Skip to content
PaperFren

Can deep learning use a whole hospital record to predict outcomes?

Open paper intelligence

Neural networks reading a patient's entire raw hospital record predicted death, readmission and long stays more accurately than standard clinical scores.

Source

Scalable and accurate deep learning with electronic health records

Rajkomar A, Oren E, Chen K, et al. · NPJ digital medicine · 2018

doi.org/10.1038/s41746-018-0029-1Read the full paper ↗1,369 citationscc by

Study at a glance

Design
Computational / modelling — Retrospective development and test of an ensemble of LSTM, attention and boosted-stump networks on raw FHIR-formatted records, compared with logistic-regression baselines, with an 80/10/10 patient split.
N
N=216221 · 216,221 hospitalizations from 114,003 unique adult patients across two US academic hospitals.
Population
Adult inpatients (stays of at least 24 hours) at two US academic medical centres
Outcome
AUROC for in-hospital mortality, 30-day unplanned readmission, long length of stay, and discharge diagnosis codes

Structured fields used in claim comparison tables when every cited study has a complete layer.

What they did

The team converted all electronic health record data from two hospitals, including doctors' and nurses' notes at one site, into a single time-ordered sequence per patient without hand-picking variables. They trained three kinds of deep network and combined them, then predicted in-hospital death, unplanned readmission within 30 days, stays of a week or more, and the full set of discharge diagnoses. Results on a held-out test set were compared with logistic-regression models built from established clinical scores.

What they found

For in-hospital death predicted 24 hours after admission, the model reached an AUROC of 0.95 and 0.93 at the two hospitals versus 0.85 and 0.86 for the baseline score, roughly halving false alerts. Readmission (0.77 and 0.76 vs 0.70 and 0.68) and long stay (0.86 and 0.85 vs 0.76 and 0.74) were also better predicted. Diagnosis prediction reached a weighted AUROC of 0.90 at discharge, though the harder micro-F1 score was only 0.40 to 0.41. An attribution method highlighted which parts of the record drove an individual prediction.

The limits

What it doesn't show

This is a retrospective study, so it does not show that using the predictions would improve patient care; prospective trials are needed. Models were trained and tested separately within each hospital and data were not harmonised across sites, so transfer to a new hospital is untested. Because only one hospital had notes and the cohorts differed, the study cannot say how much the free text contributed. The interpretability example is a single case study, and the code depends on internal Google infrastructure, limiting replication.

Key terms

Electronic health record (EHR)
The digital record of a patient's care, including labs, medications, vital signs and clinical notes.
FHIR
A standard container format for healthcare data that makes records from different systems easier to exchange.
AUROC
Area under the ROC curve, measuring how well a model ranks positive cases above negative ones; 0.5 is chance, 1 is perfect.
Ensembling
Combining the predictions of several models to get a more accurate overall prediction.
Work-up-to-detection ratio
How many flagged patients must be checked to find one true case; lower means fewer false alarms.
Attribution
A method for showing which input data most influenced a model's prediction for one case.

Flashcards

1 / 10

0 of 10 answers reviewed

Research intelligence for this paper

See its role on concept claims, tensions it is part of, placement history, and related discoveries.

Open paper intelligence

Quiz yourself

1 / 5

What was the main input to the deep learning models?

Common questions

What is new here compared with earlier clinical prediction models?

Instead of hand-selecting a few dozen variables, the model used everything in the record, including free text, with one data format for every task.

Does a higher AUROC mean patients would do better?

Not by itself. The authors stress that prospective trials are needed to show that acting on the predictions improves care.

Why was predicting diagnoses so hard?

Each stay can have many codes drawn from about 14,000 options, and many codes are nearly identical, so exact matching is demanding.

More on Clinical NLP