Event Report: AI for African Healthcare: Beyond Chatbots Toward Contextually Informed InnovationEvent reportEvent Report: AI for African Healthcare: Beyond Chatbots Toward Contextually Informed Innovation

Event Report: AI for African Healthcare: Beyond Chatbots Toward Contextually Informed Innovation

On 11th July 2025, doctoral candidate Joseph Mugauri attended a seminar presented by Professor Judy Gichoya (Emory University) who is visiting the University of Pretoria and hosted by Data Science for Social Impact. Joseph penned a report of that insightful seminar.

Event Report: AI for African Healthcare: Beyond Chatbots Toward Contextually Informed Innovation

by Joseph JTR Mugauri

The seminar “AI for African Healthcare: Beyond Chatbots Toward Contextually Informed Innovation” was delivered by Professor Judy Gichoya on 11 July 2025 at the University of Pretoria. In her talk, Professor Gichoya explored the broad landscape of AI in healthcare and honed in on what this means for Africa. She acknowledged that African healthcare faces limited data, constrained computing infrastructure, and diverse local health challenges. Yet instead of simply lamenting these constraints, she highlighted a silver lining carrying the possibility of creating innovative AI tools that are custom-built for African contexts rather than directly transplanting solutions from elsewhere. This idea that Africa’s unique needs can inspire bespoke technological advancements sets a thoughtful tone for the following discussion.

Professor Gichoya observed that the global market for AI technologies in healthcare is growing, but cautioned that widespread success is not guaranteed. To illustrate her points, she turned to radiology as a case study. Radiology is one of the fields often heralded as ripe for AI breakthroughs. Imaging generates lots of data, and pattern recognition seems a natural fit for machine learning. Gichoya’s approach was to walk through the basics of how AI is supposed to work in radiology, examine where these systems have failed, and then consider how AI might still be leveraged meaningfully to improve health outcomes. The narrative is straightforward; understanding AI’s limitations is as vital as celebrating its promises.

To ground these ideas, Gichoya briefly sketched how deep learning works. She explained that any AI algorithm needs labelled examples to learn from. In medical imaging, experts must annotate countless images (for example, marking which chest X-rays show a condition like pneumonia). The algorithm then sifts through this labelled data to detect patterns. If the training data is consistent and high-quality, predictions can be impressive. However, the process is brittle if the labels are off, incomplete, or the dataset is not representative, because the AI’s predictions will suffer. In short, deep learning is powerful but also fragile. This underscores a subtle lesson that building an AI model is not like flipping a switch; it demands painstaking preparation and ongoing validation.

Diving deeper into radiology, the presentation explained how AI typically attempts to identify diseases in images. In essence, many models use either classification or segmentation techniques. For instance, Gichoya described a chest X-ray scenario where the AI is tasked with diagnosing pneumothorax (a collapsed lung). The most straightforward approach is a yes-or-no classification where the model reviews the image and answers whether pneumothorax is present. In other cases, the model might produce a bounding box around the area it thinks is affected. Complete segmentation is even more advanced; the AI outlines the exact region where the disease appears. Each method has its uses and limitations. For example, a yes/no answer is fast but tells the radiologist little about where to look, whereas segmentation can provide visual evidence of the model’s guess. Through this example, Gichoya subtly pointed out that AI tools should assist, not replace, human experts. The radiologist still guides the interpretation, using the AI’s output as one piece of information among many.

Still on the fear that AI might replace radiologists, and why that fear may be unfounded, Gichoya emphasised that AI training hinges on pattern recognition. AI learns to associate visual patterns with diagnoses, but those patterns are not always reliable proxies for disease. In practice, a model may spot a pattern that correlates with an illness in a training dataset, yet that pattern might not appear in perfectly healthy scans, or vice versa. This mismatch can cause the AI to miss the disease or raise false alarms. Gichoya noted that “technology is not the solution for problems that existed a long time ago,” recalling that issues like the shortage of doctors are complex. Even if people hope AI will magically fill these gaps, AI alone cannot overcome fundamental systemic problems in healthcare staffing. This insight invites a measured perspective, rather than thinking of AI as a quick fix, it should be viewed as one piece in a larger puzzle.

