Rui Cao, Seyed Amir Hossein Aqajari, Emad Kasaeyan Naeini, Amir M. Rahmani; Original publication date: November 2021; DOI: 10.1109/EMBC46164.2021.9630002

Abstract

Pain, as a multivalent, dynamic and ambiguous phenomenon is difficult to objectively quantify, in particular, in real clinical settings due to several uncontrollable factors. Respiratory rate is one of the bio-signals whose fluctuations strongly correlates with pain, however, it has been often neglected due to its monitoring difficulties. In this paper, to the best of our knowledge for the first time, we propose an objective pain assessment method using respiratory rate derived from wristband-recorded Photoplethysmography (PPG) signals collected from real post-operative patients (in contrast to the existing studies analyzing stimulated pain). We first derive respiratory rate from post-operative patients’ PPG signals using an Empirical Mode Decomposition (EMD) based method and extract several statistical features from it. We then implement a feature selection method to identify the top most significant features, and exploit a weak supervision method to address the unbalanced nature of the collected labels in real settings. Several machine learning algorithms are applied to perform binary classification of no pain (NP) vs. three distinct pain levels (PL1 through PL3). We obtain prediction accuracy of up to 81.41% (NP vs. PL1), 80.36% (NP vs. PL2) and 79.48% (NP vs. PL3) which outperform the results reported by the state-of-the-art, despite obtained from data collected from real post-operative patients.