonymous.

The inclusion criteria were as follows: an individual who smokes, comes from a socioeconomically disadvantaged background, speaks English, who is active in care (having at least 2 clinical visits in the past 2 years), and who has a texting-enabled mobile l phone. We used insurance status to identify participants as socioeconomically disadvantaged [28]. In the United States, people can receive health insurance or health care payment coverage from several sources including through their employer, other private health insurance companies, or through government-supported programs. Individuals who have these government-supported programs are more likely to be from socioeconomically disadvantaged groups. For our study, we used Medicare (a government-funded program for older adults and those with disabilities) and Medicaid (a state and federal program for low-income individuals) to identify individuals from socioeconomically disadvantaged backgrounds. We determined smoking status by asking participants if they had smoked at least 100 cigarettes and their current smoking frequency (Multimedia Appendix 1). Those who responded “yes” to 100 cigarettes and “some” or “every day” were included.

The exclusion criteria were as follows: adults who were unable to consent, individuals who were not yet adults (ie, aged <18 years), prisoners, people with Alzheimer disease, and pilot study participants.

Recruitment was spread over the 3.5 years of the study. Study participants were recruited at urban and rural health care settings across 3 health care systems in the United States (UMass Memorial Medical Center, Baystate Health, and Johns Hopkins University). We used an opt-out recruitment strategy, with some variations between the three sites. At all 3 sites, after obtaining a Health Insurance Portability and Accountability Act waiver, potential participants who met our inclusion criteria in the clinical data warehouse were identified. At the UMass Chan and Baystate Medical Center, an initial invitation letter with a self-addressed, stamped opt-out postcard was sent to those eligible from the local site principal investigator outlining the study’s purpose. Johns Hopkins University’s site used emailed study invitation letters with instructions to opt-out via email instead of or in addition to mailed letters and “opt-out” postcards for those without an email address. Instructions were provided on how to opt-out by email. At all sites, individuals who opted out or whose email was returned undelivered within 2 weeks were removed from the recruitment list.

Potential participants were phoned up to 4 times to determine their interest and were screened for eligibility. Before starting the first recruitment call, the research staff sent an automated introductory text message to potential participants who would be called that day. During the phone call, participant eligibility was assessed (Multimedia Appendix 1). Eligible participants were asked to provide informed consent for study participation, were asked to complete the baseline data, and were then randomized to treatment. The screening session, informed consent, and baseline data collection were conducted via Zoom (Zoom Communications, Inc) or by phone.

Once randomized, after completion of the baseline survey (Multimedia Appendix 1), the Adapt2Quit intervention participants received 5 initial motivation messages and were asked to provide a rating. The recruitment staff viewed the rating in real time to confirm whether the participant understood and was comfortable with the system. The staff emphasized the importance of rating their messages to intervention participants. The research staff explained the quitline facilitation process and smoking status texting assessment to both groups and reiterated the importance of completing the ratings to intervention participants. Those randomized to the control group received only the texting facilitation and smoking status but did not receive the motivational messages, and instructions were provided accordingly.

Separate randomization tables were created for each site to enhance the balance between the two groups. Participant allocation to study arms was based on a permuted block scheme in which treatment assignments were made within blocks so that numbers assigned to each treatment arm are equal after a block has been filled. Blocks of various sizes (2, 4, and 6) were used in random order to facilitate allocation concealment. As in our prior technology-based trials, we embedded randomization within the technology [29,30]. After completing the screening, informed consent, and baseline data collection, the research staff entered the participant ID and participant mobile phone number from the survey into a web-based system, which then assigned the allocation based on the randomization table. Using this technique, participants and staff were blinded to allocation during the initial session. The research staff were subsequently unblinded to provide personalized training for the intervention and control groups, as noted.

To isolate the effect of the Adapt2Quit recommender system, comparison participants do not receive tailored motivational messages but receive the texting quitline facilitation and the texting assessment of smoking status with the same frequency as intervention participants.

Data are collected throughout the study at several time points via text-based ratings, quitline use, assessments, questionnaires, and CO readings. Our main outcomes are described subsequently.

This study protocol describes a randomized controlled trial to test a machine learning algorithm–based, computer-tailored health communication system for smoking cessation. We hypothesized that the intervention will impact mechanisms of action—specifically engagement, competence (a theoretical construct), and adoption of cessation-supportive actions—as well as demonstrate effectiveness in promoting smoking cessation. Our system will test these hypotheses among individuals who are socioeconomically disadvantaged, a group disproportionately affected by smoking.

