Learning to Predict Coronary Perfusion Pressure During Cardiopulmonary Resuscitation

The goal of this work is to advance the capability of automated, mechanical cardiopulmonary resuscitation (CPR) by predicting Coronary Perfusion Pressure (CPP) within 5 mmHg at a given moment in time. We aim to utilize methods from machine learning in order to model the CPP of a porcine patient subjected to automated chest compressions. During preprocessing of the data, we show how data sampling rate, delays and moving average filtering can improve predictions. We demonstrate state of the art modeling performance utilizing a variety of algorithms, and analyze the performance of each algorithm for single-step and long-term predictions. The results indicate that a delayed linear system achieves this target CPP within 0.25 mmHg. For longer time horizons, a more complex model is required. We demonstrate that the Long-short-term-memory (LSTM) network has the best single run performance, while the Sparse Spectrum Gaussian Process (SSGP) has the best average performance.