Support Vector Machines to Define and Detect Agitation Transition

Abstract

The need to automate the detection of agitation and the detection of agitation transition for dementia patients is a significant facilitator for caregivers. This research aims at detecting the transitional phase toward agitation, as well as agitation detection of subjects, using soft computing techniques that do not require supervision beyond the training phase. Three vital signs are monitored: Heart Rate (HR), Galvanic Skin Response (GSR), and Skin Temperature (ST). These measures are fed into two proposed SVM architectures which are based on the definition of a new confidence measure: “Confidence-Based SVM” and “Confidence-Based Multilevel SVM.” Results show very high detection accuracy of agitation and agitation transition, a quick adaptation to the subject, and a strong correlation between the physiological signals monitored and the emotional states of the subjects. Another challenge that is successfully addressed in this paper is the ability to train the classifier on a limited group of subjects, and then test it on subjects not belonging to the training group. The result is a learning algorithm that is “Subject-Independent.”