Multi-horizon event detection for in-hospital clinical deterioration using dual-channel graph attention network

In hospitals globally, the occurrence of clinical deterioration within the hospital setting poses a significant healthcare burden. Rapid clinical intervention becomes a crucial task in such cases. In this research, we propose an end-to-end deep learning architecture that interpolates high-dimensional sequential data for the early detection of clinical deterioration events.

We consider the problem of detecting deterioration events with two stages: predicting the "detection" status, a pre-event state; and predicting the event from detection time. Our approach involves the development of dual-channel graph attention networks with multi-task learning strategy by jointly learning task relatedness with a shared model for multiple prediction in multivariate time-series.

The experiments are conducted on two clinical time-series datasets collected from intensive care units (ICUs). Our model has shown the potential performance compared to other state-of-the-art methods, in terms of the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC).

The proposed dual-channel graph attention networks can explicitly learn the correlations in both features and time domains of multivariate time-series. Our proposed objective function also can handle the problems of learning task relations and reducing task imbalance effects in multi-task learning.

Applying our proposed framework architecture could facilitate the implementation of early detecting in-hospital deterioration events.