A Recurrent Neural Network for Real-Time Heart Attack Forecasting

Cardiac arrest has a hospital mortality rate of ~80% and is a serious concern for many ICU patients. A paper (Ho et al., 2013) published in the Journal of Machine Learning Research made a first attempt at creating a dynamic cardiac risk estimation model which incorporates different temporal signatures in a patient’s risk trajectory. Their model, DYNACARE, allows for real-time interpretability and predictability of a cardiac arrest event. They have two flavors of DYNACARE, a switching model with two latent variables, one controlling the stochastic volatility and the other capturing the binary state of the process, along with a threshold model with a single latent variable controlling the risk trajectory.

We seek to improve upon their work by trying a variety of models including recurrent neural networks and other higher-dimensional graphical models, as well as various ensembles of these models. The general intuition is that by increasing the complexity of the model along with its ability to learn long term time dependencies, we will be able to increase predictive power and give a better hazard function which conveys the risk of the patient. The paper can be found here.