1932-4553 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSTSP.2016.2535962, IEEE Journal of Selected Topics in Signal Processing IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 13, NO. 9, SEPTEMBER 2014 1 Methods for Person-Centered Continuous Pain Intensity Assessment from Bio-Physiological Channels Markus K¨ achele, Patrick Thiam, Mohammadreza Amirian, Friedhelm Schwenker and G¨ unther Palm Abstract—In this work, we present methods for the personal- ization of a system for the continuous estimation of pain intensity from bio-physiological channels. We investigate various ways to estimate the similarity of persons and to retrieve the most informative ones using meta information, personality traits and machine learning techniques. Given this information, specialized classifiers can be created that are both, more efficient in terms of complexity and training times and also more accurate than classifiers trained on the complete data. To capture the most information in the different bio-physiological channels, we cover a broad spectrum of different feature extraction algorithms. Furthermore, we show that the system is capable of running in real-time and discuss issues that arise when dealing with incremental data processing. In extensive experiments we verify the validity of our approach. Index Terms—pain intensity estimation, personalization, semi- supervised learning, machine learning I. I NTRODUCTION F ROM an evolutionary point of view, pain can be con- sidered as a mechanism to prevent harmful behaviour and as an indicator of a medical condition. Nowadays, the evolutionary aspect has shifted to more clinical settings, where pain is an important symptom that is used to measure the level and acuteness of a patient’s injury or illness. The automatic recognition of pain would help in situations in which self assessment is not possible due to one or several factors that prevent pain perception and successful communication (such as neurodegenerate diseases, for example). An automatic assessment of pain intensity in such situations would provide valuable insights that could be exploited in order to eventually localize the source of pain and accordingly to define an appropriate therapy. Thus, it would improve the health-related quality of life of patients. Recent advances in automated pain recognition have mainly focused on prediction from facial expressions [1]–[5]. The face serves as a straightforward means to communicate pain to other people, since a huge amount of information about an individual’s affective state can be retrieved just by ob- serving the face [6], [7]. However, facial expression-driven pain assessment implies the tracking of the facial region of the individual, which can be very cumbersome in a clin- ical setting. Meanwhile, pain assessment is also possible from other signals. Thus research effort has been targeted The authors are with the Institute of Neural Information Processing, Ulm University, Ulm, Germany Manuscript received April 19, 2005; revised September 17, 2014. towards other modalities which ultimately led to multi-modal recognition systems. Such recognition systems incorporate a multitude of different modalities such as bio-physiological channels (e.g. heart rate, electrodermal activity) [8], linguistic and paralinguistic measures or video signals, into complex classification systems. Pain recognition based on multi-modal signals has recently been proven to be very effective, most often outperforming unimodal systems significantly [9], [10]. With the increasing availability of smart devices and wear- ables such as fitness trackers, many people seem to have gained awareness of their health and body. The non-invasive integration of such technologies in our daily live activities provides new insights into health monitoring. As many of them are equipped with sensors to capture bio-physiological measurements such as the heart rate or the galvanic skin response, they are an affordable alternative for each individual to monitor his/her body functions. Moreover, recent studies have pointed out the need for providing quality medical monitoring to a continuously increasing population. One of the proposed solutions resides in the use of wearable sensor systems [11], [12]. The capabilities of such systems include, among others, motion, physiological and biochemical sensing. An appropriate assessment and interpretation of the data recorded by such sensors might help monitor, assess and eventually solve a large variety of health related problems. The focus of this paper is the continuous recognition of pain intensity from multiple bio-physiological channels as they might as well be found on wearable devices on the end customer market. To date, there are only few works that have addressed the problem of automatic pain detection from bio- physiological channels. Chu et al. [13] used blood volume pulse (BVP), electrocardiogram (ECG) and skin conductance level (SCL) in combination with linear discriminant analysis (LDA) as a classifier in order to discriminate between seven physiological states including five levels of pain. Olugbade et al. [14] used electromyography (EMG) paired with the analysis of body movements in combination with Random Forests (RF) and Support Vector Machines (SVM) as classifiers to distinguish between three different pain states. However, the latter focuses on an offline scenario where the entire dataset is completely pre-processed before the analysis and classification processes are undertaken. Several research groups began to couple wearable devices with computational intelligence and developed interesting ap- plications especially in the medical sector. In [15], EMG is used as input for a neurofuzzy controller to allow paralyzed