Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice PLOS Digit Health . 2026 Apr 10;5(4):e0001322. doi: 10.1371/journal.pdig.0001322 Search in PMC Search in PubMed View in NLM Catalog Add to search Reducing alcohol consumption in UK armed forces veterans: Feasibility of using personalized push notifications with AI Daniel Leightley Daniel Leightley 1 Department of Population Health Sciences, School of Population Health Sciences, King’s College London, London, United Kingdom Conceptualization, Formal analysis, Investigation, Methodology, Software, Writing – original draft, Writing – review & editing Find articles by Daniel Leightley 1, * , Charlotte Williamson Charlotte Williamson 1 Department of Population Health Sciences, School of Population Health Sciences, King’s College London, London, United Kingdom Data curation, Project administration, Writing – review & editing Find articles by Charlotte Williamson 1 , Iain J Marshall Iain J Marshall 1 Department of Population Health Sciences, School of Population Health Sciences, King’s College London, London, United Kingdom Writing – review & editing Find articles by Iain J Marshall 1 , Vasa Curcin Vasa Curcin 1 Department of Population Health Sciences, School of Population Health Sciences, King’s College London, London, United Kingdom Writing – review & editing Find articles by Vasa Curcin 1 , Roberto J Rona Roberto J Rona 2 King’s Centre for Military Health Research, King’s College London, London, United Kingdom Investigation, Methodology, Writing – review & editing Find articles by Roberto J Rona 2 , Dominic Murphy Dominic Murphy 2 King’s Centre for Military Health Research, King’s College London, London, United Kingdom 3 Combat Stress Centre for Applied Military Health Research, Leatherhead, United Kingdom Investigation, Methodology, Writing – review & editing Find articles by Dominic Murphy 2, 3 , Nicola T Fear Nicola T Fear 2 King’s Centre for Military Health Research, King’s College London, London, United Kingdom Investigation, Methodology, Writing – review & editing Find articles by Nicola T Fear 2, † , Laura Goodwin Laura Goodwin 4 Faculty of Health and Medicine, Lancaster University, London, United Kingdom Conceptualization, Investigation, Methodology, Writing – review & editing Find articles by Laura Goodwin 4 Editor: Benjamin Filtjens 5 Author information Article notes Copyright and License information 1 Department of Population Health Sciences, School of Population Health Sciences, King’s College London, London, United Kingdom 2 King’s Centre for Military Health Research, King’s College London, London, United Kingdom 3 Combat Stress Centre for Applied Military Health Research, Leatherhead, United Kingdom 4 Faculty of Health and Medicine, Lancaster University, London, United Kingdom 5 Delft University of Technology: Technische Universiteit Delft, NETHERLANDS, KINGDOM OF THE DL is a reservist in the UK Armed Forces. This work has been undertaken as part of his civilian employment. VC is supported by EPSRC-funded King’s Health Partners Digital Health Hub (EP/X030628/1). DM is a trustee at the Forces in Mind Trust who funded an earlier trial of DrinksRation. NTF is partly funded by a grant from the UK Ministry of Defence and is a trustee (non-paid) of a charity supporting the health and wellbeing of service personnel, veterans and their families. † Deceased. ✉ * E-mail: [email protected] Roles Daniel Leightley : Conceptualization, Formal analysis, Investigation, Methodology, Software, Writing – original draft, Writing – review & editing Charlotte Williamson : Data curation, Project administration, Writing – review & editing Iain J Marshall : Writing – review & editing Vasa Curcin : Writing – review & editing Roberto J Rona : Investigation, Methodology, Writing – review & editing Dominic Murphy : Investigation, Methodology, Writing – review & editing Nicola T Fear : Investigation, Methodology, Writing – review & editing Laura Goodwin : Conceptualization, Investigation, Methodology, Writing – review & editing Benjamin Filtjens : Editor Received 2025 Jul 10; Accepted 2026 Mar 8; Collection date 2026 Apr. © 2026 Leightley et al This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice PMCID: PMC13068231 PMID: 41961898 Abstract This study assessed the feasibility of deploying RationAI, a personalized AI-supported messaging framework, to reduce alcohol consumption among UK Armed Forces veterans. Participants were given DrinksRation, a mobile phone app, and allocated to receive either personalized or generic behavior change messages over 12 weeks. A total of 2,871 participants registered for an account during the study period. Feasibility was evaluated through recruitment (n = 2,871), retention (25.4% met engagement criteria), app usage and message delivery. Of those eligible, 343 participants were allocated to the personalized messaging group and 385 to the generic message group. The personalized group had higher early engagement, with app usage peaking at 212.4 (95% CI: 207.32 to 217.45) seconds in Week 2 compared to 183.7 seconds in the control group (95% CI: 178.90 to 188.46; p < 0.001) and received more notifications on average, reflecting additional personalized and event-triggered messages delivered as part of the intervention (47.7 [SD = 18.8] vs 16.3 [SD = 5.3]). Alcohol consumption declined in both groups over the 12-week period, with the personalized group showing a greater reduction from 31.08 to 13.20 units per week, compared to 31.24 to 15.17 units in the control group. Statistically significant between-group differences were observed at Week 2 (p = 0.027), Week 3 (p = 0.041), Week 4 (p = 0.008), and Week 10 (p = 0.049), favoring the personalized group, although between-group differences attenuated towards Week 12. Despite high attrition, the app engaged participants from an important population. These findings suggest the feasibility of personalized digital interventions for alcohol reduction, but there is a need for improved strategies to enhance long-term engagement. Author summary Alcohol misuse remains a major health concern among UK Armed Forces veterans, yet many experience barriers to seeking support through traditional services, including stigma, poor treatment access, and negative perceptions of care. In this study, we examined whether it was feasible to deliver RationAI, an artificial intelligence-supported messaging framework, through the DrinksRation smartphone app to veterans drinking above recommended UK guidelines. Participants received either standard generic behavior change messages or a more personalized package of messages tailored using app activity, alcohol use patterns, and self-reported wellbeing data. Over the 12-week study period, alcohol consumption declined in both groups, with the personalized group showing greater reductions during the earlier weeks of follow-up. Participants receiving personalized messages also showed higher initial app engagement and received more tailored and event-triggered notifications. However, engagement reduced over time across both groups, and many users did not remain active long enough to meet the study retention threshold. Overall, our findings suggest that AI-supported personalized messaging can be delivered feasibly to an important veteran population and may enhance early behavior change, although strategies to improve longer-term engagement are still needed. Introduction There are an estimated 1.8 million Armed Forces (AF) veterans in England and Wales (United Kingdom; UK), defined by the UK Government as those serving in the military for at least 1 paid day [ 1 ], and circa 160,000 personnel currently undertaking active duty [ 2 , 3 ]. Among AF serving and ex-severing population, research has shown that alcohol misuse is double the rate of that seen in the UK general population [ 4 ]. To place this into context, the UK Chief Medical Officer recommends that an individual should not regularly consume more than 14 units of alcohol per week (UK standard alcohol unit equates to 8g of pure alcohol). Consuming any more than this could be harmful or hazardous to health [ 5 ]. Research has showed that alcohol misuse persists after personnel leave service [ 6 ], highlighting a need for improved care services. In the United States (US), alcohol misuse among veterans is also high, with studies showing that approximately 56.6% of veterans score for alcohol misuse, and 7.5% for problematic alcohol misuse, like patterns observed in UK veterans [ 7 ]. Research has shown that those who seek support for alcohol misuse in a clinical setting attend fewer health appointments and are more likely to have a negative perception of mental health treatment [ 8 ]. In a UK and US context, one explanation may be that those misusing alcohol are often denied access to mental health treatment services until they have reduced their hazardous drinking [ 9 ]. Therefore, interventions that target drinking behavior are needed to enhance engagement, change behaviors and promote engagement with mental health services and improve mental health outcomes and quality of life. Since the Covid-19 pandemic, people in the UK are waiting longer for mental health referrals for alcohol use and treatment [ 10 ]. Traditional interventions such as face-to-face, drinking advice and leaflets have demonstrated some efficacy in reducing alcohol intake. However, the one-size-fits-all approach limits the ability to address individual needs in areas such as behavior, motivations, mental health and environmental factors [ 11 ]. In the UK, digital alcohol reduction has focused on the use of self-monitoring apps at a population level (e.g., Drink Less [ 12 ], Drinkaware [ 13 ]), where users are encouraged to regularly record and monitor their alcohol consumption using visualizations, dashboards and trackers [ 14 ]. Digital self-monitoring has been found to be associated with improved outcomes and is an effective Behavior Change Technique (BCT) for reducing alcohol misuse. A review of personalized digital interventions found that techniques such as behavior substitution, problem solving and providing a credible source of information, were associated with reductions in hazardous and harmful alcohol consumption [ 15 ]. Another review identified that personalized notifications, using language customized to the individual, play in a key role in promoting positive changes in behavior [ 16 ]. However, current smartphone interventions focused on the general population do not target the AF community in areas such as military language, prevailing social context, comorbid mental health problems, military service experience and perceptions of consumption [ 17 ]. One approach to addressing the lack of tailored care for AF populations is the use of artificial intelligence (AI) to personalize messaging, tailoring support to the specific individual needs, behaviors and contexts [ 18 , 19 ]. AI-supported interventions have the potential to enhance engagement, increase behavior change approaches, and improve health outcomes by providing timely, adaptive, and contextually relevant content which includes messages and in-app visualizations. This could be particularly beneficial to AF personnel who encounter stigma and barriers to seeking help, and digital health resources may help overcome these barriers [ 9 , 20 ]. By using self-monitoring, goal setting, action planning, and social support elements as part of digital interventions, AI can support engagement and habit formation and maintenance [ 21 , 22 ]. Behavioral data, user preferences and real-time factors can be used with AI to adjust interventions, ensuring greater relevance and effectiveness. Further, personalized interventions can deliver timely intervention prior to crisis or chronic manifestation, ensuring that behavior change prompts and supportive messages are delivered at the right moment. This study aimed to assess the feasibility of RationAI, a personalized AI-driven messaging framework designed to reduce alcohol consumption among UK AF veterans. Feasibility was evaluated as participant engagement, app usage over time, retention and message delivery success, with preliminary analysis of behavioral outcomes. The framework was evaluated by comparing veterans who received personalized AI-supported messages against those who received generic behavior change messaging, both delivered through the DrinksRation [ 23 ] smartphone app. While DrinksRation is publicly available and can be used by any individual in the UK, it was specifically developed through co-design with the AF community to address their unique needs. For the purposes of this study, inclusion was restricted to AF veterans who met eligibility and engagement criteria. The RationAI framework integrates BCTs and adaptive user personas to tailor messages in real time, aligning content with user-reported behaviors, motivations, and mental health indicators. This adaptive approach aims to enhance adherence, reduce hazardous drinking, and support sustained behavior change. Methods Ethics This study was approved by the King’s College London Research Ethics Committee (reference: HR-19/20–17438). Informed consent was obtained from all participants. Data used in this project can be viewed on the Open Science Framework ( https://osf.io/7s6j8 ). Trial design Data used in this analysis was collected from the DrinksRation app [ 23 – 25 ], which is publicly available to users in the UK (efficacy of DrinksRation has been reported elsewhere [ 23 , 24 ]). All messages were delivered via push notifications. Participants were allocated in the order of signup to either: 1) an intervention group receiving personalized messages, or 2) a control group without personalized messaging but receiving standard generic messaging. Participants DrinksRation is available via Google Play Store or Apple iOS Store. No limits were placed on who could download or use the app. The app was promoted via social media, advertisements on social media, email and leaflets. Participants were included in this analysis if they had served in the UK AF and identified as a veteran, provided baseline demographics data, consented to data collection, consumed more than 14 units of alcohol per week at baseline and used the app for 6 weeks (based on a minimum of 3 interactions per week) out of a possible 12 weeks from baseline. 14 units of alcohol per week was selected as it is the recommended maximum consumption by the UK Chief Medical Officer. The 12-week duration was selected based on existing literature, which shows this time frame as an appropriate comparator for evaluating brief interventions, and the 6 week point for use in analysis was selected to reflect participants engaging more than half the time of the study period [ 26 – 28 ]. Sample size As this was a feasibility study, no formal sample size calculation was performed. The intent was on assessing the feasibility of recruitment over a 12-month period, during which the recruitment window remained open to observe the number of participants who could be enrolled. Allocation and masking Participants were allocated using a single allocation process at the point of account creation (signup). Upon consent, each participant was assigned to a group based on a pre-determined allocation sequence (in order of sign up). No stratification or blocks were used in the allocation process. Due to the nature of the study, the research team was not blinded to participant allocation. Intervention DrinksRation (formerly called InDEx [ 23 , 29 , 30 ]) app was developed following the Medical Research Council Complex Intervention Guidelines [ 31 ] and used a co-design approach. It was developed by the King’s Centre for Military Health Research and Lancaster University, supported by experts in app development, epidemiology, addiction psychiatry, and military mental health. The app was designed as a Research Viable Product [ 32 ] to support veterans drinking at a hazardous or harmful level. The app seeks to enhance participants’ motivation and self-efficacy in modifying their alcohol consumption using BCT presented via the in-app content and through push notification messages sent to participants. The iterative development process, theoretical framework, feasibility trial, pilot and Randomized Controlled Trial are reported elsewhere [ 23 , 25 , 29 , 30 ]. DrinksRation was compatible with iOS and Android-supported devices released after 2017. Participants complete questionnaires via the app weekly, including questions on their mood and general mental