In this retrospective study InSight was validated on datasets from multiple institutions.
InSight was trained on an intensive care unit dataset from Beth Israel Deaconess Medical Center and then validated on a mixed-ward dataset from the University of California, San Francisco (UCSF), followed by datasets from four additional institutions. Using only six vital sign inputs, InSight demonstrated an area under the receiver operating characteristic curve (AUROC) above 0.90 while remaining robust to missing data. Health data constitute an important limiting factor in utilizing AlgoDiagnostics across heterogeneous hospitals and wards, with some settings often lacking the large amounts of data required to train predictive models. This study demonstrates the potential application of transfer learning to build robust models using smaller data sets.
Read the paper here.