Improving Prediction of Low-Prior Clinical Events with Simultaneous\n General Patient-State Representation Learning

Abstract

Low-prior targets are common among many important clinical events, which\nintroduces the challenge of having enough data to support learning of their\npredictive models. Many prior works have addressed this problem by first\nbuilding a general patient-state representation model, and then adapting it to\na new low-prior prediction target. In this schema, there is potential for the\npredictive performance to be hindered by the misalignment between the general\npatient-state model and the target task. To overcome this challenge, we propose\na new method that simultaneously optimizes a shared model through multi-task\nlearning of both the low-prior supervised target and general purpose\npatient-state representation (GPSR). More specifically, our method improves\nprediction performance of a low-prior task by jointly optimizing a shared model\nthat combines the loss of the target event and a broad range of generic\nclinical events. We study the approach in the context of Recurrent Neural\nNetworks (RNNs). Through extensive experiments on multiple clinical event\ntargets using MIMIC-III data, we show that the inclusion of general\npatient-state representation tasks during model training improves the\nprediction of individual low-prior targets.\n

Ähnliche Arbeiten