G. J. Melendez-Torres

Background: AI is a broad field encompassing various techniques, which include algorithms that learn from data to perform automated tasks without explicit human programming. Public health aims to protect and improve the health of communities and populations at the local, regional, national, and global levels, but it typically has limited resources. While AI may lead to improved efficiency or effectiveness of public health interventions, the adoption of AI interventions should be informed by evidence of effective use in practice. Objective: This review aims to summarize how AI-assisted interventions are currently used by public health professionals. Methods: We conducted a scoping review to identify the current uses of AI in public health. We searched MEDLINE (Ovid), Embase (Ovid), Web of Science (Core Collection), and Scopus for relevant reports published between January 1, 2020, and July 24, 2025. Two reviewers independently screened reports. Eligible reports were primary studies (experimental or observational), commentaries, and editorials that discussed how AI tools have been used in practice in public health contexts relevant to high-income countries, excluding outbreaks. Data were extracted by one reviewer and checked by a second. Results: We screened 5400 records and 155 full texts, including 8 reports. These reports discussed a variety of uses of AI tools across public health domains: 3 reports used chatbots in different settings, 1 report used machine learning to forecast health service demand, 2 reports used AI for symptom checking, and 2 reports used AI for evidence synthesis. While assessing the effectiveness of AI tools for public health was beyond the scope of this review, and most reports did not robustly evaluate effectiveness, the evidence for the effectiveness of AI tools was nonetheless mixed. One report described a chatbot that users found helpful, while another found a specific chatbot to be ineffective, highlighting its potential to undermine trust in the organization that developed it. Another report highlighted both benefits and drawbacks of a large language model used to communicate with socially isolated individuals. The main finding of this review was that even with a comprehensive search across multiple databases, there were few reports of how AI tools are being used in practice in public health contexts. Conclusions: The lack of evidence from public health practice, excluding outbreaks, makes it difficult to know which uses of AI, if any, would be useful to incorporate into practice in high-income countries. In the absence of evidence, new interventions, including those using AI, should be evaluated to ensure they are effective (and cost-effective). These evaluations should be widely and accessibly disseminated to create an evidence base that other public health teams could use to inform their own decisions around AI tools.

Background: Digital interventions offer a promising avenue for scalable parenting support in low- and middle-income countries, where in-person coverage remains limited. However, engagement disparities may reduce their potential for population-level impact. Objective: This study examined caregiver-level predictors of engagement with ParentApp, a smartphone-based adaptation of the Parenting for Lifelong Health (PLH) for Teens program for caregivers of adolescents aged 10 to 17 years. Methods: This study was nested within a cluster-randomized factorial trial of ParentApp in Mwanza, Tanzania, implemented according to the optimization phase of the Multiphase Optimization Strategy. Clusters (n=16), comprising 614 caregivers (n=205, 33.4% men), were randomized to 3 components: guidance (guided vs self-guided), app design (unstructured vs structured), and digital support (enhanced vs basic). Digitally tracked app engagement outcomes included the number of modules completed, the first module not started, and the first module not completed. Baseline candidate predictors included demographic, socioeconomic, behavioral, and psychological factors assessed via an app-embedded questionnaire. Poisson generalized linear mixed-effects models with random effects for clusters assessed associations between predictors and each outcome. Interaction terms were specified to test whether associations differed by caregiver gender. Results: Among 614 caregivers, 596 (97.1%) completed the first module; however, engagement declined rapidly thereafter, with only 281 (45.8%) completing module 2 and 49 (7.9%) completing all 12 modules. Overall, mean content completion was 35.6% (SD 33.1%) of the program. Women completed approximately 15% more modules than men (incidence rate ratio 1.15, 95% CI 1.05-1.27, P =.003), despite reporting higher levels of financial stress, food insecurity, child maltreatment, and depressive symptoms at baseline (all P ≤.02; Cohen d =0.19‐0.37). No other caregiver characteristics were significantly associated with engagement. Conclusions: This is the first known study to investigate caregiver-level predictors of engagement with a digital parenting intervention in a low- and middle-income country. Engagement with ParentApp was broadly consistent across caregiver subgroups, with the exception that women completed significantly more modules than men. To equitably engage men and fathers, future studies should pair father-inclusive, identity-affirming content with implementation strategies that reduce social and economic barriers to participation.

