Data-Driven Approaches to Assess Acute Pain Intensity Using Physiological and Behavioural Data

Abstract

Pain is a highly personal and stressful subjective experience linked to damage to the tissue. At present, the management of pain continues to remain ambiguous and disappointing at hospitals. In particular, the management of patients' postoperative pain has become a major medical and nursing challenge. Hospitals have taken initiatives to measure pain using self-report measures such as the Visual Analogue Scale (VAS) and the Numeric Pain Intensity Scale (NPIS). But these methods are inaccurate and subjective as it depends on the patient's input. Therefore, there is a need for an objective, quantitative method to monitor pain continuously. Thus, this work presents the various data-driven approaches to automatically measure and monitor postoperative patient's pain severity levels continuously. This work utilizes minimal raw data, i.e., two physiological signal data and one behavioral data, to determine pain. Therefore, this research work reduces the constraints imposed by multimodal signal processing and also helps to establish the field of wearable technologies. The physiological signals used for the study are Electrocardiogram (ECG) and Electro-Dermal activity (EDA), and the behavioral data used for the study are the facial expressions of the individuals. Evidence from several cohort studies has shown that physiological signals such as ECG and EDA signals and the facial expression data of individuals are the best sources of the presence of acute pain in adults (especially in postoperative patients).A filter-based method, i.e., one-way ANOVA, is applied to the data to select the best pain-associated features. Thus, the features are selected based on statistically significant values (P < 0.05) for the classification. Finally, a classification task implementation helps to classify five different levels of pain (Pain Index: 0-4, namely No pain as BL1, Mild pain as PA1, Moderate pain as PA2, Severe pain as PA3, Intolerable pain as PA4) using supervised ML algorithms such as Neural Network (NN), Support Vector Machine (SVM), and Random Forest (RF) and DL algorithms like a hybrid Convolutional Neural Network Long Short-Term Memory Network (CNN_LSTM). The algorithm's performance is tested using the following metrics: classification accuracy, recall, precision, f1-score, and confusion matrix. This work utilizes the BioVid Heat Pain database (BVHP DB) and the UNBC-McMaster Shoulder Pain Expression Archive database (UNBC DB)