health, which are used to personalize the app content and push notifications. Participants receive push notifications on the day that each questionnaire is due to prompt completion with further remainders every 24 hours for 3 days. DrinksRation uses a BCT framework to promote positive changes in alcohol consumption [ 23 , 33 ]. The framework uses evidence-based strategies for enhancing motivation and self-efficacy. A library of generic, personalized, and personalized (event-based) message segments was developed through focus groups, stakeholder engagement, and a review of relevant literature (see Fig 1 ), aiming to promote adherence to DrinksRation and reduce alcohol consumption using the BCT framework [ 30 , 34 , 35 ]. Example message bank segments are available from the corresponding author. Fig 1. Example message segments categorization. Open in a new tab Both groups received the same schedule of generic notifications (including questionnaire prompts and reminders). The personalized group additionally received personalized and personalized (event) notifications generated by the RationAI framework; this higher overall notification volume is therefore a deliberate component of the intervention being evaluated rather than an unplanned imbalance between groups. Personalization framework The personalization framework used Reinforcement Learning (RL) to model changes in behavior with the aim of reducing alcohol consumption via the use of push notifications. The framework is tested against adaptive user personas which are constructed for each user based on a combination of demographic information, app interaction and analytics data, alcohol consumption data, drink free day counts, and self-reported outcomes (e.g., mood, anxiety, depression and loneliness status) collected via questionnaires. This persona updates on a rolling 28-day period window. A full list of features is presented in Table 1 and is used as the input vector for the personalization framework. Active engagement is required, if a user does not engage, the adaptive persona is not generated, and generic messaging is used. Table 1. Feature set for modelling RationAI. Feature domain Features Demographic information Age, gender, service status, service length App interactions and analytics data Number of sessions, number of page views, view duration [average per day, std per day], number of times goals progress is viewed, number of messages interacted with, number of AM interactions, number of PM interactions [daily, weekly, 28-days]. This is recorded using in-app Google Analytics Alcohol consumption Weekly consumption in units, daily consumption in units, number of binge drink episodes [more than 6 units per day over 3 consecutive days], number of weekday consumption, number of weekends of consumption, number of drink free days, longest drink free streak, longest consumption streak. Alcohol consumption is measured via the drink’s diary Self-reported questionnaire responses Weekly depression score, weekly anxiety score, weekly loneliness score, number of weeks with depression, number of weeks with anxiety, number of weeks with loneliness Goal setting Number of goals set, average time taken to achieve goal, number of goals missed, number of goal successes, number of engagements with app after being sent a notification related to goals [within 1 hour] Notifications Number of notifications sent, number of personalized notifications sent, number of generic notifications sent, number of personalized (event) notifications sent, average time between notification and app engagement (limited to 168 hours), number of notification engagements Change trends For all features, except ‘Demographic Information’, change trend features are created which present the change compared to the prior week Open in a new tab Training The RL model was trained using a combination of historical app usage data, synthetic data, and researcher-curated message interactions, with intended outcomes informed by user personas. The agent was a Deep Q Network (DQN) [ 36 ], mapping the adaptive persona and feature set ( Table 1 ) to a message category decision. Synthetic data generation followed previously published approaches [ 37 , 38 ]. Synthetic data was generated to simulate different user engagement options. This was done to enable the model to generalize to a wide range of real-world behaviors. This approach was necessary due to the lack of training data. The training dataset included prior user engagement metrics defined in Table 1 , as well as alcohol consumption trends extracted from self-reported drink diaries from prior studies. In total, the training corpus represented over 8,000 drinking patterns. The reward function was designed around key behavioral outcomes, assigning positive reinforcement to actions that increased engagement and reduced alcohol consumption, while applying negative penalties to behaviors that led to increased alcohol consumption. Training diagnostics (average episodic reward progression and mean temporal difference loss) are provided in S1 Fig . The reward function operated as follows: Positive reward (+1 to +10) for app interactions following a message such as opening the app within 1 hour of receiving a notification, completing a goal after receiving encouragement. Stronger positive reward (+10 to +20) for reductions in weekly alcohol consumption or increases drink-free days over consecutive days or weeks. Negative reward (-5 to -10) for disengagement after multiple notifications, preventing message fatigue, increasing consumption, or changing drinking pattern. To prevent overfitting, the model was trained using k -fold cross-validation. The dataset was split into training (80%) and validation (20%) sets, where each fold used multiple iterations to evaluate performance. Hyperparameters were fine-tuned using Bayesian optimization, which iteratively adjusted parameter values to maximize the expected cumulative reward. Training progression was monitored using average episodic reward and mean temporal difference loss; early stopping was applied when validation reward plateaued. The learning rate was optimized within a range of 0.001 to 0.1. The discount factor, which controls the weight of future rewards, was tuned between 0.7 and 0.95 to balance short-term engagement gains with long-term behavioral change. The exploration decay rate was optimized between 0.95 and 0.99, ensuring a gradual shift from exploratory behavior (randomized message selection) to exploitative behavior (prioritizing high-performing messages). Early stopping was implemented to terminate training when improvements in reward accumulation plateaued over a predefined number of iterations. To assess learning progression and policy effectiveness, the average episodic reward was tracked throughout training, providing a performance metric for evaluating the model’s ability to maximize long-term engagement and behavior change. Exploration versus exploitation strategy : We adopted RL because message selection is a sequential decision problem requiring an explicit exploration/exploitation trade-off to optimize longer-term cumulative reward, rather than relying on static rule-based scheduling. The RL model seeks to balance exploration and exploitation using an e-greedy strategy with decay. Initially, a high ε value enabled broad exploration. Over time, ε decays, allowing the model to prioritize exploitation of high-performing message strategies while still testing new approaches. Thompson Sampling [ 39 ] was incorporated to refine message selection, adjusting probabilities based on the estimated effectiveness of different message types. Manual thresholding rules were developed by the research team to ensure certain types of messages did not appear together. These were deployed alongside the framework. Implementation : The framework was implemented as a cloud-based service. DQN was trained offline and deployed; model parameters were not updated. Adaptation during deployment occurred through updates to the rolling 28-day adaptive persona and the exploration/exploitation strategy, with fallback to generic messaging where insufficient engagement data was present (e.g., when the user did not interact with the app). The pipeline was coded in Python, using