The Behavior Change Independent Domestic Violence Advisor Program is a Domestic Abuse Perpetrator Program operating in England. This evaluation drew upon semistructured interviews with program professionals ( n = 7), perpetrators ( n = 7) and victims/survivors ( n = 5) to develop a realist model of how perpetrator behavior change is thought to be enabled through this program. Key contextual factors were identified which moderate program delivery and lead to the conditions required to initiate six mechanisms of change which could lead to outcomes of a reduction in abusive behaviors, reoffending, and harm to victims/survivors.

Background: A 12‐month trial economic evaluation was conducted to compare the Triple P–Positive Parenting Program (Triple P) versus treatment as usual (TAU) and the Circle of Security Parenting Program (COSP) versus TAU from a societal perspective. Methods: Trial data were used to determine the costs of delivering the interventions and the 12‐month costs associated with children's health and social care service use based on the Child and Adolescent Service Use Schedule questionnaire, which was completed by caregivers of children aged 2–6. All costs were expressed in 2024 Canadian dollars. To calculate quality‐adjusted‐life‐years (QALYs), caregivers documented their children's quality of life using the Health Status Classification System Preschool Version (HSCSPV) and the Health Utility Index 3 for children less or more than 5 years of age, respectively. Cost‐effectiveness acceptability curves were used to represent the uncertainty. Several sensitivity analyses were conducted (e.g., change in Triple P intervention costs). Results: The trial recruited 502 caregivers (93% female, mean age [standard deviation] of 36 years [6.7]) of 502 children (41% female, mean age of 3.8 years [SD = 1.4]). The differences in 12‐month societal costs and QALYs between Triple P and TAU were $121 (95% Confidence interval [CI]: −$258, $486) and 0.009 (95% CI: −0.018, 0.039), respectively. The differences in 12‐month societal costs and QALYs between COSP and TAU were −$253 (95% CI: −$552, $40) and 0.004 (95% CI: −0.026, 0.035), respectively. Assuming that society was willing to spend up to $100,000 for each additional QALY gained, the probabilities that Triple P and COSP were cost‐effective compared to TAU were 0.71 and 0.68, respectively. The results of the sensitivity analyses were consistent with the main analyses. Conclusions: The trial results suggest that compared to TAU, Triple P and COSP represent efficient use of public resources.

BACKGROUND Artificial intelligence (AI) is a broad field encompassing various techniques, which include algorithms that learn from data to perform automated tasks without explicit human programming. Public health aims to protect and improve the health of communities and populations at local, regional, national, and global level, but typically suffers from limited resources. AI may lead to improved efficiency or effectiveness of public health interventions, but there needs to be appropriate consideration of the risks of implementation. OBJECTIVE This scoping review aims to identify the potential roles and risks of AI in public health. METHODS We searched Medline, Embase, Web of Science (Core Collection), and Scopus for relevant reports published between 01 January 2020 and 08 September 2025. Two reviewers independently screened reports. Eligible reports were primary studies, reviews, opinion pieces, commentaries, and editorials relevant to high income countries. Data were extracted by one reviewer and checked by a second. We combined substantively similar potential roles and risks by narrative theme across reports. We refined the protocol and gained feedback on initial findings from stakeholder consultation with 9 people from Health Determinants Research Collaborations, and 9 people from data science institutes, public health agencies, and national Local Government networks. RESULTS We screened 1,805 records and 320 full texts, including 67 reports. These reports considered 44 potential roles of AI across eight public health contexts. Potential roles primarily included using machine learning to make predictions, analyse data, and make recommendations, and using large language models to help tailor educational content or communications. Most frequently indicated risks were bias and generalisability, confidentiality and ethical concerns, and exacerbation of inequalities. Specifically, AI could lead to perpetuation or reinforcement of health inequalities and discrimination, a reduction in public trust for public health communications and interventions, and inaccurate predictions, unintended consequences, or ineffective interventions. CONCLUSIONS While we did not consider how plausible or beneficial the identified potential roles of AI in public health may ultimately be, when adopting any new intervention, the intervention should be evaluated for anticipated effectiveness (and cost-effectiveness), with careful consideration of the risks. Future research needs to evaluate the effectiveness of AI interventions across the range of public health domains, with a focus on assessing the potential of inequity generating impacts and harms. There needs to be meaningful community engagement and co-production with diverse and underrepresented people to help mitigate against inequity (and improve trust) when adopting AI tools. Stakeholders emphasised the importance of examining how interventions can be optimally diffused and implemented within public health systems to maximise the likelihood of any realising any positive effects. CLINICALTRIAL Zenodo 15971614

