Frequency and Time Domain Feature Engineering and Predictive Modeling Based on ECG, SpO2, and Respiration Signals

Abstract

Today, there exists a challenge in simplifying biosignals into features that are well suited for machine learning and clinician understanding. This work reports the feature engineering exercise involved with such challenge, along with the predictive modeling. We primarily tackle ECG, Respiration (Thoracic Impedance), and SpO2 (Plethsmographic) signals extracted from a proprietary dataset used by GE Healthcare. Throughout the study, we analyze biosignals while searching for general characteristics which may help describe (and even highlight) human function for a machine learning model, while maintaining clinical value. Wave Morphology Analysis in the Time Domain, Wavelet Decomposition and Fast Fourier Transforms were the main methods explored for feature engineering. Finally, results from a Convolutional Neural Network and a Random Forest model are reported, whereby the best performing model is able to predict Sepsis with 77% accuracy at least three (3) hours in advance.