A patient’s health state can be characterized by a multitude of signals from many different data modalities. This high-dimensional, personalized data stream aggregated over patients’ lives has spurred interest in developing new clinical AI models. One of the rate-limiting factors in developing AI models that generalize to real-world scenarios is the very attribute that makes the data exciting—their high-dimensional nature.
At DiMe’s #AskMeAnything Journal, author Visar Berisha, PhD led a discussion on how “the curse of dimensionality” can doom models to failure, even when they seem to work well during development. We explored the key highlights of his Nature publication “Digital medicine and the curse of dimensionality“, Visar also provided some suggestions on how to develop clinical AI models that are more likely to fare well during prospective validation.
