Dascena publishes on mortality prediction in a data-scarce hospital setting
In this study Dascena demonstrates a transfer learning technique to be used in specializing machine learning algorithm. The findings are published in Biomedical Informatics Insights.
Algorithm based clinical decision support (CDS) systems require sufficient amounts of training data to produce quality associations between patient health data. This presents an obstacle to widespread adoption and high performance of these systems, as not every target site possesses a sufficient amount of patient data. To overcome this obstacle, the machine learning mortality prediction algorithm AutoTriage was trained on a large set of source data before being applied to the target site. This significantly reduced the amount of data needed at the target site to achieve an area under the receiver operating characteristic curve (AUROC) of 0.80, which represents a greater sensitivity and specificity than that of the existing Modified Early Warning Score (MEWS) system with an AUROC of 0.76. This source training method reduced the number of patients needed at the target site from over 4,000 to about 220, and it reduced the time required to collect this data from an average of six months to ten days. This study highlights the usefulness of the transfer learning technique in the specialization of machine learning algorithms to new hospitals.