BACKGROUND Prevalence of physical inactivity is high and increasing across high-income countries, and interventions to increase physical activity can include behaviour change delivered digitally. Artificial intelligence (AI) is a broad field encompassing various techniques, which include algorithms that learn from data to perform automated tasks without explicit human programming, and could potentially increase the effectiveness of digital behaviour change interventions through personalisation and reducing barriers to engagement. OBJECTIVE This review aims to summarise the evidence for the effectiveness of AI-assisted interventions to reduce physical inactivity. METHODS We conducted a systematic review to identify and summarise evidence from randomised controlled trials (RCTs) of AI-assisted interventions for reducing physical inactivity. Eligible trials were RCTs that reported results of an AI-assisted public health intervention for reducing physical inactivity in a high-income country. We searched Medline (Ovid), Embase (Ovid), Web of Science (Core collection), and Scopus for relevant trials published between 2010 and 14 January 2025. We also searched for reviews of public health interventions for physical inactivity published between 2023 and 14 January 2025, and extracted all references from relevant reviews for screening. We conducted forward and backward citation searching on all included trials (dates of searches: September 2025 to June 2026). Screening for trials was conducted independently by two reviewers. Two reviewers independently assessed risk of bias using the Cochrane Risk of Bias 2 tool. As the included trials were heterogeneous in terms of interventions, outcomes, and timepoints, we synthesised the results narratively. RESULTS We included 17 trials (comprising 48 reports and 3,282 randomised participants). Five trials estimated the effectiveness of apps with chatbots for reducing physical inactivity (one with some concerns of bias, four with high risks of bias), with little evidence to suggest that chatbots increase physical activity. Twelve trials estimated the effectiveness of selection of motivational or other messages using recommender systems, reinforcement learning, machine learning, or case-based reasoning for reducing physical inactivity (four with some concerns of bias, eight with a high risk of bias), with little evidence to suggest an increase in physical activity generally, though some evidence to suggest an increase in step count specifically. All trials had relatively few participants, so results were generally imprecise. There were no trials using large language models. CONCLUSIONS There is no strong evidence of a beneficial effect of AI-assistance in public health interventions for reducing physical inactivity in high-income countries, though AI-assisted interventions may increase step count. Future research should better describe public health interventions that use AI and embed equity considerations into their design and analysis to ensure already disadvantaged groups are not harmed further by the adoption of AI-assisted interventions in public health. CLINICALTRIAL PROSPERO CRD42025642339