Azure SDK for model prediction, deployment, and API engagement. The pipeline uses Azure Functions for inference with backend systems. Data storage and retrieval is managed through Google Cloud Services. Feasibility outcomes In line with the aim of assessing feasibility, we examined recruitment and retention over the 12-month period, patterns of app usage (e.g., weekly interaction time) and participant responsiveness to push notifications (e.g., number received and engaged with). Retention was defined as active use of the app for at least 6 of the 12 study weeks. These feasibility indicators were chosen to inform the acceptability, usability, and practical delivery of the intervention. Participants logged their daily alcohol intake, and the app aggregated weekly consumption data. All outcome measures were collected via the DrinksRation app and stored in a cloud database. Weekly alcohol consumption and drink-free days were extracted from the drink’s diary store. Statistical analysis Statistical analyses were performed using STATA, and statistical significance was assessed at p < 0.05 where reported. Descriptive statistics were used to summarize participant characteristics at baseline and the number of notifications received over the 12-week period. Continuous variables are reported as means with standard deviations (SD), while categorical variables are presented as frequencies and percentages. No inferential statistical tests were conducted for baseline group differences [ 40 ]. The primary outcome was weekly alcohol consumption (UK units), self-reported via the in-app drinks diary. This was analyzed using a linear mixed effects model (LMM) with random intercepts for participants to account for within-subject correlation across repeated measures. Fixed effects included intervention group (personalized vs generic messaging), study week (categorical), and their interaction (group * week) to assess differential temporal effects. The model adjusted for age, gender, and history of prior alcohol treatment. Socioeconomic variables were not available in the data, and mental health symptom measures were not included in the primary outcome models due to missingness and because they may plausibly lie on the causal pathway, such that adjustment could risk over-adjustment. Estimated marginal means with 95% confidence intervals (CIs) were calculated for each week (weeks 1–12) within each group. Between-group differences in EMMs were estimated to assess the effect of the personalized intervention over time. Where suitable, pairwise group comparisons by week were conducted using post hoc tests with adjustment. The intent was to assess differences at each outcome point. A secondary analysis evaluated weekly app engagement (total app usage in seconds per week), using the same LMM structure described previously comparing both groups. Fixed effects and interactions were modelled to examine group differences over time in user engagement. To support planning of a future adequately powered trial, we report week-specific between-group effect size estimates (personalized minus no personalization) with 95% CI for alcohol units and app usage in S2 Table . The study was also reported following CONSORT (Consolidated Standards of Reporting Trials [ 41 ] and eHealth version) checklist [ 42 ]. Results Demographics Recruitment took place between May 2020 and July 2021 (see Fig 2 ). A total of 2871 participants were considered for inclusion. Of these, 107 (3.7%) were excluded for not providing baseline demographic data, 235 (8.2%) were excluded for not serving in the UK AF, and 409 (14.2%) were excluded for reporting alcohol consumption below 14 units per week. This resulted in 2120 (73.8%) participants being eligible for allocation, with 1218 (42.4%) assigned to the no personalization group and 902 (31.4%) to the personalization group. Following allocation, 559 (25.6%) participants in the no personalization group and 833 (39.3%) in the personalization group were excluded for having fewer than six weeks of app usage. The final analytic sample was 728 (25.4%) participants, with 385 (13.4%) in the no personalization group and 343 (11.9%) in the personalization group. Fig 2. CONSORT diagram for recruitment into analysis sample. Open in a new tab Table 2 summarizes baseline characteristics of participants. Most of the sample were male ( n = 631, 86.7%), served in the Army ( n = 556, 76.4%) and had not sought any prior treatment for alcohol use/disorder ( n = 636, 87.4%). Baseline characteristics by ≥6-week eligibility are provided in S2 Table . Table 2. Participant demographics stratified by allocation. Variable Overall ( n = 728) Personalization ( n = 343) No Personalization ( n = 385) Age in years (Mean, SD) 48.4 (19.3) 48.3 (19.2) 48.5 (14.4) Gender (n, %) Male 631 (86.7) 298 (86.9) 333 (86.5) Female 97 (13.3) 45 (13.1) 52 (13.5) Service branch (n, %) Army 556 (76.4) 250 (72.9) 306 (79.5) Royal Air Force 82 (11.3) 45 (13.1) 37 (9.6) Royal Navy 90 (12.4) 48 (14.0) 42 (10.9) Phone OS (n, %) iOS 433 (59.5) 204 (59.5) 229 (59.5) Android 295 (40.5) 139 (40.5 156 (40.5) Prior alcohol treatment (n, %) No Yes 636 (87.4) 92 (12.6) 293 (85.4) 50 (14.6) 343 (89.1) 42 (10.9) Open in a new tab Differences over time Table 3 summarizes the estimated marginal means of weekly alcohol consumption across both groups over the 12-week periods. As shown in Fig 3 , alcohol consumption declined over time in both groups. In the personalization group, consumption reduced from 31.08 units (95% CI: 29.41 to 32.74) in Week 1 to 13.20 units (95% CI: 10.23 to 16.18) by Week 12. In the no personalization group, consumption reduced from 31.24 units (95% CI: 29.66 to 32.82) to 15.17 units (95% CI: 11.76 to 18.58) over the same period. Table 3. Estimated mean weekly alcohol consumption (units) by intervention group, with 95% CI and between-group differences. Week Personalized (mean [95% CI]) No personalization (mean [95% CI]) Difference p -value 1 31.08 [29.41 to 32.74] 31.24 [29.66 to 32.82] -0.16 0.888 2 25.25 [23.58 to 26.92] 27.81 [26.23 to 29.39] -2.56 0.027 3 20.87 [19.20 to 22.54] 25.62 [24.04 to 27.20] -4.76 0.041 4 17.83 [16.16 to 19.50] 20.91 [19.33 to 22.49] -3.08 0.008 5 16.13 [14.46 to 17.80] 18.20 [16.61 to 19.79] -2.07 0.074 6 13.91 [12.22 to 15.59] 15.69 [14.10 to 17.28] -1.78 0.126 7 16.75 [14.75 to 18.75] 17.24 [15.31 to 19.17] -0.49 0.729 8 16.47 [14.27 to 18.66] 18.44 [16.18 to 20.70] -1.98 0.214 9 15.06 [12.58 to 17.53] 16.68 [14.04 to 19.32] -1.62 0.376 10 14.21 [11.57 to 16.85] 18.20 [15.21 to 21.20] -4.02 0.049 11 14.12 [11.13 to 17.11] 17.25 [13.86 to 20.63] -3.12 0.174 12 13.20 [10.23 to 16.18] 15.17 [11.76 to 18.58] -1.97 0.392 Open in a new tab Fig 3. Change of alcohol consumption with 95% CI provided for each week. Open in a new tab Significant between-group differences were observed at Week 2 ( p = 0.027), Week 3 ( p = 0.041), Week 4 ( p = 0.008), and Week 10 ( p = 0.049), with the personalization group having lower alcohol unit consumption. However, the trajectories converged by Week 12, with a smaller and non-significant between-group difference at the end of follow-up ( p = 0.392). App usage declined over time in both groups (see Table 4 ). In the personalized group, mean usage decreased from 207.0 seconds (95% CI: 202.0 to 212.1) at Week 1 to 116.7 seconds (95% CI: 107.2 to 126.2) at Week 12. In the no personalization group, app usage reduced from 183.9 seconds (95% CI: 179.1 to 188.8) to 104.1 seconds (95% CI: 93.2 to 115.1) over the same period. Although the personalized group initially had higher engagement, the between-group differences diminished over time. The interaction p -values (week 1 = 0.888, week 6 = 0.126, week 12 = 0.392) suggest that the overall rate of decline in app usage was comparable between the two groups. Table 4. Estimated mean weekly app usage (seconds) by intervention group, with 95% CI and between-group differences. Week Personalized (mean [95% CI]) No personalization (mean [95% CI]) Difference p -value 1 207.04 [201.98 to 212.10] 177.92 [173.14 to 182.70] 29.12 <0.001 2 212.39 [207.32 to 217.45] 183.68 [178.90 to 188.46] 28.71 <0.001 3 192.09 [187.03 to 197.15] 178.14 [173.36 to 182.92] 13.95 <0.001 4 203.12 [198.05 to 208.19] 179.21 [174.43 to 183.99] 23.91 <0.001 5 176.22 [171.15 to 181.29] 169.17 [164.35 to 173.98] 7.06 0.048 6 194.27 [189.15 to 199.40] 171.68 [166.87 to 176.49] 22.6 <0.001 7 171.00 [164.81 to 177.20] 165.59 [159.59 to 171.58] 5.42 0.218 8 144.06 [137.21 to 150.91] 155.92 [148.82 to 163.02] -11.86 0.018 9 159.27 [151.46 to 167.07] 150.47 [142.09 to 158.85] 8.8 0.132 10 128.93 [120.58 to 137.29] 130.98 [121.40 to 140.55] -2.05 0.752 11 117.15 [107.62 to 126.68] 101.04 [90.18 to 111.90] 16.11 0.029 12 116.43 [106.95 to 125.90] 103.90 [92.96 to 114.83] 12.53 0.09 Open in a new tab Notification differences Table 5 summarizes the number of notifications sent to participants. This between-group difference reflects the intervention definition where generic notifications were comparable across groups, and the higher total in the personalization group is attributable to the additional personalized and event-triggered notifications. Participants in the personalization group received more notifications overall (mean = 47.7, SD = 18.8) compared to the no personalization group (mean = 16.3, SD = 5.3). While both groups received a similar number of generic notifications, those in the personalization group also received personalized (mean = 19.8, SD = 11.9) and event-triggered notifications (mean = 11.4, SD = 6.7). Table 5. Descriptive on the number and type of notifications sent stratified by allocation. Overall ( n = 157) Personalization ( n = 343) No Personalization ( n = 385) Generic (Mean, SD) 16.5 (5.4) 16.6 (5.3) Personalized (Mean, SD) 19.8 (11.9) – Personalized (Event) (Mean, SD) 11.4 (6.7) – Total (Mean, SD) 47.7 (18.8) 16.3 (5.3) Open in a new tab Discussion This study evaluated the feasibility of RationAI, a personalized AI-supported messaging framework, integrated within the DrinksRation app, to support alcohol reduction among UK AF veterans. While the behavioral efficacy of the intervention has been reported in a previous RCT [ 23 ], the focus of this work is on the feasibility of delivering this personalized framework in a naturalistic real-world setting over a 12-week period compared to those who did not receive personalization. Feasibility was defined across several dimensions, including participant engagement, retention, consistency of app usage, and delivery of personalized messaging. Of the 2,871 individuals who registered, 25.4% ( n = 728) met the inclusion threshold of at least six weeks of active app use, each with a minimum of three interactions per week. This means that nearly 70% of participants did not meet this threshold, this should be considered in light a brief alcohol intervention and known long term adherence challenges [ 43 , 44 ]. Among those who met the inclusion threshold, app engagement was generally high. Participants in the personalized group received and engaged with a higher volume of notifications, including tailored and event-triggered messages, compared to the control group. This indicates that the personalized framework was able to deliver messages successfully and that these messages were accessed by participants. It is important to note that app usage across the two groups remained similar. This indicates that personalization did not influence app usage. Alcohol consumption decreased in both groups, with earlier between-group differences at several weeks but attenuation towards the end of follow-up (Week 12 difference non-significant). This pattern suggests the personalized messaging package may accelerate early reductions rather than produce a sustained separation at 12 weeks. Convergence may reflect reduced engagement over time, habituation to notifications, diminishing marginal effects of repeated prompts, and/or regression towards a stable level of consumption in both groups. The observed drop-off in engagement, both pre-inclusion and across the 12-week period, reflects a broader challenge in digital health interventions, especially when deployed at scale and without clinician support. Future research should explore strategies to increase retention and promote sustained engagement. These may include social or peer support features, gamified elements, or embedded feedback loops to increase interactivity and perceived value [ 45 ]. Additionally, the use of adaptive notification schedules or behavioral nudges may prevent notification fatigue and maintain motivation throughout the intervention. While this study did not aim to isolate specific behavioral mechanisms, differences between the groups suggest that personalized messaging may contribute to earlier reductions in alcohol use and support feasibility of delivery in a real-world setting, although sustained separation at 12 weeks was not observed. These findings build upon prior evidence that personalization and timing are critical components of digital behavior change interventions [ 15 , 16 , 21 ]. In this study, personalization was implemented via a machine learning algorithm informed by adaptive user personas. The successful deployment and real-time operation of this system support its feasibility and scalability for broader use. There are several limitations of this work. First, because the AI-supported intervention comprises both tailored content and an expanded notification strategy (including event-triggered delivery), this feasibility study cannot disentangle the independent contribution of personalization content versus message dose; a factorial design would be required to isolate these components. Second, the inclusion criteria meant that only a subset of all app users were analyzed, and these individuals may differ systematically from those who disengaged early. While this is common in digital intervention studies [ 46 ], it raises questions about generalizability. Third, alcohol consumption was self-reported via in-app drink diaries, which may be subject to recall or social desirability bias. Future studies may consider integrating passive monitoring tools, such as wearable sensors or linked breathalyzer data, to support objective outcome measurement. Fourth, although this feasibility study was not powered to detect subgroup effects, further work is needed to understand how different groups, by gender, age, or comorbidity, engage with and respond to personalized interventions. The sample was drawn from a real-world population of veterans, many of whom likely experience complex comorbidities, including mental health problems, which are known to affect treatment engagement [ 8 , 9 , 17 ]. Finally, the absence of financial incentives and the fully remote delivery model may also have contributed to higher attrition. Nevertheless, the fact that over 700 participants remained engaged at a meaningful level, interacting with the app regularly over multiple weeks, suggests that the intervention was acceptable to a substantial subset of users. Despite these limitations, the study demonstrates that approximately 30% of veterans were willing to engage with a digital intervention aimed at reducing alcohol consumption, and that nearly half of those remained engaged throughout the intervention period. Given the barriers to help-seeking in the veteran community [ 9 , 47 – 50 ], this level of engagement is encouraging. Our findings suggest that AI-supported messaging strategies hold promise as scalable, acceptable tools for delivering behavior change support in populations with high levels of unmet need. Future research should focus on refining personalization algorithms, extending duration of follow-up, and testing strategies to support longer-term engagement and behavior change. These findings have implications for the design and implementation of digital health interventions targeting alcohol use in the AF community. Generic messaging, while useful for broad implementation, does not adequately address the complex social, psychological and occupational factors influencing alcohol consumption in veteran. AI supported models, provide a data-driven method for optimizing message delivery based on real-world user behavior. The integration of AI supported personalization frameworks allows for more precise, adaptive, and user focused interventions that can better address individual needs and behavioral patterns. Given the high prevalence of alcohol misuse in the AF, such approaches could provide a scalable and