A search was conducted in the CDC’s Prevention Research Synthesis (PRS) Project database. The PRS database is a collection of HIV prevention literature focused on behavioral risk reduction, medication adherence, linkage/retention/re-engagement in HIV care, structural interventions, PrEP, and systematic reviews.¹⁶ By the end of August 2023, the PRS database had amassed ~123,000 citations. The automated search is implemented annually using the following databases: MEDLINE (OVID), EMBASE (OVID), CINAHL (EBSCOhost), Global Health (OVID), PsycINFO (OVID), and Sociological Abstracts (ProQuest) (Appendix). The search was developed in the MEDLINE (OVID) database using indexing and keyword terms cross-referenced by using Boolean logic with no language limits. The finalized search was tailored to other databases to adhere to each proprietary indexing system.¹⁷ Supplementary searches included a manual search of journals (available from the PRS website), online non-indexed databases (e.g., Scopus), gray literature (e.g., Google Scholar), and reference lists from relevant HIV literature.

For this review, the librarian searched the PRS database on August 29, 2023, for literature published 2012 to the present. There were two components used in the search. The first component looked at AI and included phrases centered on “machine learning” and “natural language.” The second component focused on PrEP. The search query looked specifically at the title, abstract, and indexing terms for any concepts related to PrEP and AI. Additionally, we searched RePORTER (https://reporter.nih.gov/) for pipeline and potential publications and ClinicalTrials.gov (https://clinicaltrials.gov/) and International Standard Randomized Controlled Trial Number registry (https://www.isrctn.com/) for studies that were registered but had not been completed or published. We also searched for any newly published literature in PubMed, Scopus, and Google Scholar by using the same search terms (searched November 2, 2023). See Appendix for full PrEP annual searches, PRS database queries, and supplementary searches conducted for this review.

Studies published in English that reported using AI/automation in PrEP interventions were included. We excluded studies that used AI/automation to improve HIV prevention in general and did not focus on PrEP. This review especially emphasized the use of AI/automation in behavioral interventions to enhance PrEP care.

A two-step approach was applied to select studies for review. First, two reviewers independently screened the citations by title and abstract. Second, two reviewers independently reviewed the full text of included citations to confirm study eligibility. Disagreements were resolved through discussion. Reviewers were trained, and all screening forms were pilot tested and revised, as necessary. Identified citations were exported to DistillerSR (a systematic review software developed by DistillerSR Inc, Ottawa, Canada) for data management and to screen studies and identify eligible studies.

For all eligible studies, two reviewers used a standard data abstraction form to extract data on population characteristics and AI/automation program-related information and outcomes. Study population characteristics were abstracted from the primary study’s participant eligibility criteria.

For studies with outcomes to evaluate interventions, risk of bias assessment was conducted by using either the revised Cochrane risk-of-bias tool for randomized trials (RoB 2) for randomized controlled trials (RCT)^18,19 or an adapted Newcastle-Ottawa Quality Assessment Scale (NOS) for non-randomized studies.^20-22 RoB 2 consisted of 28 questions with “yes,” “probably yes,” “probably no,” “no,” or “no information,” and NOS consisted of 5 (for cohort study) to 6 (for cross-sectional study) questions with “yes” or “no” responses. For RoB 2, reviewers assessed risk of bias (“low risk of bias,” “some concerns,” “high risk of bias”) based on the answers to the questions. For NOS, a total score of 5 or 6 was possible, with 3 or higher being considered as “low risk of bias” for cohort study and 4 or higher for cross-sectional study. We narratively synthesized characteristics and findings of the included studies.

Among 12 behavioral interventional, 4 types of AI/automation were identified: chatbot (AI-based or non-AI-based) (n=8),^23-30 ML (n=1),³¹ natural language processing (NLP; building machines that can understand human language and generate text and speech) (n=1),³² and other types of AI (n=2).^33,34

Eight of the 12 intervention studies did not report evaluation outcomes; they were either ongoing studies (i.e., registered in trial registry but have not yet published; n=3),^28-30 qualitative studies to assess perception/preference (n=2),^27,31 study protocols (n=2),^23,24 or an intervention description (n=1).³² Four were US-based studies,^24,27,28,31 three were non-US-based,^23,29,30 and one study’s location was unspecified.³² Six used chatbot,^23,24,27-30 one used NLP,³² and one used ML³¹ (Table 1).

