Are you aware that Alzheimer's disease (AD) patients exhibit distinct speech characteristics when compared to normal older adults? There have been ongoing efforts to utilize these unique speech features in speech-based AD detection. Numerous prior studies are dedicated to enhancing diagnostic algorithms for more precise diagnoses. However, is there any other way to improve detection models? For instance, what if we use speech tasks that accentuate the speech patterns of AD patients? 

To discover speech tasks that highlight the distinctive speech characteristics of AD patients, it was crucial to first comprehend how AD induces alterations in the acoustic features of speech. Previous research has shown that multiple factors can influence our acoustic features, not limited to AD alone. One such factor is cognitive load, which has been observed to bring about changes in our acoustic features when applied. What's particularly intriguing is that cognitively vulnerable populations, such as older adults, children, and individuals with neurodegenerative disorders, appear to be more susceptible to the effect of cognitive load. This leads to the question: Could it be that the speech characteristics of AD patients become more prominent under the increased cognitive load?

We designed three speech tasks with varying levels of cognitive load: Interview, Sentence Repetition, and Story Recall tasks. In the Interview task, participants were asked questions about their basic personal information, recently watched TV programs, and their current mood. For the Sentence Repetition task, participants were presented with sentences and asked to repeat them exactly as they heard them. In the Story Recall task, participants listened to a story and were then asked to recall as much of it as they could remember. The cognitive load was manipulated by increasing the length of to-be-remembered items.

We collected speech data from both AD patients and normal older adults. The collected speech was segmented at the utterance level and analyzed using a feature set called eGeMAPS, which measures aspects of voice such as pitch, intensity, and speed. We compared speech characteristics between the two groups. Consistent with prior research, AD patients and normal older adults exhibited distinct speech features. AD patients showed increased pitch variability, decreased loudness, monotonous loudness, and slower speech compared to normal older adults. What's intriguing is that these characteristics varied depending on the speech task used! The difference in speech characteristics between AD patients and normal older adults was most pronounced when the story recall task was employed.

Could these differences in speech tasks also impact the performance of our detection models? To explore this, we created three AD classification models and three cognitive severity prediction models, each trained on different sets of training data. When comparing the performance of these models, the one trained on Speech Recall task data stood out with a classification accuracy of 81.4% and a prediction error of 4.62, demonstrating the best results.

This study highlights the crucial role of selecting appropriate speech tasks when collecting data for the development of speech biomarkers. In the context of AD, it has been demonstrated that increased cognitive load enhances speech characteristics, leading to improved diagnostic accuracy. Speech production is a complex cognitive process, encompassing higher-level elements such as conceptualization, lexical selection, syntactic and phonological encoding, as well as lower-level processes like speech motor control for articulation. For AD patients with limited cognitive resources, the increased cognitive load induced by speech tasks intensifies its impact on speech production, resulting in observable alterations in acoustic features. These findings underscore the importance of thoroughly investigating the unique speech patterns associated with specific disorders, and understanding the underlying causes.

Original Source Article:

Bae, M., Seo, M. G., Ko, H., Ham, H., Kim, K. Y., & Lee, J. Y. (2023). The efficacy of memory load on speech-based detection of Alzheimer’s disease. Frontiers in Aging Neuroscience, 15, 1186786. (Link)