Session S91.3

PhysioNet/CinC Challenge

JH Henriques*

University of Coimbra
Coimbra, Portugal

This work proposes the application of neural network multi-models to the prediction of adverse acute hypotensive episodes (AHE) occurring in intensive care units. Contrasting to classical auto regressive representation, multi-models do not use recursively model outputs as inputs for step ahead predictions. As result, prediction errors are not propagated over the forecast horizon and long-term predictions can be accurately estimated. Applied to vital signs time-series, considered in the PhysioNet/CinC challenge, the proposed strategy enables to adequately capture arterial blood pressure dynamics and, consequently, to predict the onset of hypotensive events over a defined forecast window.
Regression representations are common techniques for modeling and prediction tasks. By means of auto-regressive representations, information from past instants can be used to estimate future values, usually one-step ahead. Using a multi-model strategy, one independent sub-model is employed for each sampling instant within the prediction horizon. Thus, since independent sub-models are used for consecutive sampling instant, future predictions do not depend on previous predictions. Among regression models, neural networks have shown considerable capabilities to learn and generalize from non-linear environments, enabling to capture the fundamental data dynamics. Moreover, multi-models can be trained by means of standard backpropagation algorithms. In fact, each independent neural sub-model is used for each sampling instant and does not depend on previous predictions.
In this work, General Regression Neural Networks models, integrated into a multi-model structure, were employed. These neural sub-models were trained and validated using H and C data sets (sampled once per minute) available in the “numerics record”. For this purpose, time-series values obtained from an appropriate period of time, immediately before and after the beginning of the defined forecast windows, were used. Data was previously pre-processed, namely to deal with non-existent information and noise reduction. No information from “clinical records” was used. To address the prediction challenge, the “b” segments of A and B data sets were resampled (to one sampling per minute). The forecast was made using the trained neural multi-model structure, only considering the information available before the forecast window. The occurrence of an AHE within the forecast window (one hour) was assessed according to AHE definition (see PhysioNet/CinC Challenge).

(Abstract Control Number: 266)