BACKGROUND Artificial intelligence (AI) is a broad field encompassing various techniques, which include algorithms that learn from data to perform automated tasks without explicit human programming. Public health aims to protect and improve the health of communities and populations at local, regional, national, and global level, but typically suffers from limited resources. While AI may lead to improved efficiency or effectiveness of public health interventions, adoption of AI interventions should be informed by evidence of effective use in practice. OBJECTIVE This review aims to summarise how AI-assisted interventions are currently used by public health professionals. METHODS We conducted a scoping review to identify the current uses of AI in public health. We searched Medline (Ovid), Embase (Ovid), Web of Science (Core Collection), and Scopus for relevant reports published between 01 January 2020 and 24 July 2025. Two reviewers independently screened reports. Eligible reports were primary studies (experimental or observational), commentaries, and editorials that discussed how AI tools have been used in practice in public health contexts relevant to high income countries. Data were extracted by one reviewer and checked by a second. RESULTS We screened 5,400 records and 155 full texts, including 8 reports. These reports discussed a variety of uses of AI tools across public health domains: three reports used chatbots in different settings, one report used machine learning to forecast health service demand, two reports used AI for symptom checking, and two reports used AI for evidence synthesis. While assessing the effectiveness of AI tools for public health was beyond the scope of this review, and most reports did not robustly evaluate effectiveness, the evidence for effectiveness of AI tools was nonetheless mixed. One report described a chatbot that users found helpful, while another found a specific chatbot to be ineffective, highlighting its potential to undermine trust in the organisation that developed it. Another report highlighted both benefits and drawbacks of a large language model used to communicate with socially isolated individuals. The main finding of this review was that even with a comprehensive search across multiple databases, there were few reports of how AI tools are being used in practice in public health contexts. CONCLUSIONS The lack of evidence from public health practice makes it difficult to know which uses of AI, if any, would be useful to adopt into practice. In the absence of evidence, new interventions, including those using AI, should be evaluated to ensure they are effective (and cost-effective). These evaluations should be widely and accessibly disseminated to create an evidence base that other public health teams could use to inform their own decisions around AI tools. CLINICALTRIAL Zenodo 15971109

BACKGROUND Prevalence of alcohol consumption and smoking is high across high-income countries, and interventions to decrease either can include behaviour change delivered digitally. Artificial intelligence (AI) is a broad field encompassing various techniques, which include algorithms that learn from data to perform automated tasks without explicit human programming, and could potentially increase the effectiveness of digital behaviour change interventions through personalisation and reducing barriers to engagement. OBJECTIVE This systematic review aims to summarise the evidence for the effectiveness of AI-assisted interventions to reduce alcohol consumption and smoking. METHODS We conducted a systematic review to identify and summarise evidence from randomised controlled trials (RCTs) of AI-assisted interventions for reducing alcohol consumption or smoking. Eligible trials were RCTs that reported results of an AI-assisted public health intervention for reducing engagement with alcohol consumption or smoking in a high-income country. We searched Medline (Ovid), Embase (Ovid), Web of Science (Core collection), and Scopus for relevant trials published between 2010 and 14 January 2025. We also searched for reviews of public health interventions for alcohol consumption or smoking published between 2023 and 14 January 2025, and extracted all references from relevant reviews for screening. We conducted forward and backward citation searching on all included trials (dates of searches: July to September 2025). Screening for trials was conducted independently by two reviewers. Two reviewers independently assessed risk of bias using the Cochrane Risk of Bias 2 tool. As the included trials were heterogeneous in terms of interventions, outcomes, and timepoints, we synthesised the results narratively. RESULTS We included 4 trials for reducing alcohol consumption (comprising 16 reports and 4,718 randomised participants): 3 trials used apps with chatbots and reported mixed evidence, and one trial, which was effective, used a rules-based AI that tailored website content. We included 12 trials for stopping smoking (comprising 31 reports and 68,659 randomised participants): 8 trials of apps or messenger chatbots and 4 trials of recommender systems, all of which had mixed evidence. For many trials, the interventions had multiple components, of which AI-assistance was only one, meaning the effectiveness of AI-assistance specifically could not be determined. Additionally, high risks of bias across almost all trials reduced confidence in the results. There were no trials using large language models (LLMs). CONCLUSIONS There is no strong evidence of a beneficial effect of AI-assistance in public health interventions for reducing alcohol consumption or smoking in high-income countries. Future research should better describe public health interventions that use AI and embed equity considerations into their design and analysis to ensure already disadvantaged groups are not harmed further by the adoption of AI-assisted interventions in public health. CLINICALTRIAL PROSPERO CRD42025642317 and CRD42025642334.