The talk also ventured beyond traditional imaging AI into the realm of generative AI (GenAI) in medicine. Professor Gichoya shared a vivid example involving Google’s Gemini. In her scenario, a doctor analyses a CT scan for signs of pancreatitis. Using an interactive setup, the physician hovers a cursor over areas of interest on the image and queries Gemini in real time: “Is there a sign of pancreatitis here?” The GenAI tool responds immediately. It’s a compelling vision of the future with an intelligent assistant that can highlight regions of concern and explain its reasoning on the fly. Gichoya’s example illustrates how even new forms of AI beyond simple pattern recognition are creeping into clinical workflows. Yet, she presented this with a hint of caution because technology that seems almost magical is still under human guidance. The doctor is in the driver’s seat, asking questions and interpreting answers. This interactive use of AI is intriguing, but like all the tools discussed, it must be validated carefully.

Throughout the discussion, a recurring question was: What unique factors must we consider when deploying AI in African healthcare? Gichoya first pointed to one of the most glaring issues: the lack of African data. Virtually all leading AI models in radiology have been trained on datasets from North America, Europe, or parts of Asia. If an algorithm learned to spot tuberculosis patterns using mostly Asian patient scans, it might miss atypical presentations common in African populations. She reminded us that AI’s accuracy critically depends on the relevance of its training data. If a model never sees examples of how a particular disease manifests in African patients, it’s likely to perform poorly on that population. This raises practical alarms since deploying an untested model could mean missed diagnoses or misdiagnoses. By highlighting this, she encouraged the audience to think about data diversity. It’s not just a technical issue but an ethical one; equitable healthcare AI demands datasets that reflect the people it serves. This point also suggests an exciting opportunity for African institutions to lead the way in assembling such datasets, thus steering the innovation in a direction that benefits their communities.

Another challenging topic Professor Gichoya tackled was AI “hallucinations,” especially in the context of generative models. She explained that a hallucination occurs when an AI confidently produces false or non-existent information. For example, she recounted asking a popular language model (ChatGPT) to list her five most-cited papers. The model complied instantly, but all five were fabricated; it had “invented” paper titles that didn’t exist. Similarly, she mentioned giving an AI an X-ray image and asking it to identify underlying conditions. The AI’s response included the presence of breast implants on a scan that had no implants at all. These stories are both amusing and alarming. They show that when an AI’s design goal is to generate plausible-sounding output, it doesn’t know or care whether what it says is true. The model was filling in the blanks. She used these examples to remind us of a key lesson that AI outputs must constantly be scrutinised. We cannot fully trust AI systems, especially generative ones, without cross-checking their claims against reality and expert knowledge.

Finally, Professor Gichoya did not shy away from the broader issues of privacy and intellectual property. AI models, after all, rely on vast data, which is very sensitive. She cautioned about what happens if a model trained on medical records or research inadvertently leaks confidential information. If the model regurgitates exact phrases or patient details it saw during training, that could violate privacy laws or ethical norms. Relatedly, there is a risk of infringing on intellectual property because AI systems might absorb and reproduce copyright-protected articles. These concerns loom large in an African context where data governance and legal frameworks are still developing. Gichoya’s reflections here serve as a reminder that technical innovation doesn’t exist in a vacuum. Deploying AI in healthcare also means grappling with policy, ethics, and public trust.

Conclusively, Professor Gichoya’s presentation was not just a run-through of technical points. It was a careful exploration of how AI’s promise in healthcare is thrilling and fraught with pitfalls. By the end, several themes stood out. First, AI is not a cure-all; it works in tandem with human expertise, not as a substitute. Second, context matters greatly, whether it is the patient population or the social environment. Third, constant vigilance is needed to ensure these systems are accurate, fair, and safe. For anyone interested in the future of health technology in Africa, Gichoya’s insights are a powerful reminder that the path forward should be innovative, yes, but also thoughtful and grounded in local realities.

~Written by Joseph JTR Mugauri