Five practical examples of how AI Is already shaping healthcare in South Africa

Detecting tuberculosis through lung sounds

The team at AI Diagnostics (https://www.aidiagnostics.health/) in Cape Town have developed an algorithm that analyses lung sounds to detect possible tuberculosis using a small handheld device. A healthcare worker records sounds from six standard chest positions, and the algorithm classifies them as normal or suspicious, indicating the need for confirmatory testing.

Its performance now approaches that of computer-assisted detection applied to digital chest radiography, while being far more portable and substantially less costly. Because it can be deployed in community settings where symptom screening is common but radiography is unavailable, it enables earlier detection and reduces barriers to care. This approach illustrates how AI can broaden diagnostic reach, with important subsequent curative and preventative benefits for the screened population.

‍

Modelling testing strategies for COVID-19 risk management

During the early pandemic response, Anglo American sought ways to reduce workplace transmission while maintaining business continuity. The central question was how frequently to test workers under evolving constraints: fluctuating community incidence, variable rapid test sensitivities (ranging from ~50–90%), differing coverage of naturally and vaccine-acquired immunity.

Our team at Wimmy (https://www.wimmy.com/) developed a scenario-modelling framework to simulate infection risk under alternative testing intervals, identifying an optimal balance between operational continuity and safety. Weekly testing with the highest-performance available antigen test consistently kept infections and hospitalisations below predefined acceptability thresholds, even during rising waves.

Beyond the quantitative result, the agent-based model provided a structured, transparent decision tool. It reduced conflict between management and labour by making assumptions explicit, enabling rapid scenario adjustments and grounding decisions in evidence rather than negotiation dynamics.

‍

Neural-network-assisted detection of noise-induced hearing loss

Noise-induced hearing loss remains a major occupational health burden in mining and heavy industry. Clinicians typically look for a characteristic notch in audiograms around 4–6 kHz, but interpretation varies by experience and context.

We trained a recurrent neural network on a large, expert-curated dataset. The algorithm achieved accuracy comparable to specialist judgement and outperformed simple rule-based methods. Importantly, the model introduced an “uncertain” category that routes borderline cases to specialist review rather than forcing binary classification.

This triage structure improves case detection sensitivity while using expert capacity efficiently. It illustrates a broader pattern: AI systems can scale expert decision-making while preserving specialist oversight where it is most impactful.

‍

Evaluating AI-enabled medical scribes

Primary care providers are increasingly experimenting with AI systems that record consultations and generate transcripts, SOAP notes, referral letters or billing codes. In a controlled evaluation of six simulated consultations, we assessed several commercial and open-source combinations across the speech-to-text and language-model pipeline. Given the multilingual South African society, we stress-tested the AI scribes by building in frequent switches between Afrikaans and English.

Transcription accuracy varied substantially: the best models had ~8% word-error rates, while the least accurate exceeded ~36%. However, raw transcription accuracy was not the most meaningful metric. Clinicians rely on summaries and structured notes rather than verbatim text.

Preliminary findings show that systems with moderate transcription error rates can still produce clinically usable summaries, though reliability varies by consultation type and complexity. Additionally, the ability to store audio and transcripts offers medico-legal advantages, enabling verification of whether critical information was communicated. Ongoing work now focuses on assessing the completeness, factual consistency and clinical safety of generated summaries, as these are the outcomes most directly relevant to clinical workflows.

‍

A conversational agent for adolescent sexual and reproductive health

Audere’s (https://www.auderenow.org/) AI-driven chatbot, AIMEE, supports adolescent girls and young women seeking confidential information on sexual and reproductive health. It addresses the wide range of questions that its users have, from menstruation and contraception to sexually transmitted infections, pregnancy scares and pre-exposure prophylaxis for HIV prevention. The system is proactive and may check in with users who previously expressed concern or emotional distress.

Over 5,000 users have engaged with the chatbot, with roughly 40% of conversations including an element of emotional or psychosocial support. Approximately 30% of conversations occur outside standard clinic hours, highlighting the value of an always-available, private channel. While not a clinical replacement, the system guides users toward appropriate services, offers reassurance in times of uncertainty and flags high-risk situations that require escalation to human counsellors. Building on the success of AIMEE, Audere, in partnership with the South African National Department of Health, has launched Self-Cav, an AI-powered health companion chatbot accessible via https://www.bwisehealth.com/. It provides confidential, judgment-free guidance on sexual health, mental health, and HIV prevention.

‍

Foundational considerations

Across these examples, AI extends clinical reach, supports operational decisions and enhances the availability and comprehensiveness of care. However, sustained benefit of AI systems in healthcare crucially depends on three foundational elements:

• Robust data infrastructure that enables secure, efficient data flow across systems and facilities.
  
  
• Strong governance and privacy safeguards, aligned with POPIA, GDPR and ethical best practice.
  
  
• High-quality data at source, without which even well-designed models will produce unreliable outputs.
  
  

Ensuring strong foundations, clear accountability frameworks and a focus on human benefit will determine whether these tools scale safely and equitably. What is very clear, though, is that AI’s trajectory in healthcare is accelerating. Our team at Wimmy has had the privilege to support several healthcare organisations in their pursuit of more data-informed, more accessible, more comprehensive and more affordable healthcare. And we are eager to support more organisations as they grapple with the testing, validation, integration and adoption of AI systems to the benefit of their healthcare staff and patient populations.

‍