Tensor factorization on temporal structured EHR for deep phenotyping
We applied a state-of-arts tensor factorization method to longitudinal EHR data and identify 14 clinically relevant subphenotypes for CVD. We found that some phenotypes such as Vitamin D deficiency and depression, and Urinary infections were not captured by the conventional risk scores, and the top six prevalent subphenotypes differ in the risk of developing the subsequent myocardial infarction.
