Video-based Pain Level Assessment: Feature Selection and Inter-Subject Variability Modeling

Abstract

Automatic pain level assessment, based on video features, may provide clinically-relevant, objective measures of pain intensity. In various clinical contexts accurate pain level estimation by health care personnel is challenging. This problem is compounded by considerable inter- and intra-individual variability of both perceived pain levels and of the associated facial expressions, especially at low pain levels. Thus, providing objective video-based indices for pain level assessment is a rather computationally challenging problem. In the present work both geometric and color-based features were extracted. The most informative features were identified with lasso regression, and subject variability was modeled through a generalized linear mixed effects probit model. Video recordings from the Biovid Heat Pain Database were used with the proposed methodology, aiming to classify video samples to five levels of pain. Performance of the proposed model was comparable to the state-of-the-art random forests algorithm despite its relative simplicity and more conservative cross-validation approach adopted.