Statistical and clinical utility of multimodal dialogue-based speech and facial metrics for Parkinson's disease assessment

We present a framework for characterising the statistical and clinical relevance of speech and facial metrics in Parkinson's disease (PD) extracted by a multimodal conversational platform. 38 people with PD (pPD) and 22 controls were recruited in an ongoing study and were asked to complete four interactive sessions, a week apart from each other. In each session, a virtual conversational agent, Tina, guided participants through a battery of standard tasks designed to elicit speech and facial behaviours. Speech and facial metrics were automatically extracted in real time, several of which showed statistically significant differences between pPD and controls. We explored which of these differences were greater than measurement error, a threshold defined as the minimally detectable change (MDC). Furthermore, we computed the minimal clinically important difference (MCID) with respect to the Communicative Participation Item Bank short form (CPIB-S) scale for these select metrics. Our results show that differences in metrics like duration and fundamental frequency (F0) of speech are captured beyond measurement error. We also discuss several confounding factors that need to be taken into consideration before making any clinical interpretation of changes in these metrics.