Multi-Label and Multimodal Classifier for Affective States Recognition in Virtual Rehabilitation

Abstract

Computational systems that process multiple affective states may benefit from explicitly considering the interaction between the states to enhance their recognition performance. This work proposes the combination of a multi-label classifier, Circular Classifier Chain (CCC), with a multimodal classifier, Fusion using a Semi-Naive Bayesian classifier (FSNBC), to include explicitly the dependencies between multiple affective states during the automatic recognition process. This combination of classifiers is applied to a virtual rehabilitation context of post-stroke patients. We collected data from post-stroke patients, which include finger pressure, hand movements, and facial expressions during ten longitudinal sessions. Videos of the sessions were labelled by clinicians to recognize four states: tiredness, anxiety, pain, and engagement. Each state was modelled by the FSNBC receiving the information of finger pressure, hand movements, and facial expressions. The four FSNBCs were linked in the CCC to exploit the dependency relationships between the states. The convergence of CCC was reached by 5 iterations at most for all the patients. Results (ROC AUC) of CCC with the FSNBC are over 0.940 ± 0.045 (mean ± std. deviation) for the four states. Relationships of mutual exclusion between engagement and all the other states and co-occurrences between pain and anxiety were detected and discussed.

Affect-driven Ordinal Engagement Measurement from Video

“…The person-in-context engagement can be measured by observations of person's interactions with a particular contextual environment, e.g., reading a web page or participating in an online classroom. Except for the person-oriented engagement, for measuring the two other parts of the continuum, knowledge about context is required, e.g., the lesson being taught to the students in a virtual classroom [27], or the information about rehabilitation consulting material being given to a patient and the severity of disease of the patient in a virtual rehabilitation session [10], or audio and the topic of conversations in an audio-video communication [8].…”

“…Over the recent years, extensive research efforts have been devoted to automate engagement measurement [11], [12]. Different modalities of data are used to measure the level of engagement [12], including RGB video [15], [17], audio [8], ECG data [9], pressure sensor data [10], heart rate data [25], and user-input data [12], [25]. In this section, we study previous works on RGB video-based engagement measurement.…”

“…Different from the end-to-end approaches, in feature-based approaches, first, multi-modal handcrafted features are extracted from videos, and then the features are fed to a classifier or regressor to output engagement [6], [7], [8], [10], [11], [12], [16], [17], [38], [39], [40], [41], [42], [43], [44], [45]. Table 1 summarizes the literature of feature-based video engagement measurement approaches focusing on their features, machine-learning models, and datasets.…”

“…Various modalities have been used for automatic engagement measurement in virtual settings, including person's image [5], [6], video [7], audio [8], Electrocardiogram (ECG) [9], and pressure sensors [10]. Video cameras/webcams are mostly used in virtual programs; thus, they have been extensively used in assessing engagement in these programs.…”

Chronic Pain Protective Behavior Detection with Deep Learning

“…Beyond the greater challenge of detecting anomalous movement behaviors (compared to the recognition of physical activity types) in data from real patients, such area also faces the difficulty of obtaining large volume of training data for the positive class(es) (e.g. [43], leading to considerable skew in the datasets that exist, and also constraining the use of deep neural network models. Although LSTM networks show a lot of promise based on our review, care must be taken on how the input data is formatted, particularly in the approach taken to segment the data along the temporal dimension.…”