Deep Neuroevolution: Training Deep Neural Networks for False Alarm Detection in Intensive Care Units

We present a neuroevolution based-approach for training neural networks based on genetic algorithms, as applied to the problem of detecting false alarms in Intensive Care Units (ICU) based on physiological data. Typically, optimisation in neural networks is performed via backpropagation (BP) with stochastic gradient-based learning. Nevertheless, recent works (c.f., [1]) have shown promising results in terms of utilising gradient-free, population-based genetic algorithms, suggesting that in certain cases gradient-based optimisation is not the best approach to follow. In this paper, we empirically show that utilising evolutionary and swarm intelligence algorithms can improve the performance of deep neural networks in problems such as the detection of false alarms in ICU. In more detail, we present results that improve the state-of-the-art accuracy on the corresponding Physionet challenge, while reducing the number of suppressed true alarms by deploying and adapting Dispersive Flies Optimisation (DFO).