Factor analysis of acoustic-prosodic features as an approach to predicting risk-factors for dementia in a healthy population

Background: Language is a valuable source of clinical information in Alzheimer’s Dementia, as itdeclines concurrently with neurodegeneration. While speech and language data have been extensivelystudied in connection with diagnosis, there is increasing interest in assessing their potential for oftracking of cognitive trajectories in preclinical stages of Alzheimer’s.Method: We collected conversational speech data from 43 participants who were assigned to a low-riskgroup (n=23) and a high-risk group (n=20) depending on whether they had a family history of dementia.Their speech was collected through a novel methodology for monitoring cognitive health, Prevent-ED[1]. Data pre-processing consisted of acoustic enhancement (i.e. removing background noise,normalising speech volume and segmenting conversational turns) and feature generation (i.e. extractingthe Geneva Minimalistic Acoustic Parameter Set, eGeMAPS [2]). Subsequent data analysis included afactor analysis to reduce the extracted feature set (88 variables), and a statistical significance test(Student’s t test) to compare each of the resulting factors between the low and high-risk groups.Results: The factor analysis yielded five factors, namely pitch, voice intensity, spectral energy, voicequality and spectral dynamics (see Table 1 for details). None of these factors were significantly differentbetween the low and high-risk groups. Nonetheless, this factorial structure explains 35.46% of theoverall variance in our dataset.Conclusion: Despite the lack of statistical significance, a 35% of explained variance is a favourableresult which is likely to improve in future research, when the risk categories include other details asideform family history (i.e. genetics, CSF markers). Therefore, our analysis suggests that prosodyconstitutes a promising avenue for finding distinctive markers of dementia risk. Besides, this is acontent-independent approach which does not rely transcriptions, ensuring participant’s privacy andallowing for the method to be generalised to other languages.