The analysis of pain sensation is of high importance in numerous medical applications. For example, in physiotherapeutic treatments, pain, or especially the pain threshold, can not only determine the course or result of the treatment, but can also influence the structure and composition of the exercises from the very beginning. The quantitative assessment of one's own pain is traditionally based on a self-assessment with questionnaires. However, this method is not suitable for patients who are not able to communicate their pain (objectively).

In cooperation with Prof. Kerstin Lüdtke (Physiotherapy, University of Lübeck) we are developing a learning-based pattern recognition platform that will automatically determine the current level of pain based on data from multiple, portable sensors. The data acquisition is performed with multiple devices. Pain is induced by heat using a thermode, CHEPS (Contact Heat-Evoked Potential Stimulator), on the forearm of the non-dominant arm of the study participants. During the experiment, participants are asked to continuously report their perceived pain level using a CoVAS (Computerised Visual Analogue Scale). In order to record the physical response of people in pain, two wearables (Empatica E4 and RespiBan) record different physiological characteristics (such as BVP, EDA, EMG, ECG) during the test. The data from the wearables, CoVAS and Thermode are synchronized and adjusted for sampling rate.

A dataset with a total of 52 persons was collected in the course of the PainMonit project. To classify the data, we consider so-called 2-class problems, for example, no pain vs. high pain. In an in-depth analysis, a Random Forest method based on manually defined features was compared with generated features by different Deep Learning approaches for automatic learning based on the raw data. Here, a maximum accuracy of up to 93% (no pain vs. high pain) could be achieved. Further, it could be shown that Deep Learning approaches do not necessarily provide better results than classical pattern recognition methods for pain detection. Furthermore, the integration of the subjective sensation (CoVAS) leads to an increased performance of the system. Future work in this area will focus on improving the classification for the task no pain vs. low pain using an adapted sensor fusion. An important aspect of our research in this area is to develop such machine learning methods for solving the problem, which will provide us with new insights into the physiological and behavioural markers of pain.

In current work, we would like to further evaluate and improve our system for chronic back pain patients and by inducing a "natural" pain using movement in osteoarthritis patients.