Pain assessment is a key factor in successful pain management [1]. Inaccurate postoperative pain assessment may cause illnesses [2] and even long-term chronic issues [3]. Pain assessment tools for clinical use are in great demand. If communication ability is limited or even lost due to surgery or illness complications, it is difficult for a doctor to determine the patient's feelings. A proper pain assessment tool can offer an approximate pain level of that patient for further treatment. For now, the wide range of pain assessment methods still cannot determine the precise pain prevalence and levels for adults in hospitals [4]. That may cause incorrect treatment and lead to various problems and risks for patients. Painkillers have many side effects, and overtreatment of pain can trigger respiratory depression in the short term or substance use disorder in the long term [5]. However, undertreatment of pain may result in chronic pain, more health care costs, and physiological and psychological suffering [3,6]. All these issues mentioned are prevalent among noncommunicative patients [7]. A valid pain assessment tool would be truly transformative to health care delivery as clinicians could deliver pain treatments and assess response in real time. This would decrease unwanted side effects and recovery time from illness or a procedural intervention.

With the rapid development of Internet-of-Things (IoT) devices, including wearable devices, automated and continuous objective pain intensity assessment is possible [8]. The accuracy of these wearable devices has been evaluated in several studies. As an example, Mehrabadi et al [9] validated the accuracy of these devices in terms of sleep. Researchers try to identify nervous reaction to pain by monitoring the fluctuation in patients' physiological data, including electromyography, electrocardiography, photoplethysmography, and electrodermal activity (EDA) in real time [10,11]. Other research uses facial expression and head movement to accompany the physiological data [12,13]. These methods can quantify pain intensity, especially for poorly communicating patients [14]. However, all these methods to date use stimulated pain and were evaluated on healthy participants. Based on this observation, we developed the UCI iHurt Dataset (UCI_iHurtDB) [15]. The UCI_iHurtDB is the first multimodal dataset collected from postoperative adult patients suffering from real pain in hospitals. This dataset is planned to be released for research purposes in the near future.

Skin conductance or the EDA signal is considered a useful biomedical data point that corresponds to pain perception. Our skin produces sweat via over 3 million small tubular sweat glands. Sweat glands are distributed across the body but with the highest densities on the soles of the feet, palms and fingers, and forehead and cheeks. If a patient is exposed to a certain group of stimuli, they can be triggered to secrete moisture, termed emotional sweating. This results in a decrease in skin resistance, or in other words, an increase in skin conductance [16], which is also known as EDA or galvanic skin response (GSR). Other than pain, the EDA signal can extract a variety of valuable information from the human body. Rostami et al [17] highlighted an important insight that, by using the GSR signal, the biological impact of food on a person’s body can be captured.

Pain assessment research only using EDA data is limited. Eriksson et al [18] and Munsters et al [19] validated the relationship between EDA and pain for newborn infants, suitable for automated pain assessment due to their inability to communicate. By monitoring the EDA data during routine blood sampling or care intervention, they found EDA can differentiate between pain and no pain; however, more research is needed to achieve a clinical-grade level. Manivannan et al [20] verified whether the EDA could be used as a valid pain indicator for hypnotic analgesia with 10 participants. They used an iron disk to create mechanical pain in a laboratory setup. The experimental results show a clear relation between pain scores and EDA. None of these mentioned works used machine learning algorithms to create a classification model for pain assessment. Furthermore, their dataset includes healthy patients with various stimulus methods to cause pain. In another work, Susam et al [21] attempted to assess postoperative pain using EDA through a machine learning model. Their model could distinguish between clinical moderate-to-severe pain and no-pain conditions. However, their work only focused on children as a population.

To the best of our knowledge, for the first time in this paper, we present an automatic and versatile pain assessment tool to predict different pain levels in postoperative adults using only EDA signals as a physiological signal. In our pain assessment tool, we used 11 different time-domain features extracted from EDA signals for prediction. A feature selection algorithm is used to increase our tool's prediction accuracy and find the top-most important features related to pain intensity. To evaluate our results, we used different types of machine learning algorithms. Machine learning techniques and neural networks have been widely used in health monitoring domains. Zargari et al [22] used a combination of convolutional neural networks and recurrent neural networks to accurately track the position of in-mouth nutrient sensors. Mehrabadi et al [23] used convolutional neural networks to detect COVID-19 in patients with acute respiratory distress syndrome. These techniques are not limited only to health care domains. Ashrafiamiri et al [24] used deep neural networks to secure autonomous driving. To accurately validate our pain assessment algorithm's performance, we compared our results with the accuracy achieved for the pain models presented in [13]. Werner et al [13] used the BioVid Heat Pain database in their work in which participants were subjected to painful heat stimuli under controlled settings. This comparison aims to show that despite all the challenges in real hospital settings, our models can achieve comparable results. This also shows the feasibility of using artificial intelligence–based objective pain assessment for real patients.

