In the world of big data, selecting the “best” subset of data to learn from can be tricky and challenging, especially when data collection processes may pose subsequent biases in part of the observations. It is certainly the case in medicine, where the data represents a patient’s medical history. Choosing your data correctly is key to improve the performance of even a baseline vanilla model, and define the initial target audience of your data science product.

We will review methodologies of choosing the inclusion/exclusion criteria (a standard craft in the design of clinical studies), assessing their impact and implications, and how to deal with the selection bias that can be induced by the selection process.