Designing AI algorithms to suit local context

Charles B. Delahunt, Yunusa G. Muhammed, Tarisiro Matiza

Abstract: Artificial Intelligence holds tremendous promise to transform healthcare, and if we can effectively convert technical advances in AI into robust deployments, then we can vastly improve quality of life for billions of currently underserved people. But a substantial barrier to this opportunity is a disconnect between AI developers and clinical realities: To deploy successfully, AI development must be shaped at every stage by the specifics of the local deployment context. This integration is complex, it’s a necessary condition of success, and it falls to AI teams to carry out. However, because this task sits outside the traditional algorithm-centric focus of AI, it is poorly understood. Therefore this paper describes little-discussed but crucial aspects of an algorithm’s path to deployment, each directly relevant to AI researchers and illustrated by concrete examples drawn from experiences in the African healthcare context. We describe how AI is just one part of a much larger healthcare context, and how a core task of the AI team is to understand and translate this context into AI design choices from the very start of a project. We also include a set of actionable steps accessible to any researcher, which can markedly improve the fitness of an algorithm for future deployment.