Deep representation learning of electronic health records to unlock patient stratification at scale

• DOI: 10.1038/s41746-020-0301-z
• Publication Date: 2020-07-17T14:23:54Z
• Authors: Isotta Landi; Benjamin S. Glick...; Hao-Chih Lee; Sarah Cherng; Giulia Landi; Matteo Danieletto; Joel T. Dudley; Cesare Furlanello; Riccardo Miotto
• Keywords: Health records; Personalized Medicine; Electronic health record; Feature Learning

Deriving disease subtypes from electronic health records (EHRs) can guide next-generation personalized medicine. However, challenges in summarizing and representing patient data prevent widespread practice of scalable EHR-based stratification analysis. Here we present an unsupervised framework based on deep learning to process heterogeneous EHRs derive representations that efficiently effectively enable at scale. We considered 1,608,741 patients a diverse hospital cohort comprising total 57,464 clinical concepts. introduce representation model word embeddings, convolutional neural networks, autoencoders (i.e., ConvAE) transform trajectories into low-dimensional latent vectors. evaluated these as broadly enabling by applying hierarchical clustering different multi-disease disease-specific cohorts. ConvAE significantly outperformed several baselines task identify with complex conditions, 2.61 entropy 0.31 purity average scores. When applied stratify within certain condition, led various clinically relevant for disorders, including type 2 diabetes, Parkinson's disease, Alzheimer's largely related comorbidities, progression, symptom severity. With results, demonstrate generate lead meaningful insights. This help better understand varying etiologies sub-populations unlock patterns research the realm