Background: Effective treatments for Alzheimer’s disease (AD) remain elusive and the heterogeneity of the disease’s behavioral symptoms in its earliest stages make it difficult to detect those on a neurodegenerative trajectory with a high degree of accuracy. Beginning in 2005, digital audio recordings of spoken responses to neuropsychological tests administered to the Offspring cohort of the Framingham Heart Study were collected. We leveraged this unique database of audio recordings to test the feasibility of identifying voice biomarkers of dementia/AD.
Methods: From a pool of 7000+ recordings obtained between 2005-2016, we selected a subset that included those diagnosed with dementia by consensus review (n=107) and a control group that was not demented (n=35). The audio files were diarized (separating tester’s speech from that of the subject) and transcribed semi-automatically, and a set of acoustic and language-based features were computed on the annotated text. We posed the machine learning task as a classification problem and trained a random forest classifier using mean area under the receiver operating curve (AUC) scores across a 10-fold cross validation scheme as the performance metric.
Results: Using only context-agnostic acoustic features the model obtains an average AUC of 0.76. Adding natural language processing (NLP) based features boosts the value to 0.91.
Conclusions: While these results are preliminary, simple acoustic and language features computed over speech segments show promise for the development of accurate digital biomarkers of cognitive impairment. Future work will involve a bigger sample size and aim to determine whether these biomarkers can serve as a low-cost, scalable screening tool to accurately identify individuals-at-risk among the general population.