Self-Care from Anywhere: Evaluating the usability of an AI-powered HIV toolkit among adolescent girls and young women and healthcare providers in South Africa

An exploratory study in South Africa found that the AI-powered Self-Care from Anywhere toolkit demonstrated high usability and acceptability among adolescent girls and young women and healthcare providers, offering a promising, user-centered approach to overcoming barriers in HIV prevention, testing, and linkage to care.

The Gist

Imagine you are trying to get a health check-up, but you are shy, worried about being judged, or just too busy to visit a clinic. Now, imagine having a super-smart, non-judgmental digital friend on your phone who can answer your questions, guide you through a home test, and even talk to your doctor for you.

That is essentially what this research paper is about. It tested a new AI-powered tool called "Self-Care From Anywhere" (SCFA) in South Africa to see if it works for young women and the doctors who treat them.

Here is the story of the study, broken down into simple parts:

1. The Problem: The "Wall" Between Patients and Care

In South Africa, HIV is still a big challenge. Many young women (ages 16–24) want to get tested or get help, but they hit a "wall."

• The Wall of Shame: They are afraid of being judged by nurses or friends.
• The Wall of Distance: Clinics are far away, and they don't have time or money to travel.
• The Wall of Busyness: Doctors are overwhelmed, like chefs in a kitchen with too many orders, so they can't spend enough time with each patient.

2. The Solution: A Digital "Swiss Army Knife"

The researchers built a toolkit with two main parts, like a two-sided coin:

• Side A: The AI Companion (For the Young Women)
  Think of this as a 24/7 digital best friend that lives on WhatsApp. It's an AI chatbot that talks to young women. It doesn't get tired, it doesn't judge, and it keeps secrets. It can:

  • Answer questions about sex and health.
  • Guide them through doing an HIV self-test at home (like a video game tutorial).
  • Give them emotional support if the news is bad.
  • Connect them to a real clinic if they need one.

• Side B: The Clinical Portal (For the Doctors)
  This is a dashboard for the doctors. Imagine a control panel that shows the doctor a summary of what the young woman talked to the AI about. It's like a "pre-game briefing" for the doctor. It helps them know what the patient is worried about before they even meet, so they can be more efficient and kind.

3. The Experiment: Putting It to the Test

The researchers gathered 97 young women and 44 healthcare workers in Gauteng, South Africa.

• The Young Women chatted with the AI. Some just talked; others used the AI to guide them through a fake (simulated) HIV test at home.
• The Doctors looked at the "dashboard" (Clinical Portal) to see how easy it was to read the summaries and manage patients.

They asked everyone: "Was this easy to use? Did you like it? Did it feel helpful?"

4. The Results: A Standing Ovation (With a Few Hiccups)

For the Young Women:

• The Verdict: They loved it! The tool got a score of 87.7 out of 100 for usability (anything above 68 is considered "good," and above 80 is "excellent").
• Why? They felt safe. One participant said, "With the AI, it's like talking to a friend on the phone. If I told a human friend, they might judge me, but the AI just gives advice."
• The Magic: It felt personal. The AI remembered their name (or a fake name they chose) and spoke in a way that made them feel understood.

For the Doctors:

• The Verdict: They thought it was a great idea, but they had some worries.
• The Good: They liked having a "cheat sheet" of the patient's history. It felt like having a map before entering a maze.
• The Bad: About 23% of the doctors were worried. They thought, "This is great, but will it just add more work to my already crazy schedule?" They were also worried about keeping patient identities safe so they don't lose track of who is who.

5. The Big Takeaway: A Bridge, Not a Replacement

The study found that this AI tool is a bridge, not a replacement for human doctors.

• It builds a bridge of trust for young women who are too scared to walk into a clinic.
• It builds a bridge of information for doctors who are too busy to ask every single question.

The Catch:
There is a little tug-of-war. Young women want total anonymity (no one knowing who they are), but doctors need clear identification (knowing exactly who the patient is) to give good care. The researchers realized they need to find a way to balance "being a secret agent" with "being a real patient."

In a Nutshell

This paper says: AI can be a superhero for healthcare in places where resources are tight. It gives young women a safe, private space to learn and test themselves, and it gives doctors a superpower to understand their patients faster. But, to make it perfect, we need to keep tweaking the tool so it fits perfectly into the real, messy lives of both the patients and the doctors.

It's not about replacing the human touch; it's about using technology to make that human touch happen more often and more effectively.

Technical Summary

1. Problem Statement

Despite the availability of HIV testing and treatment services in South Africa, significant gaps persist in early diagnosis and linkage to care, driving onward transmission.

• Target Population Barriers: Adolescent Girls and Young Women (AGYW) face specific obstacles including stigma, privacy concerns, fear of discrimination, and logistical issues (travel distance, clinic hours).
• Provider Barriers: Healthcare providers operate in resource-constrained environments with high client volumes and staff shortages, limiting time for counseling and follow-up.
• Limitations of Current Digital Tools: Existing digital health solutions often rely on generic scripts, lack emotional resonance, and fail to integrate counseling, self-testing guidance, and clinical linkage into a cohesive workflow.
• The Gap: There is a need for AI-enabled tools that can complement human-delivered services, offering confidential, adaptive, and accessible support without replacing in-person care, specifically tailored to the unique realities of AGYW in South Africa.

