Detecting Dementia from Face in Human-Agent Interaction

This paper proposes an approach to automatically detect dementia from a human face. Although some works have detected dementia from speech and language attributes, there are few studies focusing on facial expression in dementia patients. We recorded the human-agent interaction data of spoken dialogues from 24 participants (12 with dementia and 12 without) and extracted the face features. Our objective was to classify dementia by L1 regularized logistic regression. The facial features and the L1 logistic regression were then used to classify the participants into two groups with 0.82 detection performance, as measured by the areas under the receiver operating characteristic curve. We also identified various contributing features, such as action units, eye gaze, and lip activity. These results demonstrate that our system has the potential to detect dementia from the face through spoken dialog systems and as such, can be of assistance to health care workers.