Augmented Curation of Unstructured Clinical Notes from a Massive EHR\n System Reveals Specific Phenotypic Signature of Impending COVID-19 Diagnosis

Understanding the temporal dynamics of COVID-19 patient phenotypes is\nnecessary to derive fine-grained resolution of pathophysiology. Here we use\nstate-of-the-art deep neural networks over an institution-wide machine\nintelligence platform for the augmented curation of 15.8 million clinical notes\nfrom 30,494 patients subjected to COVID-19 PCR diagnostic testing. By\ncontrasting the Electronic Health Record (EHR)-derived clinical phenotypes of\nCOVID-19-positive (COVIDpos, n=635) versus COVID-19-negative (COVIDneg,\nn=29,859) patients over each day of the week preceding the PCR testing date, we\nidentify anosmia/dysgeusia (37.4-fold), myalgia/arthralgia (2.6-fold), diarrhea\n(2.2-fold), fever/chills (2.1-fold), respiratory difficulty (1.9-fold), and\ncough (1.8-fold) as significantly amplified in COVIDpos over COVIDneg patients.\nThe specific combination of cough and diarrhea has a 3.2-fold amplification in\nCOVIDpos patients during the week prior to PCR testing, and along with\nanosmia/dysgeusia, constitutes the earliest EHR-derived signature of COVID-19\n(4-7 days prior to typical PCR testing date). This study introduces an\nAugmented Intelligence platform for the real-time synthesis of institutional\nknowledge captured in EHRs. The platform holds tremendous potential for scaling\nup curation throughput, with minimal need for retraining underlying neural\nnetworks, thus promising EHR-powered early diagnosis for a broad spectrum of\ndiseases.\n