2. Methodology

The study employed an exploratory mixed-methods design conducted in Gauteng Province, South Africa, between November 2024 and May 2025.

• Intervention: The Self-Care From Anywhere (SCFA) toolkit, developed by Audere, comprising two components:

  1. AI Companion: A WhatsApp-based conversational agent for clients (AGYW) providing health guidance, HIV self-testing (HIVST) instructions, empathetic pre/post-test counseling, and a private channel for provider engagement.
  2. Clinical Portal: A provider-facing dashboard featuring structured summaries of client-AI interactions, computer vision-enabled interpretation of HIV self-test results, and AI-generated predictive HIV vulnerability scores for case management.

• Study Population:

  • AGYW (n=97): Aged 16–24 years. Participants were assigned to Group A (AI Companion only) or Group B (AI Companion + simulated HIVST).
  • Healthcare Providers (n=44): From public and private sectors, primarily HIV testing and screening (HTS) counselors and nurses.

• Procedures:

  • AGYW: Completed baseline surveys (demographics, intention to engage in care), interacted with the AI Companion, and (for Group B) performed a simulated HIV self-test with AI-guided counseling. Post-intervention surveys included the System Usability Scale (SUS) and intention-to-care measures. Focus Group Discussions (FGDs) were conducted with a subset.
  • Providers: Watched introductory videos, managed simulated client cases in the Clinical Portal without formal training (to capture first impressions), completed usability surveys, and participated in FGDs/In-Depth Interviews (IDIs).

• Data Analysis:

  • Quantitative: Descriptive statistics (means, SDs, percentages) using STATA v18. SUS scores were analyzed by subgroup. Intention-to-care changes were assessed via confidence interval plots.
  • Qualitative: Thematic analysis using NVivo v12, guided by the User-Centred Design (UCD) framework.

3. Key Contributions

• Novel AI Integration: Demonstrates a dual-sided AI ecosystem where a client-facing LLM (Large Language Model) companion feeds structured data into a provider-facing clinical portal, bridging the gap between self-care and clinical management.
• Contextual Adaptation: The study specifically addresses the South African context, focusing on the high HIV vulnerability of AGYW and the resource constraints of local healthcare providers.
• Usability Benchmarking: Provides empirical usability data (SUS scores) for an AI health tool in a low-to-middle-income country (LMIC) setting, establishing a baseline for future iterations.
• Identification of Tensions: Highlights the critical design tension between client preferences for anonymity (AGYW) and provider needs for accurate client identification and continuity of care.

4. Results

Quantitative Findings

• AGYW Usability: The AI Companion achieved a mean SUS score of 87.7 (SD 12.7), classified as "excellent" (scores >80). Usability remained high across all subgroups (age, education, HIV status, PrEP use), though slightly lower scores were noted among those with unknown HIV status or lower education levels.
• Intention to Engage: Post-intervention, mean scores for "Intention to Engage in HIV Care" increased across most dimensions, though confidence intervals overlapped, indicating no statistically significant change in this small sample.
• Provider Usability: Providers rated the Clinical Portal's ease of use, comprehension, and client support as "positive" to "very positive." However, 23% expressed concerns regarding workflow efficiency and the ability to manage increased client volume.

Qualitative Findings

• AGYW Perspectives:
  • Benefits: Valued the non-judgmental, confidential, and empathetic tone of the AI. It allowed for open discussion of sensitive topics without fear of stigma or "bad attitudes" from human staff.
  • Features: Privacy (pseudonyms), personalization (language adaptation), and clarity of information were key drivers of acceptance.
  • Limitations: Some participants expressed distrust in AI's ability to handle crisis situations (e.g., suicidal ideation) compared to human counselors.
• Provider Perspectives:
  • Benefits: The portal was seen as a valuable tool for pre-consultation preparation, offering summaries of client chats and risk scores to prioritize cases.
  • Challenges: Concerns arose regarding client anonymity vs. record linkage (risk of duplicate profiles). Interface navigation issues were noted (e.g., locating chat panels, interpreting risk scores). Providers noted that while the tool is promising, it requires formal training and workflow integration to be effective in daily practice.

5. Significance and Conclusion

• Feasibility and Acceptability: The study confirms that AI-powered digital health tools are highly usable and acceptable among AGYW and healthcare providers in South Africa. The SCFA toolkit successfully demonstrated potential to enhance user engagement and support care delivery.
• User-Centered Refinement: The findings underscore the necessity of iterative, user-centered design. The toolkit is currently being refined to include more content on treatment adherence and mental health, and to address the specific needs of HIV-positive populations.
• Strategic Tension: A critical insight is the conflict between the AGYW desire for anonymity and the health system's need for identification to ensure continuity of care. Future implementations must navigate this balance to ensure safe, effective, and equitable AI-driven self-care.
• Future Directions: The authors recommend further large-scale implementation studies (e.g., the "Aimee" study) and health economic analyses (cost-effectiveness) to validate the tool's impact on clinical outcomes and its viability for scale-up in public health systems.

Limitations: The study was exploratory with a small sample size, relied on self-reported data, and was limited to English-speaking urban participants, which may affect generalizability to rural or non-English speaking populations. Additionally, the lack of formal training for providers may have skewed usability perceptions.