AI for precision medicine must keep non-random complexity in mind to support equity in outcomes

Large Models (LMs) as a new form of artificial intelligence (AI) have the potential to provide more personal insights by processing large volumes of multimodal, longitudinal data now generated by hundreds of millions of people through things like wearables, apps, and centralized systems such as electronic medical records. Medical research has historically excluded most populations (women and minorities) to the effect that treatments are routinely less effective and sometimes harmful in these populations. Machine Learning (ML) in medicine was supposed to solve these problems, but in many cases exacerbated these harmful historical biases. LMs project an illusion of understanding that might lead some to repeat the omissions of the past, and once again exacerbate unfair and harmful outcomes. But human physiology is both deeply complex and structured. Health AI researchers can incorporate insights from both biology and historical failures to help Health AIs realize the goal of equitable health support more completely than previous efforts.