Automatic recognition of pain, anxiety, engagement and tiredness for virtual rehabilitation from stroke: A marginalization approach

Abstract

Virtual rehabilitation taps affective computing to personalize therapy. States of anxiety, pain and engagement (affective) and tiredness (physical or psychological) were studied to be inferable from metrics of 3D hand location-proxy of hand movement-and fingers' pressure relevant for upper limb motor recovery. Features from the data streams characterized the motor dynamics of 2 stroke patients attending 10 sessions of motor virtual rehabilitation. Experts tagged states manifestations from videos. We aid classification contributing with a marginalization mechanism whereby absent input is reconstructed. With the hand movement information absent, marginalization statistically outperformed a base model where such input is ignored. Marginalized classification performance was (Area below ROC curve: µ ± σ) 0.880 ± 0.173 and 0.738 ± 0.177 for each patient. Marginalization aid classification sustaining performance under input failure or permitting different sensing settings.

Automatic Recognition Methods Supporting Pain Assessment: A Survey

“…In this context, body movement is also represented by selected parameters of the human skeleton obtained with motion capturing [63]- [65]. In a rehabilitation context, Rivas et al [94] extracted mean position, speed, and acceleration from hand movement and finger pressure time series.…”

“…Interestingly, Brahnam et al [66] trained their network with a genetic algorithm instead of backpropagation. In addition to this, researchers have used a Gaussian Mixture Model (GMM) [71], [85], a variant of naive Bayes classifier [94], and Genetic Selection of a Fuzzy Model (GSFM) [70].…”

“…The supervised and unsupervised methods are marked PMS and PMU respectively in Tables 3, 4, and 7 (not specifically marked in 2). Several authors [4], [56], [57], [60], [61], [73], [94], [102], [105] trained person-specific models only using data of the person of interest, some comparing with generic models and other methods. Lopez-Martinez and Picard [50] and Lopez-Martinez et al [51] used multi-task learning developing a set of person-specific classifiers that share a common hidden neural network layer.…”

“…Most other work used the applied stimulus as ground truth, either with [78], [80], [88], [101] or without personspecific calibration [75], [77], [79], [86], [92]. Few used annotation by trained observers [94], [103], [104], or did not (or not clearly) report on the origin of the used ground truth labels [82], [89], [90], [99], [100], [102], [105]. Jiang et al [83] segmented their tests of increasing pain intensity into different levels based on self-report, similar to the the calibration procedure used for the BioVid database [168].…”

“…Wang et al[102] and Yang et al[105] recognized pain from various physiological parameters obtained from hospitals' electronic flow sheets. Hand movement and finger pressure were exploited by Rivas et al[94] to detect pain and other states of stroke patients doing rehabilitation exercises.…”