A Novel Pain Level Assessment Approach Using Multi Modal Bayesian Network Modeling

Effective pain assessment is essential for the treatment and care of patients in complicated and uncertain clinical surroundings. In order to compensate for the low accuracy of single-mode pain level assessment, a pain recognition method based on multi modal Bayesian network (MMBN) was proposed. In order to simplify the model, mutual information is used to judge the correlation of multi modal pain features and redundant feature vectors are eliminated. Then, a BN structure for pain recognition is proposed. To address the issue of small available data for modeling in many real-world applications, a Bayesian network parameter learning algorithm using Maximum A Posteriori and Convex Optimization (MACV) is also proposed. Based on convex optimization, the informative qualitative prior knowledge is transformed into constraints to generate a class of candidate parameter sets. Then BN parameters are estimated by fusing the candidate parameter sets with the one from samples. Finally, BN reasoning method is used to achieve the final assessment of pain level. Experimental results show that the learned accuracy of MACV is superior to MLE or QMAP algorithm which is the state-of-art. And MMBN provides higher recognition accuracy than unimodal methods, and can effectively improve the assessment performance for pain level.