cost-effective alternative to conventional interventions. This could include in areas such as common mental health disorders [ 51 ], gambling [ 52 , 53 ] and PTSD [ 54 , 55 ] interventions which do not currently have any personalization approaches tailored to high intensity occupations such as the military. This study provides preliminary evidence supporting the use of AI supported personalized messaging for reducing alcohol consumption in the UK AF. By integrating behavioural data, adaptive user personas, and Just-In-Time interventions, AI can enhance engagement and optimize behaviour change strategies beyond conventional static interventions. These findings highlight the potential of AI-supported digital health solutions in tackling alcohol misuse, particularly in populations facing barriers to traditional care. Future research should focus on long-term efficacy, objective outcome validation, and scalability. Supporting information S1 Fig. Training results. (DOCX) pdig.0001322.s001.docx (418.3KB, docx) S1 Table. Baseline characteristics of ≥6 weeks vs < 6 weeks. (DOCX) pdig.0001322.s002.docx (195.5KB, docx) S2 Table. Effect size estimates. (DOCX) pdig.0001322.s003.docx (197.4KB, docx) Acknowledgments We note with great sadness the passing of our co-author, Professor Nicola T. Fear, before the final version of this publication. Professor Fear made an important contribution to this work, and we are grateful for her scholarship, guidance, and enduring impact on the field. Dr Daniel Leightley accepts responsibility for the integrity and validity of the data collected and analyzed. Data Availability The data has also been uploaded to the Open Science Framework. This can be accessed via https://osf.io/7s6j8 . Funding Statement The author(s) received no specific funding for this work. References 1. Veterans factsheet 2020. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/874821/6.6409_CO_Armed-Forces_Veterans-Factsheet_v9_web.pdf 2. UK Government. UK armed forces biannual diversity statistics: 1 April 2020. 2020. 3. Population projections: UK armed forces veterans residing in Great Britain, 2016 to 2028. UK Government; 2019. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/775151/20190107_Enclosure_1_Population_Projections_-_UK_Armed_Forces_Veterans_residing_in_Great_Britain_-_2016_to_2028.pdf [ Google Scholar ] 4. Stevelink SAM, Jones M, Hull L, Pernet D, MacCrimmon S, Goodwin L, et al. Mental health outcomes at the end of the British involvement in the Iraq and Afghanistan conflicts: a cohort study. Br J Psychiatry. 2018;213(6):690–7. doi: 10.1192/bjp.2018.175 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. UK Government. Delivering better oral health: an evidence-based toolkit for prevention - alcohol guidelines. 2021. https://www.gov.uk/government/publications/delivering-better-oral-health-an-evidence-based-toolkit-for-prevention/chapter-12-alcohol [ Google Scholar ] 6. Palmer L, Norton S, Jones M, Rona RJ, Goodwin L, Fear NT. Trajectories of alcohol misuse among the UK armed forces over a 12-year period. Addiction. 2022;117(1):57–67. doi: 10.1111/add.15592 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Na PJ, Montalvo-Ortiz J, Petrakis I, Krystal JH, Polimanti R, Gelernter J, et al. Trajectories of alcohol consumption in U.S. military veterans: Results from a 10-year population-based longitudinal study. Drug Alcohol Depend. 2023;246:109833. doi: 10.1016/j.drugalcdep.2023.109833 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Elbogen EB, Wagner HR, Johnson SC, Kinneer P, Kang H, Vasterling JJ, et al. Are Iraq and Afghanistan veterans using mental health services? New data from a national random-sample survey. Psychiatr Serv. 2013;64(2):134–41. doi: 10.1176/appi.ps.004792011 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Coleman SJ, Stevelink SAM, Hatch SL, Denny JA, Greenberg N. Stigma-related barriers and facilitators to help seeking for mental health issues in the armed forces: a systematic review and thematic synthesis of qualitative literature. Psychol Med. 2017;47(11):1880–92. doi: 10.1017/S0033291717000356 [ DOI ] [ PubMed ] [ Google Scholar ] 10. NHS Providers. NHS providers’ written evidence to the house of commons general committee on the health and care bill. 2021. 11. Fowler LA, Holt SL, Joshi D. Mobile technology-based interventions for adult users of alcohol: a systematic review of the literature. Addict Behav. 2016;62:25–34. doi: 10.1016/j.addbeh.2016.06.008 [ DOI ] [ PubMed ] [ Google Scholar ] 12. Garnett C, Crane D, West R, Brown J, Michie S. The development of Drink Less: an alcohol reduction smartphone app for excessive drinkers. Transl Behav Med. 2019;9(2):296–307. doi: 10.1093/tbm/iby043 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Attwood S, Parke H, Larsen J, Morton KL. Using a mobile health application to reduce alcohol consumption: a mixed-methods evaluation of the drinkaware track and calculate units application. BMC Public Health. 2017;17(1):394. doi: 10.1186/s12889-017-4358-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Williamson G, Khatun T, King K, Simms A, Dymond S, Goodwin L, et al. Digital health interventions for women in frontline public service roles: a systematic review of effectiveness in reducing substance use. PLOS Digit Health. 2026;5(1):e0001154. doi: 10.1371/journal.pdig.0001154 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Kaner EF, Beyer FR, Garnett C, Crane D, Brown J, Muirhead C, et al. Personalised digital interventions for reducing hazardous and harmful alcohol consumption in community-dwelling populations. Cochrane Database Syst Rev. 2017;9(9):CD011479. doi: 10.1002/14651858.CD011479.pub2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Williamson C, White K, Rona RJ, Simms A, Fear NT, Goodwin L, et al. Smartphone-based alcohol interventions: a systematic review on the role of notifications in changing behaviors toward alcohol. Subst Abus. 2022;43(1):1231–44. doi: 10.1080/08897077.2022.2074595 [ DOI ] [ PubMed ] [ Google Scholar ] 17. Jones E, Fear NT. Alcohol use and misuse within the military: a review. Int Rev Psychiatry. 2011;23(2):166–72. doi: 10.3109/09540261.2010.550868 [ DOI ] [ PubMed ] [ Google Scholar ] 18. Middleton SE, et al. AI for defence: readiness, resilience and mental health. RUSI Journal. 2024;169:52–62. [ Google Scholar ] 19. Aggarwal A, Tam CC, Wu D, Li X, Qiao S. Artificial intelligence-based chatbots for promoting health behavioral changes: systematic review. J Med Internet Res. 2023;25:e40789. doi: 10.2196/40789 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Leightley D, Murphy D. Personalised digital technology for mental health in the armed forces: the potential, the hype and the dangers. BMJ Mil Health. 2023;169(1):406–8. doi: 10.1136/military-2022-002279 [ DOI ] [ PubMed ] [ Google Scholar ] 21. Steenstra I, Nouraei F, Arjmand M, Bickmore TW. Virtual agents for alcohol use counseling: exploring LLM-powered motivational interviewing. 2024. doi: 10.1145/3652988.3673932 [ DOI ] [ Google Scholar ] 22. Garnett CV, et al. Behavior change techniques used in digital behavior change interventions to reduce excessive alcohol consumption: a meta-regression. Annalf Behav Med. 2018;52:530–43. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Leightley D, Williamson C, Rona RJ, Carr E, Shearer J, Davis JP, et al. Evaluating the efficacy of the drinks:ration mobile app to reduce alcohol consumption in a help-seeking military veteran population: randomized controlled trial. JMIR Mhealth Uhealth. 2022;10(6):e38991. doi: 10.2196/38991 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Leightley D, Rona RJ, Shearer J, Williamson C, Gunasinghe C, Simms A, et al. Evaluating the efficacy of a mobile app (Drinks:Ration) and personalized text and push messaging to reduce alcohol consumption in a veteran population: protocol for a randomized controlled trial. JMIR Res Protoc. 2020;9(10):e19720. doi: 10.2196/19720 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Williamson C, Dryden D, Palmer L, Rona R, Simms A, Fear NT, et al. An expert and veteran user assessment of the usability of an alcohol reduction app for military veterans, drinks: ration: a mixed-methods pilot study. Military Behavioral Health. 