Speech is a rich source of clinical information when assessing and monitoring Alzheimer’s Disease. Natural language processing technologies, like those developed by Winterlight Labs, can objectively quantify speech and enable its use as a novel endpoint in drug development.

Using speech from clinical interviews collected in the TAURIEL Phase 2 study, Genentech collaborated with Winterlight to identify aspects of speech and language that showed progressive change with prodromal to mild Alzheimer’s disease over an 18-month period. Results of this work were recently published in Alzheimer’s and Dementia: Diagnosis Assessment & Disease Monitoring.

To create this composite, the team took samples of participant speech collected passively during the Clinical Dementia Rating (CDR) Interview. By focusing in on the portion of the interview which discussed a recent experience, the team was able to collect ~2-4 minutes of speech per participant. This speech was then processed using Winterlight’s feature extraction platform and then a detailed, data driven selection process was used to select which elements of speech to include in the composite.

FIGURE 1
Longitudinal change in speech features. Mean change from baseline at each assessment for selected speech features with significant linear change over the course of the study. Error bars represent 95% confidence intervals.

Overall the composite had a similar trajectory and effect size to the existing gold standards used in AD research, including the CDR Sum of Boxes, which was the primary endpoint in this trial. Leveraging speech that is already being collected as part of the trial process represents a low burden way of gathering additional treatment efficacy data. Future work will focus on validating this composite in other datasets and optimizing it so it can be leveraged in the clinical drug development process.

FIGURE 2
Speech composite score. Mean change from baseline at each assessment for the speech composite score compared with clinical assessment total scores.