HIV Self-Testing (HIVST)-chatbot was a fully automated HIVST service with HIV risk assessment provided by an AI chatbot for gay, bisexual, and other men who have sex with men (collectively referred to as MSM) in Hong Kong.²³ Their previous project, HIVST with online instruction and counseling (HIVST-OIC),³⁵ increased HIV testing and has been included in CDC’s PRS Evidence-Based Intervention (EBI) compendium.³⁶ However, HIVST-OIC had an implementation issue: it required intensive resources and capacity development (e.g., online real-time HIV testing instruction and counseling by a nurse, a 15-minute motivational interviewing phone call, and immediate online psychological support).^23,35 Thus, the proposed study would explore whether a fully automated HIVST-chatbot is as efficacious, if not more efficacious, as HIVST-OIC in increasing HIVST uptake and the proportion of HIVST users receiving counseling, including HIV risk assessment, and increasing PrEP use. The study was anticipated to start recruitment in April 2023 and enroll 528 Chinese-speaking MSM in Hong Kong (264 in HIVST-chatbot group, 264 in HIVST-OIC group).²³

Developing an AI chatbot intervention can be labor intensive and expensive, but PrEPBot showed that AI chatbot development does not need to be labor intensive nor expensive.²⁴ PrEPBot was a low-cost, short message service (SMS) text messaging-based chatbot. The chatbot was tailored to sexual and gender minority adolescents and young adults in Louisiana to support patient navigation and disseminate PrEP-related information as part of their TelePrEP program.²⁴ The intervention developer chose a commercial, readily available, easily deployable chatbot platform instead of an independently designed app. Additionally, PrEPBot used SMS text messaging with rule-based conversations that were scripted but had limited capacity for recognizing free text input. This study suggested the possibility of implementing an AI chatbot intervention with a limited budget by using a commercially available tool that can be programmed by researchers with no programming experience. Acceptability and usability of PrEPBot will be evaluated in their pilot trial.²⁴

Amith (2020) assessed functionality of a computer-based ontology-driven approach (an automated method to group responses and tailor follow-up questions) to manage conversations on PrEP and post-exposure prophylaxis (PEP).³² The dialogue could provide automated counseling and improve patient-provider communication by using NLP to develop the ontology-based method for handling dialogue and automating the communication of PrEP. The study showed that high functionality of automated conversations may provide real-time counseling and high availability to inform patients, but more studies are needed to explore further possibilities.

While the above 3 studies were about developing AI tools, Zhang (2022) and van den Berg (2021) assessed perceptions or preferences of using AI chatbots or ML among PrEP-eligible women²⁷ and primary care providers (PCPs).³¹ Almost a quarter of PrEP-eligible women participants (7 out of 29 women) in New York expressed enthusiasm about chatbots in PrEP care; however, about one-fifth of the women were strongly opposed to using chatbots.²⁷ A sample of 42 PCPs in Massachusetts reported high acceptance of using automated HIV risk prediction models to identify PrEP candidates. The models used ML algorithms to detect patterns in electronic health records (EHR) data indicative of risk for HIV acquisition.³¹ Moreover, PCPs stated that automated HIV risk prediction models might help patient-provider communication in general primary care settings by making HIV risk assessment more routine, reducing stigma, and empowering PCPs to prescribe PrEP instead of referring to HIV specialists.³¹ However, PCPs reported skepticism about using ML without knowing more about how the models worked, and some PCPs worried that patients might react negatively if they discovered that their HIV risk was predicted by using automated computer algorithms.³¹

The three ongoing chatbot studies are focused on MSM. Two studies (MyTestBot and TestBot) would be implemented in Malaysia,^29,30 and the other (Chatbot in Southern US) would focus on Black or African American (hereafter referred to as Black) MSM in the Southern US.²⁸ Both MyTestBot and TestBot promote HIV testing, and Chatbot in Southern US promotes PrEP awareness and uptake. While the two studies in Malaysia would either develop an AI chatbot or assess feasibility of the chatbot, Chatbot in Southern US would develop and pilot test to assess acceptability and feasibility of the chatbot (possibly not AI-driven).