2022;11(1–2):14–27. doi: 10.1080/21635781.2022.2151532 [ DOI ] [ Google Scholar ] 26. Mattoo SK, Prasad S, Ghosh A. Brief intervention in substance use disorders. Indian J Psychiatry. 2018;60(Suppl 4):S466–72. doi: 10.4103/0019-5545.224352 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Beyer F, Lynch E, Kaner E. Brief Interventions in primary care: an evidence overview of practitioner and digital intervention programmes. Curr Addict Rep. 2018;5(2):265–73. doi: 10.1007/s40429-018-0198-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Doherty AM, Mason C, Fear NT, Rona R, Greenberg N, Goodwin L. Are brief alcohol interventions targeting alcohol use efficacious in military and veteran populations? A meta-analysis. Drug Alcohol Depend. 2017;178:571–8. doi: 10.1016/j.drugalcdep.2017.05.029 [ DOI ] [ PubMed ] [ Google Scholar ] 29. Puddephatt J-A, Leightley D, Palmer L, Jones N, Mahmoodi T, Drummond C, et al. A Qualitative evaluation of the acceptability of a tailored smartphone alcohol intervention for a military population: information about drinking for ex-serving personnel (InDEx) app. JMIR Mhealth Uhealth. 2019;7(5):e12267. doi: 10.2196/12267 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Leightley D, et al. A smartphone app and personalized text messaging framework (InDEx) to monitor and reduce alcohol use in ex-serving personnel: Development and feasibility study. JMIR Mhealth Uhealth. 2018. doi: 10.2196/10074 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Craig P. Developing and evaluating complex interventions: new guidance. 2019. https://mrc.ukri.org/documents/pdf/complex-interventions-guidance [ Google Scholar ] 32. Leightley D. Navigating the software development landscape: Introducing the term research viable product (RVP). 2024. doi: 10.17605/OSF.IO/5N37J [ DOI ] [ Google Scholar ] 33. Williamson G, Carr E, Fear NT, Dymond S, King K, Simms A, et al. Digital therapeutic intervention for women in the UK armed forces who consume alcohol at a hazardous or harmful level: protocol for a randomized controlled trial. JMIR Res Protoc. 2023;12:e51531. doi: 10.2196/51531 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Irizar P, Leightley D, Stevelink S, Rona R, Jones N, Gouni K, et al. Drinking motivations in UK serving and ex-serving military personnel. Occup Med (Lond). 2020;70(4):259–67. doi: 10.1093/occmed/kqaa003 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Puddephatt J-A, Leightley D, Palmer L, Jones N, Mahmoodi T, Drummond C, et al. A Qualitative evaluation of the acceptability of a tailored smartphone alcohol intervention for a military population: information about drinking for ex-serving personnel (InDEx) app. JMIR Mhealth Uhealth. 2019;7(5):e12267. doi: 10.2196/12267 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Roderick M, MacGlashan J, Tellex S. Implementing the deep Q-network. 2017. 37. Dankar FK, Ibrahim M. Fake it till you make it: guidelines for effective synthetic data generation. Appl Sci. 2021;11(5):2158. doi: 10.3390/app11052158 [ DOI ] [ Google Scholar ] 38. Jordon J. Synthetic data -- what, why and how?. ArXiv. 2022. doi: 10.48550/arXiv.2205.03257 [ DOI ] [ Google Scholar ] 39. Daniel JR, Benjamin VR, Abbas K, Ian O, Zheng W. A tutorial on thompson sampling. Found Trends Machine Learn. 2018;11(1):1–99. doi: 10.1561/2200000070 [ DOI ] [ Google Scholar ] 40. de Boer MR, Waterlander WE, Kuijper LDJ, Steenhuis IHM, Twisk JWR. Testing for baseline differences in randomized controlled trials: an unhealthy research behavior that is hard to eradicate. Int J Behav Nutr Phys Act. 2015;12:4. doi: 10.1186/s12966-015-0162-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Schulz KF, Altman DG, Moher D. CONSORT 2010 statement: updated guidelines for reporting parallel group randomised trials. BMJ. 2010;340(mar23 1):c332–c332. doi: 10.1136/bmj.c332 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Eysenbach G, CONSORT-EHEALTH Group. CONSORT-EHEALTH: improving and standardizing evaluation reports of Web-based and mobile health interventions. J Med Internet Res. 2011;13(4):e126. doi: 10.2196/jmir.1923 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Christensen H, Griffiths KM, Farrer L. Adherence in internet interventions for anxiety and depression. J Med Internet Res. 2009;11(2):e13. doi: 10.2196/jmir.1194 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Linardon J, Fuller-Tyszkiewicz M. Attrition and adherence in smartphone-delivered interventions for mental health problems: a systematic and meta-analytic review. J Consult Clin Psychol. 2020;88(1):1–13. doi: 10.1037/ccp0000459 [ DOI ] [ PubMed ] [ Google Scholar ] 45. Gonul S, Namli T, Huisman S, Laleci Erturkmen GB, Toroslu IH, Cosar A. An expandable approach for design and personalization of digital, just-in-time adaptive interventions. J Am Med Inform Assoc. 2019;26(3):198–210. doi: 10.1093/jamia/ocy160 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Kaner EF. Effectiveness of brief alcohol interventions in primary care populations. Cochrane Database Syst Rev. 2018. doi: 10.1002/14651858.CD004148.pub4 [ DOI ] [ PubMed ] [ Google Scholar ] 47. Betton V, Borschmann R, Docherty M, Coleman S, Brown M, Henderson C. The role of social media in reducing stigma and discrimination. Br J Psychiatry. 2015;206(6):443–4. doi: 10.1192/bjp.bp.114.152835 [ DOI ] [ PubMed ] [ Google Scholar ] 48. Murphy D, Busuttil W. PTSD, stigma and barriers to help-seeking within the UK armed forces. J R Army Med Corps. 2015;161(4):322–6. doi: 10.1136/jramc-2014-000344 [ DOI ] [ PubMed ] [ Google Scholar ] 49. Sharp M-L, Fear NT, Rona RJ, Wessely S, Greenberg N, Jones N, et al. Stigma as a barrier to seeking health care among military personnel with mental health problems. Epidemiol Rev. 2015;37:144–62. doi: 10.1093/epirev/mxu012 [ DOI ] [ PubMed ] [ Google Scholar ] 50. Williamson V, Greenberg N, Stevelink SAM. Perceived stigma and barriers to care in UK Armed Forces personnel and veterans with and without probable mental disorders. BMC Psychol. 2019;7(1):75. doi: 10.1186/s40359-019-0351-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Parkes S, Croak B, Brooks SK, Stevelink SAM, Leightley D, Fear NT, et al. Evaluating a smartphone app (MeT4VeT) to support the mental health of UK armed forces veterans: feasibility randomized controlled trial. JMIR Ment Health. 2023;10:e46508. doi: 10.2196/46508 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Dymond S, Dighton G, Wood K. The United Kingdom armed forces veterans’ health and gambling study. 2021. 53. Hitch C, Leightley D, Murphy D, Trompeter N, Dymond S. Acceptance and commitment therapy for co-occurring gambling disorder and posttraumatic stress disorder in veterans: a narrative review. Eur J Psychotraumatol. 2023;14(1):2178203. doi: 10.1080/20008066.2023.2178203 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. National Center for PTSD. Mobile app: PTSD coach. https://www.ptsd.va.gov/appvid/mobile/ptsdcoach_app.asp 55. Leightley D, Williamson V, Darby J, Fear NT. Identifying probable post-traumatic stress disorder: applying supervised machine learning to data from a UK military cohort. J Ment Health. 2019;28(1):34–41. doi: 10.1080/09638237.2018.1521946 [ DOI ] [ PubMed ] [ Google Scholar ] PLOS Digit Health. doi: 10.1371/journal.pdig.0001322.r001 Decision Letter 0 Benjamin Filtjens Benjamin Filtjens Academic Editor Find articles by Benjamin Filtjens Author information Copyright and License information Roles Benjamin Filtjens : Academic Editor © 2026 Benjamin FiltjensBenjamin FiltjensBenjamin FiltjensBenjamin Filtjens This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 15 Dec 2025 Response to Reviewers '. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.'. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.-->-->* A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled ' Revised Manuscript with Track Changes '.'.-->-->* An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled ' Manuscript '.'.-->--> -->-->If you would like to make changes to your financial disclosure, competing interests statement, or data availability statement, please make these updates within the submission form at the time of resubmission. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.-->--> -->-->We look forward to receiving your revised manuscript.