All methods of the study, including the data collection and pain assessment, were approved by the University of California Irvine Institutional Review Board (IRB, HS: 2017-3747). Potential candidates were screened for eligibility using the Acute Pain Service schedule and provided with a copy of the consent form to review for at least 24 hours before participation in research procedures.

A total of 25 patients with acute pain were engaged by the Acute Pain Service and recruited for this study. Of these 25 participants, 5 were removed from our study due to problems in the data collection process due to rapid hand movements or unreliable EDA signal due to conditions such as skin rash and itching resulting from wearing the Empatica E4 watch on their IV arm. The average age of patients in this study was 54.45 (SD 17.44, range 23-89) years; 55% (11/20) of the patients were male, and 45% (9/20) of patients were female. All patients were taking the standard-of-care postoperative pain medications at the time of the study. Enrolled participants agreed to perform the research protocol for a median of 4 (IQR 3-6) days after surgical intervention. The nature of the procedures for each participant included the following domains: 45% (9/20) general surgery (diagnostic laparoscopy, exploratory laparotomy, and vascular), 15% (3/20) trauma (thoracic pain and rib plating), and 10% (2/20) urology (cystectomy and bladder augmentation). Also, 40% (8/20) of enrolled participants received standard-of-care epidural analgesia provided by the acute pain service team at the time of research participation. The remaining participants were receiving oral and intravenous analgesics for pain control (Table 4).

To show that our pain assessment algorithm can achieve comparable results to the work by Werner et al [13], we used identical settings as their work. Werner et al [13] used 5 different pain levels, including the baseline. They also considered 5.5-second windows for the GSR data. Therefore, in the Data Preparation part of our pipeline, we considered 5.5-second windows for the slices of the GSR data for feature extraction (2.75 seconds before and after each timestamp). Furthermore, as discussed in the Methods section, we down-sampled the pain levels from 11 classes to 5 classes to make them similar with their labels.

At first, we used the set of features that was used in the work by Werner et al [13] for classification without any feature selection. The maximum value of the peaks, range, standard deviation, interquartile range, root mean square, mean value of local maxima, mean value of local minima, mean of the absolute values of the first differences, and mean of the absolute values of the second differences are the features that were used here.

We used leave-one-person-out cross-validation using k-nearest-neighbor and random forest algorithms. We reported the accuracy based on 4 different pain models (BL vs PL1, BL vs PL2, BL vs PL3, and BL vs PL4). Table 5 shows the comparison of the validation accuracy achieved by our classifiers with that by the pain models of Werner et al [13].

According to these data, for the first 2 pain models (BL vs PL1 and BL vs PL2), we achieved a higher accuracy using both of our classifiers in comparison with Werner et al [13]. For the third pain model, our accuracy is also close to their models, with less than 1% difference using the k-nearest-neighbor classifier. For the fourth model, the accuracy of our models was noticeably lower than their models. As the next step, we added 2 more features (the number of peaks and the mean of the GSR data) to our set of features and then selected the most important ones using the mean decrease impurity method to improve the accuracy.

To obtain the best set of classification features, we ran leave-one-person-out cross-validation on different pain models using random forest classifiers. We computed the Gini importance of the features on the training data and selected the top k number of features for training the model and classification (2-11 were considered to be possible values for k). Since we had a different number of folds, we could have different sets of features for each fold. We considered the set of features that was used in most of the folds as the final set of features for the current pain model. Table 6 shows the selected features for each pain model. Descriptions for each of these features can be found in Table 1.

According to this table, the maximum value and root mean square of the signal are 2 features that were selected in all the pain models. The mean value of the local maxima and the mean value of the signal were also selected for all the pain models except the third one. The difference between upper and lower quartiles of the signal (IQR) is a feature that was selected as an important feature for classification only for the BL vs PL4 pain model.

After the set of features for each pain model were obtained, we ran one-person-leave-out cross-validation for k-nearest neighbor and random forest algorithms using these sets of new selected features to achieve the final results.

Table 7 shows the validation accuracy comparison of our models with those by Werner et al [13] using feature selection for 5.5-second windows. As shown in this table, by using feature selection, our classifiers' accuracy improved compared to those shown in Table 5. For all the pain models except the fourth one, we were able to achieve higher accuracy than the pain models by Werner et al [13]. For BL vs PL4, our accuracy was still about 10% less than their work. In the Discussion section, we explain the potential reasons for this difference in our model.