Sepsis Detection Using Features Extracted from Photoplethysmography

Sepsis is a life-threatening clinical condition associated with high mortality rates. It is caused by an overreaction of the immune response to an infection that can lead to impaired microcirculatory function with a critical reduction in tissue perfusion and subsequent organ failure. Early identification and treatment are fundamental to improve patients’ chances of survival. One of the techniques that can provide important information on microcirculatory function is photoplethysmography (PPG). This technique is applied in the pulse oximeter device, a non-invasive and low-cost solution that is used in the clinical setting to measure blood oxygenation. In this study, we propose a method for sepsis detection from the photoplethysmographic signal using machine learning algorithms. For the development of the method, a dataset of PPG signals belonging to both septic and non-septic subjects was obtained from the MIMIC III Database. From the collected signals, a set of features were selected and extracted and used as input for the XGBoost classifier with the aim of discriminating between septic and non-septic subjects. Our method achieved an accuracy of 70.96%, a sensitivity of 64.12% and a specificity of 76.85% in PPG samples classification and an accuracy of 72.97%, sensitivity of 64.71% and specificity of 80% in subject classification. This method was compared with our previous study in which a deep learning approach was used on the same dataset. The results indicate that the deep learning approach outperforms the classic approach.