Vocal Biomarkers for Dementia Patient Monitoring

Patient monitoring systems are becoming important in patient care and can provide useful feedback for clinicians and caregivers on health-related issues. These systems can also be beneficial for the patients themselves and improve their quality of life and care. In an ongoing European project, called Dem@Care, we are dealing with assessment of dementia signs in monitored subjects using multiple sensing devices including wearable microphones, static and wearable cameras and some physiological sensors. In one use case, the patient is monitored at home in natural settings. In another use case, the system is used in a lab during clinical assessment while the patient is asked to perform certain cognitive and physical exercises. Working on the audio analysis component, we are looking for vocal biomarkers of Dementia, focusing on early detection and monitoring of the patients’ conditions. We base our work on recordings that took place recently in the Alzheimer Medical Center at Thessaloniki in Greece. 27 early-stage Alzheimer patients, 43 Mild-Cognitive-Impairment patients and 19 normal control subjects were recorded, while performing several tasks. The tasks included free speech, repeating spoken sentences and Diadochokinetic test. We demonstrate vocal biomarker extraction and automatic classification from the recorded data. The vocal biomarkers are based on prosodic, content similarity and speech quality features. They are used with SVM (Support Vector Machine) classifier to demonstrate classification between the different groups. The classification error is shown to be below 20%.