An economic approach to machine learning in health policy

We consider the effects of 'precision' screening policies for cancer guided by algorithms. We first show that complex machine learning models can indeed predict cancer better than simpler models that use established risk factors. We then tackle the evaluation challenge: an algorithm that can predict cancer in a hold-out set only establishes predictability; it does not imply an algorithmic screening rule built on it would improve social welfare. Using a series of policy evaluation methods we show that targeting screening via algorithm could in fact lead to large health benefits. Moreover, we show the choice of prediction target is key - not all models with high accuracy can be used to construct beneficial screening policies.