3.1.2. PainMonit Database (PMDB)

The PainMonit Database (PMDB) was acquired at the Institute of Medical Informatics, University of Lübeck, Germany, following the findings of a preliminary study investigating heat-induced pain in a small dataset containing 10 subjects in Gouverneur et al. [34]. A Pathway CHEPS (Contact Heat-Evoked Potential Stimulator thermode, Medoc, Ramat Yishay, Israel) with a 27 mm diameter contact surface was attached to the non-dominant forearm interior site (10 cm below the elbow) of participants to induce pain by thermal stimuli as it is one of the most commonly used stimuli to induce experimental pain. In total 55 subjects (21 male and 33 female with an average age of 27.35±6.88) participated in the study. Healthy people between the age of 18 and 65 were recruited. In contrast, chronic pain disorders, acute pain, skin diseases that could be a contraindication to the thermode, pregnancy, neurological, psychiatric or psychological diseases, and regular use of medications (except contraceptives) were defined as exclusion criteria. The main difference between PMDB and existing benchmark in the literature is the presence of subjective pain annotations in addition to the objective temperature-based ones. Subjective feedback was obtained using a Computerised Visual Analogue Scale (CoVAS) slider (Computerized Visual Analogue Scale, Medoc, Ramat Yishay, Israel), a simple slider whose position is digitalised and returns ratings between 0 and 100. Like an ordinary VAS, the far left location represents no pain while the far right is associated with the worst pain imaginable. The pain induction machine, the thermode and slider can be seen in Figure 2a–c, respectively.

The Medoc devices used during data acquisition. (a) Medoc Pathway system. (b) Cheps thermode.(c) Computerised Visual Analogue Scale (CoVAS) slider.

The new data acquisition protocol includes a calibration and induction phase. The calibration is based on recording the parameters pain threshold (TP, threshold when heat stimulus becomes painful) and pain tolerance threshold (TT, threshold when pain becomes unbearable) individually for each subject. Following a staircase calibration method, increasing 10-s temperature stimuli with 5 s pause in-between were given to the subject. The protocol started with a temperature of 40 ∘C and increased the stimulations by 1 ∘C each time up to a maximum of 49 ∘C. Participants were asked to continuously rate their pain perception utilising the CoVAS. Temperatures exceeding 0 and 90 in CoVAS rating for the first time were noted as TP and TT, respectively. To ensure further robustness of the calibration, these parameters were recorded twice for each subject and averaged. Subsequently, the thresholds were further tested to check their validity. During a calibration check, TP and TT were once again applied and rated by participants. If the former threshold was perceived as vigorously painful (CoVAS above 10), it was adjusted by reducing it by 1 ∘C. Equally, TT was raised by 1 ∘C if its initial value did not retrieve CoVAS values above 90. Four painful temperature stimuli Pi were then defined using the thresholds TP and TT with the following equation:

with i∈{1,2,3,4} and R=(TT−TP)/4. Moreover, a non-painful temperature NP was defined by NP=TP−R. Figure 3 illustrates baseline, non-painful and painful temperatures with the associated thresholds.

The 5 temperature intervals defined by the temperature range R in dependency of TP and TT. While the first temperature resembles a non-painful stimulus, the last 4 are meant to evoke pain.

During the pain induction phase, eight 10-s stimuli were applied for each of the temperatures Pi defined previously. Between randomised stimuli, the temperature returned to the baseline at 32 ∘C, and a resting phase of random duration between 20 to 30 s was applied. Participants were asked to rate their pain continuously using the CoVAS. To avoid any possibility of harm and sensitisation or habituation effects the thermode was repositioned after half of the impulses.

During the pain induction phase, various physiological sensors were recorded by two different wearable devices. Data of both were transferred via Bluetooth to one machine in real-time. On the one side, the wristband Empatica E4 (E4) (Empatica E4, Empatica Inc., Boston, United States) was worn on the non-dominant arm to avoid movement artefacts and recorded Blood Volume Pulse (BVP) from which Heart Rate (HR) and Inter-Beats-Interval (IBI) are computed, EDA, Accelerometer (ACC), and skin temperature in 64, 4, and 32 Hz, respectively. On the other side, respiBAN Professional (RB) (respiBAN Professional, Plux, Lisbon, Portugal), a chest-worn device registering respiration and various physiological modalities with a sampling rate of 1000 Hz, was included. Two electrodes were placed at the medial phalanx of the index and middle finger of the non-dominant arm to capture EDA. Moreover, the activity of the heart was measured by monitoring ECG with a positive electrode at the upper left, a negative electrode at the upper right pectoral and a reference electrode placed at the right waist. In addition, an electrode placed on the skin above the trapezius muscle recorded its activity via Electromyography (sEMG). To further reduce noise and artefacts a reference electrode was placed above the 7th cervical vertebrae. Moreover, the same self-adhesive disposable electrodes (Kendall Covidien H124SG ⌀ 24 mm, Wolfram Droh GmbH, Mainz, Germany) were stuck on all sites. No additional gel or other substance was needed, as the Ag/AgCl sensor is embedded in an adhesive ad conductive hydrogel. While the application takes a little longer than simple dry electrodes with a Velcro strap, Ag/AgCl hydrogel electrodes stick safely and provide the most reliable EDA signals [35]. As washing hands decreases skin conductance because it removes sweat and other conductance increasing substances, participants were asked to wash their hands with simple soap immediately before the procedure to standardise the time since the last handwashing ([36], p. 657). To further restrict the noise level, EMG and ECG sites were cleansed with simple alcohol pads [37].

In addition to the physiological sensors, behaviour responses were also recorded. An HD Webcam Pro C920 (C920, Logitech, Lausanne, Switzerland) placed in front of the subjects, captured RBG-video information incorporating facial movements. Moreover, depth information was gathered using an Intel^® RealSense™ D435 (D435, Intel Corporation, Santa Clara, CA, USA) camera. To ensure consistent and sufficient lighting, two light boxes were setup in a 45∘ to the left and right in front of the participant and shutters of the windows were closed. While aiming for a comparable setup, the same video modalities like BVDB (RGB and depth information) were recorded, but differences in devices (e.g., D435 vs. Kinect) could introduce differences in the resulting datasets. Subjects were asked to sit comfortably, rest the non-dominant arm and use the other to rate their pain using the CoVAS. The study took roughly one hour for each subject. Because of technical issues during the recording or flawed conduction of the experimental setup, three subjects were removed, creating a final dataset of 52 subjects in total.