-->--> -->-->Kind regards,-->--> -->-->Benjamin Filtjens-->-->Academic Editor-->-->PLOS Digital Health-->--> -->-->Leo Anthony Celi-->-->Editor-in-Chief-->-->PLOS Digital Health-->-->orcid.org/0000-0001-6712-6626-->--> -->--> Journal Requirements: -->--> -->-->If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. -->--> -->--> Additional Editor Comments (if provided): -->--> -->--> -->--> -->-->[Note: HTML markup is below. Please do not edit.]-->--> -->--> Reviewers' Comments: -->--> -->-->Reviewer's Responses to Questions Comments to the Author 1. Does this manuscript meet PLOS Digital Health’s publication criteria ? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.-->?> Reviewer #1: Partly Reviewer #2: Yes ********** 2. Has the statistical analysis been performed appropriately and rigorously?-->?> Reviewer #1: I don't know Reviewer #2: Yes ********** 3. Have the authors made all data underlying the findings in their manuscript fully available (please refer to the Data Availability Statement at the start of the manuscript PDF file)??> The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: No Reviewer #2: Yes ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English??> Reviewer #1: Yes Reviewer #2: Yes ********** Reviewer #1: The paper presents a feasibility study for RationAI, a personalized AI-based message recommendation system, integrated with the DrinksRation mobile phone app. The goal of the AI framework is to increase the effectiveness of the DrinksRation app, with a study conducted among UK Armed Forces veterans. Given my profile, I will mainly comment on the AI part. In general the paper is well written, however I have two main concerns. 1. Table 5 highlights that the `No Personalization` group received much fewer notifications w.r.t. the `Personalization` group. This creates a problem: is the higher effectiveness in lowering alcohol consumption seen in the `Personalization` group given by the PERSONALIZATION of messages, or simply by the higher amount of received messages? Unfortunately, this makes it difficult to assess the effectiveness of the RL model. 2. The training and deployment of the RL model can be detailed more. What's the overall size of the training set? Are the more details on the used synthetic data, as well as the researcher-curated data? Is there any training metric available, like loss, etc? Which specific RL algorithm is used and: is RL really necessary in this setting? Was the model adapted during deployment, or was it trained completely offline before the actual use? Reviewer #2: The authors intend to address an important public health issue on alcohol misuse among UK Armed Forces veterans by evaluating the feasibility of an AI-driven personalized messaging framework integrated into a mobile app. The study is innovative by leveraging reinforcement learning for personalization. The paper is overall well-written and structured. While the authors provide strong technical rigor and clear reporting, there are some limitations in the method, and the results need to be addressed. 1) There is a lack of description on the AI personalization algorithm performance metrics, for example, accuracy, reward progression. 2) It is unclear why there is no comparison between the characteristics of participants who dropped out and those who completed at least 6 weeks of app usage. 3) It is unclear why known factors such as mental health comorbidities and socioeconomic factors impacting alcohol usage were not included and adjusted beyond age, gender, and prior treatment. 4) It would be interesting to include subgroup analyses (for example, gender, age, prior treatment) to explore differential effects. 5) While the authors reported significant differences at some weeks, the overall trend converges by Week 12. This needs to be further discussed. 6) It would be beneficial to report effect sizes for the results, which is necessary for future large clinical trials to estimate the sample size. 7) Are there any data on user satisfaction or perceived relevance of messages? It would be helpful to examine the qualitative feedback on the app usage or non-use. ********** what does this mean? ). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files. Do you want your identity to be public for this peer review? If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public. For information about this choice, including consent withdrawal, please see our Privacy Policy ..--> Reviewer #1: No Reviewer #2: Yes ********** Figure resubmission: -->--> -->--> -->While revising your submission, we strongly recommend that you use PLOS’s NAAS tool ( https://ngplosjournals.pagemajik.ai/artanalysis ) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix.--> --> Reproducibility: -->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols-->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols-->?> PLOS Digit Health. 2026 Apr 10;5(4):e0001322. doi: 10.1371/journal.pdig.0001322.r002 Author response to Decision Letter 1 Article notes Copyright and License information Collection date 2026 Apr. PMC Copyright notice 23 Jan 2026 Attachment Submitted filename: response-reviewers.docx pdig.0001322.s005.docx (44.7KB, docx) PLOS Digit Health. doi: 10.1371/journal.pdig.0001322.r003 Decision Letter 1 Benjamin Filtjens Benjamin Filtjens Academic Editor Find articles by Benjamin Filtjens Author information Copyright and License information Roles Benjamin Filtjens : Academic Editor © 2026 Benjamin FiltjensBenjamin FiltjensBenjamin FiltjensBenjamin Filtjens This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 8 Mar 2026 Reducing alcohol consumption in UK Armed Forces veterans: feasibility of using personalized push notifications using AI PDIG-D-25-00448R1 Dear Dr. Leightley, We are pleased to inform you that your manuscript 'Reducing alcohol consumption in UK Armed Forces veterans: feasibility of using personalized push notifications using AI' has been provisionally accepted for publication in PLOS Digital Health. Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow-up email from a member of our team. Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated. IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they'll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact [email protected]. Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Digital Health. Best regards, Benjamin Filtjens Academic Editor PLOS Digital Health *********************************************************** Additional Editor Comments (if provided): Reviewer Comments (if any, and for reference): Reviewer's Responses to Questions Comments to the Author Reviewer #1: All comments have been addressed ********** publication criteria ? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.-->?> Reviewer #1: Yes ********** 3. Has the statistical analysis been performed appropriately and rigorously?-->?> Reviewer #1: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available (please refer to the Data Availability Statement at the start of the manuscript PDF file)??> The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: Yes ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English??> Reviewer #1: Yes ********** Reviewer #1: (No Response) ********** what does this mean? ). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files. Do you want your identity to be public for this peer review? If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public. For information about this choice, including consent withdrawal, please see our Privacy Policy ..--> Reviewer #1: No ********** Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials S1 Fig. Training results. (DOCX) pdig.0001322.s001.docx (418.3KB, docx) S1 Table. Baseline characteristics of ≥6 weeks vs < 6 weeks. (DOCX) pdig.0001322.s002.docx (195.5KB, docx) S2 Table. Effect size estimates. (DOCX) pdig.0001322.s003.docx (197.4KB, docx) Attachment Submitted filename: response-reviewers.docx pdig.0001322.s005.docx (44.7KB, docx) Data Availability Statement The data has also been uploaded to the Open Science Framework. This can be accessed via https://osf.io/7s6j8 . Articles from PLOS Digital Health are provided here courtesy of PLOS ACTIONS View on publisher site PDF (921.9 KB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top