Because of the potential for wide reach and effectiveness, texting programs (especially texting) have been increasingly adopted in real-world settings worldwide (public health and within health care systems) [48-50]. For example, the World Health Organization has helped establish texting programs for smoking cessation in several countries, including Costa Rica, Tunisia, and India [51]. The US Department of Health and Human Services established a Text4Health Task Force in 2010 to promote health text messaging programs in the United States [48]. In 2017, 55% of state quitlines offered texting programs (up from 20% in 2012) [34,52]. There is now growing evidence that these programs can be effective. An evaluation of SmokefreeTXT (a free evidence-based smoking cessation program offered by the US National Cancer Institute) reported that point-prevalence abstinence for smokers who initiated treatment at completion of the 42-day program was 7.2% (3% at the 6-month follow-up) [53]. Leaders have called for continued innovations to increase the effectiveness of these health messaging programs, especially to increase effectiveness with smokers from socioeconomically disadvantaged backgrounds.

Our study will provide important evidence on whether a more personalized approach using machine learning can encourage quitline use and support smoking cessation, compared to a text-based referral to quitline. In previous papers, we described our pilot studies testing the Adapt2Quit intervention [16,35,54]. Furthermore, our prior large randomized controlled trial compared the Adapt2Quit intervention to a system that delivered messages based on participants’ baseline readiness to quit [17]. In that trial, we recruited participants online as well as through referrals from friends and family. In addition, messages were delivered by email rather than by text, as is being done in this trial. The previous study also did not include biochemical verification of smoking cessation. Beyond the potential benefit of computer-tailored messaging, our system aims to actively engage participants with the quitline. Quitlines have proven to be effective in supporting people to quit smoking, yet using a quitline and attending counseling sessions is a significant step for individuals from socioeconomically disadvantaged backgrounds who smoke. Currently, only 2% to 3% of people who smoke use the quitline, with usage rates even lower among individuals from socioeconomically disadvantaged backgrounds [34]. Prior research has shown that it is possible to use a low-intensity intervention such as texting to increase motivation and the use of cessation resources among people who smoke [55].

In addition to our system, one group compared a similar recommender system in a pilot trial to a system that selected messages based on a knowledge base. They compared the message appreciation, engagement with the system, and one’s own self-reported smoking cessation status (7-day point prevalence) [18]. Another group developed a recommender system for smoking cessation and promoting physical activity [56]. This group described an experiment they conducted to identify the optimal approach for selecting messages. In addition, a systematic review examined the potential of using another machine learning approach—reinforcement learning—for similar purposes [57]. Together, these approaches highlight the potential of the use of machine learning approaches to enhance a widely used behavioral intervention: computer-tailored health communication.

Because we have completed recruitment for our study, we also provide the characteristics of our participants. Our recruitment method, using an opt-out strategy at 3 different health care systems, was able to successfully recruit a diverse sample. A third of our sample (265/757, 35%) are Black, and 16% (121/757) self-identified as Hispanic or Latino. There may be several factors for our successes, including that our intervention is a low-intensity intervention using a texting model that is widely available. Furthermore, our procedures for screening, baseline, and follow-up are all remote, which may have reduced barriers to recruitment. Participants in our study indicated higher motivation levels than the general population. This is an artifact of our recruitment strategy that people who are more motivated participate in a study about tobacco cessation. This more representative sample implies that our study findings will be more generalizable to the US population.

We describe the protocol for a rigorous evaluation of behavioral intervention among individuals from socioeconomically disadvantaged backgrounds. Since recruitment for our study is complete, we also present participant characteristics. Using an opt-out recruitment strategy across 3 different health care systems, we successfully recruited a diverse sample, which has been challenging to achieve for clinical trials. Our fully remote procedures for screening, baseline, and follow-up may have reduced barriers to recruitment. Notably, 35% (265/757) of our sample identified as Black and 16% (121/757) as Hispanic or Latino. This more representative sample suggests that our study findings will be more generalizable to the US population. One limitation is that our participants reported higher motivation levels for quitting smoking than the general population, which is likely due to the recruitment strategy attracting those who were more motivated to participate in a tobacco cessation study. Because we only used insurance status to determine socioeconomically disadvantaged, some participants may not align with all characteristics that may describe socioeconomically disadvantaged.

Our project is the first to rigorously test the use of a recommender system for smoking cessation among individuals from socioeconomically disadvantaged backgrounds who smoke. In this paper, we have reported on protocols for conducting a rigorous trial. In addition to our primary analyses for the 3 aims, we will conduct exploratory analyses to assess the heterogeneity of treatment effects across these subgroups, which could yield valuable insights for future tobacco cessation efforts. Our follow-up papers will report on the results of our trial following the analysis plans described in this paper. Future studies will also explore implementing texting interventions in various clinical and public health settings.