Four intervention studies reported outcomes data that were used to evaluate the AI/automation interventions. Types of AI/automation used in these interventions were either chatbot (n=2)^25,26 or other AI (n=2).^33,34 Two studies evaluated the same intervention, Directory-observed therapy (DOT) Diary, thus this review actually looked at three, and not four, unique AI/automation interventions (Table 2). DOT Diary was US-based^33,34 and the other two interventions were non-US-based (Brazil²⁵ and Zambia²⁶). Only one of the DOT Diary³³ studies was RCT. A majority of the studies (n=3)^25,33,34 had low risk of bias.

Both Amanda Selfie²⁵ and Waiting-area Chatbot²⁶ were developed and tested for acceptability and usefulness. Amanda Selfie was a culturally tailored, AI chatbot tested among 1,288 adolescent MSM (AMSM) and adolescent transgender women (ATGW) aged 15-19 in Brazil.²⁵ The target population was chosen because of concerns about their behavioral changes (e.g., increased unprotected anal intercourse, reduced HIV risk perceptions) and high usage and values toward technologies including AI and chatbot. Participants were recruited from the demonstration cohort study which explored acceptability, use, and PrEP persistence.³⁷ Eligible participants interacted with Amanda Selfie, transgender chatbot using AI, who provided sex education and enabled them to link up to the PrEP clinics or other HIV testing and care services as needed. All interactions were monitored and assessed to analyze the consistency and accuracy of Amanda Selfie’s. The qualitative evaluation demonstrated that Amanda Selfie was highly usable, functionable, and acceptable, especially among ATGW. While some participants reported feeling comfortable talking to Amanda Selfie about their sexuality and emphasized that they felt safer and less exposed to judgment talking to a chatbot than to humans, others reported the importance of migrating to a dialog with a “real” person on topics that required greater depth (e.g., family problems, doubts about PrEP medications). Limitations reported by participants include inaccurate answers, moving too quickly or taking too long to read through all information, and using medical terms that were hard to understand.

The other chatbot intervention, Waiting-area Chatbot, was developed to provide information on dual protections against HIV and pregnancy while patients were in waiting areas in family planning (FP) clinics in Zambia.²⁶ This chatbot was web-based and the chat content was scripted as a series of closed-ended questions, so AI was not required. The non-AI chatbot was chosen because of its relative simplicity, geographic availability, cost, and available human resources. Because FP providers may not be able to spend enough time with each patient to address their concerns or patients may feel too uncomfortable to discuss HIV vulnerability, the Waiting-area Chatbot enabled women to engage in digital conversations about such topics as PrEP while waiting for their appointments.²⁶ High feasibility, acceptability, and positive effect on knowledge and provider interactions were reported among the sample of 30 women. Limitations of the study include providing a tablet with chatbot and mobile data for participants to use. Further studies are needed to determine if participants would use their own mobile devices and their own mobile data plan to proactively engage with the chat.

DOT Diary used AI to improve adherence by correctly capturing PrEP ingestion.^33,34 The DOT Diary app contained an automated DOT (aDOT) monitor and supported PrEP adherence by capturing data through the front-facing camera of the mobile device and analyzing the visual data using ML. High feasibility, acceptability, and usability of the app, as well as high PrEP adherence, were reported during the 8-week pilot among young, Black and Hispanic/Latino MSM in San Francisco and Atlanta.³⁴ However, a separate study of an RCT among 100 young MSM in these cities showed no significant difference in PrEP adherence between groups (i.e., the intervention [Dot Diary] and control [standard of care]), even though the accuracy of aDOT to measure PrEP ingestion was established.³³ In the pilot study, some participants reported that using the aDOT app was time consuming.³⁴ In the RCT, the number of PrEP pills recorded in the app showed a monotonic decline over time, but such a decline was not observed in tenofovir-diphosphate levels in dried blood spots of the participants. This discrepancy may suggest that the use of aDOT while taking PrEP decreases over time.