Over the last year, it has been hard to ignore the hype around artificial intelligence (AI) and what the rapidly advancing technology could do to transform the health care sector. Commentators vary in predictions from AI being a valuable assistant for doctors, to the technology being poised to replace human professionals such as radiologists—a prediction made in 2018 that remains as of yet unrealized!
While my passion is to advance understanding of how to ethically and inclusively utilize applications of AI to improve human lives, I will also be the first to acknowledge that high-level political engagements and research can sometimes neglect the complexities of real-world health care settings. For this reason, as PATH has embarked on a new initiative in sub-Saharan Africa to explore appropriate applications of AI for primary health care settings, we have looked to our partner academic institutions, ministries of health, and technologists to help us more deeply understand the key challenges to effectively and ethically applying AI tools in their countries.
Based on current research and my engagement with diverse stakeholders, I offer this list of five things for the global health community to keep in mind as it explores the potential to integrate AI-enabled applications into public health programs and health service delivery on the African continent.
1. Localization will be critical to the effective and ethical use of AI in African health care settings. Many of the AI tools we are familiar with today, such as ChatGPT and Google’s Gemini, are large language models (LLMs). They are trained using huge amounts of data to be able to provide intelligent and creative responses to prompts, like a human. Given that these models pick up information from the environment where they have been primarily used and trained, you can’t just assume they will work in any context. Existing LLMs developed to give medical advice to individuals or provide diagnostic assistance to physicians are trained using data sources that reflect and bake in the biases of WEIRD (Western, educated, industrialized, rich, and democratic) societies. For LLMs to be useful for community health workers in African countries, they need to be trained based on an understanding of local medical practices and vernacular.
Health care service provision can vary profoundly from country to country due to differences in burdens of disease, cultures, and local enabling environments. To think we can drag and drop a solution from a high-income setting into a low-resource environment is at best naive, and at worst negligent. To get the best results from AI, we have to think about developing LLMs for the many, not the (privileged) few, and start with intentional engagement of end users on day one of the product development cycle.
2. African markets and governments need to invest in their digital (enabling) infrastructure if they are going to benefit from AI. We are reaching a breakthrough moment for AI as the practical barriers to implementation rapidly get broken down. For example, advancements in physical infrastructure, such as solar energy and satellite internet access, are making it more feasible to digitally enable health facilities at the last mile. However, there are many issues around digital infrastructure which we have yet to adequately address, and which continue to undermine the path to scale. For example, the data capture mechanisms required to facilitate reimbursement for AI-enabled tools—a vital incentive to procure a specific technology in insurance-based health care systems—are effectively nonexistent in many countries; they are weak at best in many high-income countries as well. AI-solution vendors instead must create bespoke mechanisms for getting paid, and thus, unsurprisingly, the result is a fragmented market of AI tools which struggles to thrive because the costs of scaling are prohibitively high.
We need government officials and policymakers to understand that there are a series of enabling investments that need to be made to support a well-functioning public and private marketplace for AI-based solutions. Without them, the local entrepreneurial ecosystem will fail to thrive, and local health care systems will have access to limited options.
“To get the best results from AI, we have to think about developing LLMs for the many, not the (privileged) few, and start with intentional engagement of end users on day one of the product development cycle.”
3. Oversight is critical, and without appropriate safeguards, we will do more harm than good. While AI-enabled tools can be useful assistants for health care workers, we are not yet ready for autonomous applications for diagnosis and treatment planning. The jury is still out on whether LLMs are superior to physicians in clinical decision-making, so we need to support regulators and other critical policymaking stakeholders to understand how LLMs work and how to safely and ethically oversee them. More importantly, we need to ensure that as countries rapidly develop AI policies and other safeguards that they share their successes and failures openly and create a culture of shared learning. This is something that PATH has been incredibly vocal about at the G20, and the latest draft of the ministerial statement (to be published in October 2024) suggests that key stakeholders might be listening!
Without the trust and goodwill of our patients, it's impossible to ethically deploy any novel or innovative technology. Developing effective safeguards is one of the few ways we can build and maintain that trust, and thus, taking time to create robust regulation isn’t an impediment to innovation, it’s a critical part of enabling it.
4. AI tools require significant resources, and we need to consider the trade-offs. When it comes to determining if an AI-enabled tool is an appropriate solution, we need to look not just at the accuracy and effectiveness of the tool, but also at cost-effectiveness. In any given context, the value generated by the AI tool might not be the most efficient use of health care funding and resources. Unfortunately, only a handful of African randomized controlled trials of an AI-enabled diagnostic tool have been published, and the one that included an assessment of cost-effectiveness concluded that computer-aided diagnosis of TB on chest X-rays was not cost-effective (based on mortality-related benefits).
When evaluating a use case for an AI-enabled tool, we need to consider the opportunity cost of implementing that solution. We can’t be driven by a desire to digitize for the sake of it; rather, we need to stay singularly focused on the people we’re trying to help, and where there is value in AI, the onus is on us technologists to prove it.
“Developing effective safeguards is one of the few ways we can build and maintain trust, and thus, taking time to create robust regulation isn’t an impediment to innovation, it’s a critical part of enabling it.”
5. African innovators can and will lead the way in combatting AI bias and developing applications of AI that improve lives. Around the world there is a large youth population that is hungry to do something to improve their countries and communities. Our partners in countries such as Rwanda have brilliant teams of technologists who know the limitations of existing LLMs and are doing something about it. For example, Digital Umuganda is an organization dedicated to creating AI models and tools that make it possible for LLMs to work for marginalized communities that speak local African languages. They have been successful in creating voice and text datasets in Kinyarwanda and a translation module that makes it possible for people to ask ChatGPT questions in Kinyarwanda and receive back intelligent answers. The translations are still not perfect, but they have made incredible progress in making this game-changing technology accessible to all Rwandans.
There already is a thriving AI for Health entrepreneurial ecosystem on the African continent. At PATH our role isn’t to compete, but rather to find ways to platform and accelerate the impact of those innovators. I’m hoping that in a few years they’ll have put me out of a job, or at least that’s my marker of success.
