Scalable and accurate deep learning with electronic health records

Abstract

Predictive modeling with electronic health record (EHR) data is anticipated to drive personalized medicine and improve healthcare quality. Constructing predictive statistical models typically requires extraction of curated predictor variables from normalized EHR data, a labor-intensive process that discards the vast majority of information in each patient’s record. We propose a representation of patients’ entire raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format. We demonstrate that deep learning methods using this representation are capable of accurately predicting multiple medical events from multiple centers without site-specific data harmonization. We validated our approach using de-identified EHR data from two US academic medical centers with 216,221 adult patients hospitalized for at least 24 h. In the sequential format we propose, this volume of EHR data unrolled into a total of 46,864,534,945 data points, including clinical notes. Deep learning models achieved high accuracy for tasks such as predicting: in-hospital mortality (area under the receiver operator curve [AUROC] across sites 0.93–0.94), 30-day unplanned readmission (AUROC 0.75–0.76), prolonged length of stay (AUROC 0.85–0.86), and all of a patient’s final discharge diagnoses (frequency-weighted AUROC 0.90). These models outperformed traditional, clinically-used predictive models in all cases. We believe that this approach can be used to create accurate and scalable predictions for a variety of clinical scenarios. In a case study of a particular prediction, we demonstrate that neural networks can be used to identify relevant information from the patient’s chart.

Scalable Machine Learning for Predicting At-Risk Profiles Upon Hospital Admission

“…Our prediction system is thus much more flexible and general, since it can be used with patients admitted for the first time with no known history. Second, [30] and [31] use deep neural network models whereas the current article presents results obtained with more classical models that: (i) perform faster on CPU machines, and (ii) are more intelligible: they offer more insights into the factors that influence the predictions, as emphasized in [32]. For instance, the weights computed by our models are stable and intelligible, allowing further interpretation (and fix if needed) by clinicians.…”

“…Closely related works in terms of research objectives are the independent works performed simultaneously found in Avati et al [30], Rajkomar et al [31]. Authors develop a model for predicting all-cause mortality of patients from EHR data.…”

“…The research objectives pursued are similar, and the methodology and the data used exhibit some similarities but with several important differences. First, [30] and [31] require "longitudinal" data in the sense that for making a prediction, their systems require historical data for the patient: it relies on historical observation windows and several observation slices: [30] uses 4 observation slices and [31] uses data from 24 hours before hospital admission (and also after admission). Both systems require the knowledge of the patient situation at several moments in time.…”

“…Last but not least, to be able to operate without the need for historical data of patients, our models were trained with much more "transversal" data. [30] studies a population of 221 284 patients (216 221 hospitalizations were studied in [31]), whereas our models are built from the automated analysis of more than 1.2 million of patients.…”

Models and Machines: How Deep Learning Will Take Clinical Pharmacology to the Next Level

“…Applications of ML in the context of pharmaceutical research typically showcase algorithms trained on huge data sets that identify correlations between features predictive of a given outcome . A few highlights from the drug discovery and development pipeline in which ML methods have added value are predictions of ligand‐protein binding from chemical properties, automatic classification of biopsy images, and prediction of medical events from electronic health records …”