Best Practices for Validating Machine Learning in Medicine
Friday, September 18, 2020 — 2:00PM - 3:15PM
Many students build classifiers and perform regressions in data-driven courses including machine learning, data science, and applied statistics. However, even for more advanced students there are particular mistakes made when applying those predictive modeling skills in health care settings in which data can be scarce and uncertain with significant consequences for errors. In this panel, we explore those issues with a range of perspectives, seeking practical advice for computer scientists along with illustrative cautionary tales. Despite all that artificial intelligence has accomplished there is a considerable degree of skepticism among clinicians about the real-world applicability of AI in medical contexts. We address a variety of techniques that can remedy this by using proper validation strategies - some clinically oriented during data collection and a few computational approaches. This is particularly important as many complex models may be less interpretable in how they function (for example, ensemble methods or deep learning), but can be useful to sift through large data sets as recommendation systems for clinical decision making. With healthcare providers exposed to more properly trained and validated models machine learning strategies will be easier to adopt and guide clinical decisions in practice.
Riyad Bin Rafiq, Mark V. Albert, Megan O’Brien, Sarah Moudy