Scalable and Accurate Deep Learning with Electronic Health Records
Comprehensive Overview
Published in npj Digital Medicine in 2018, this study asked whether deep learning could accurately predict a wide range of important clinical outcomes directly from a patient's entire electronic health record, including free-text notes, without the labor-intensive process of hand-selecting a small set of predictor variables the way traditional models require. Researchers from Google, UCSF, and the University of Chicago Medicine developed a method that represents a patient's complete raw record, structured data and clinical notes alike, as a single chronological sequence using the FHIR data standard, then trained neural networks on de-identified records from 216,221 hospitalizations involving 114,003 adult patients across two academic medical centers. The resulting dataset totaled nearly 47 billion individual data points. The models were trained to predict four distinct outcomes: in-hospital mortality, unplanned 30-day readmission, prolonged hospital stays of a week or longer, and a patient's complete set of discharge diagnoses, each compared directly against traditional, currently used clinical prediction scores. The topline result was that these deep learning models achieved substantially higher accuracy than traditional models on every single prediction task, at both hospitals, without requiring any manual selection of which data to consider.
Outcomes and Implications
The single most important finding was that the deep learning models substantially outperformed traditional clinical prediction tools across all four outcomes: for in-hospital mortality predicted 24 hours after admission, accuracy reached 0.93 to 0.95 versus 0.85 to 0.86 for the standard Early Warning Score, achieving that same accuracy 24 to 48 hours earlier while roughly halving the rate of false alarms clinicians would need to investigate. For 30-day unplanned readmission, accuracy was 0.76 to 0.77 versus 0.68 to 0.70 for the standard HOSPITAL score, and for prolonged length of stay, 0.85 to 0.86 versus 0.74 to 0.76 for a comparable traditional model. The models also predicted a patient's entire set of discharge diagnoses, drawn from over 14,000 possible codes, with 0.90 accuracy at discharge. In a case study addressing interpretability, a central concern with deep learning in medicine, the model highlighted exactly which notes, medications, and nursing observations it relied upon for an individual mortality prediction, correctly flagging evidence of a malignant pleural effusion and infection a clinician would also recognize as significant. Because the approach uses the entire raw record rather than curated variables, it could in principle scale to new hospitals without the months of manual data-harmonization traditional modeling requires. The authors conclude that deep learning can produce more accurate, earlier, and more scalable clinical predictions than existing tools, though this was a retrospective study, and prospective trials are still needed to confirm these predictions actually improve care when acted upon.