ImageNet classification with deep convolutional neural networks
Alex Krizhevsky, Ilya Sutskever et al.
2.8k
Citations
120
Influential Citations
npj Digital Medicine
Venue
2018
Year
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.
This paper is a landmark in applying deep learning to electronic health records (EHRs) because it tackles two major barriers: the labor-intensive process of feature engineering and the difficulty of harmonizing data across different hospital systems. By using the Fast Healthcare Interoperability Resources (FHIR) format, the authors show that raw, sequential patient data can be fed directly into neural networks, preserving the richness of the entire record. This approach is scalable and generalizable, as validated across two independent academic medical centers without site-specific tuning.
The significance lies in its potential to democratize predictive modeling in healthcare. Traditional models require curated variables and often fail to transfer between institutions. This work suggests that deep learning can learn directly from the raw data, capturing subtle patterns that might be lost in manual feature extraction. The high accuracy for tasks like in-hospital mortality (AUROC 0.93-0.94) and discharge diagnoses (AUROC 0.90) demonstrates clinical utility, while outperforming existing models sets a new benchmark.
This paper has had a major impact on the AI in healthcare field, with over 2,800 citations. It demonstrated that deep learning can work directly on raw EHR data, bypassing the need for manual feature engineering and site-specific harmonization. This has inspired subsequent work on foundation models for healthcare, such as BEHRT and Med-BERT, and has influenced the design of clinical decision support systems. The approach is now a standard baseline for EHR-based predictive modeling, and its success has accelerated the adoption of AI in hospital systems.
Alex Krizhevsky, Ilya Sutskever et al.
Ashish Vaswani, Noam Shazeer et al.
Douglas M. Bates, Martin Mächler et al.
Diederik P. Kingma, Jimmy Ba