Summary
The episode works through four questions drawn from the lesson on AI in public health practice. It uses the BlueDot alert and the record of Canada's Global Public Health Intelligence Network to show that automated epidemic intelligence depends on the people and organization around it, and it estimates the recall of a hypothetical news-monitoring filter from a sample of discarded items. With a hypothetical admission-risk model, it shows how the positive predictive value falls when a model moves to a population where the outcome is rarer while the AUC stays the same, and it links this to external validation and to proxy labels. The hosts then place studies of language models in health on the lesson's levels of evidence, debate whether a public health team should offer a generative chatbot to the public, and use Cohen's kappa to show how raw agreement can overstate an assistant's coding of a rare code.