Figure 3.

Diagram for submitting and distributing containerized prediction models in a protected environment. Dockerized models were submitted to Synapse by a model developer to an evaluation queue. The Synapse Workflow Hook pulled in the submitted Docker image and built it inside the protected University of Washington (UW) environment. The model trained on the available electronic health record data and then made inferences on the evaluation dataset patients, outputting a prediction file with mortality probability scores for each patient. The prediction file was compared with a gold standard benchmark. The model’s performance, measured by area under the receiver-operating characteristic curve, was returned to the model developer. CWL: Common Workflow Language.