Contrasting the Association of episodic memory using traditional neuropsychological test and acoustic features with PET imaging markers

Preeti Sunderaraman, Chenglin Lyu, Christina B. Young, Cody Karjadi, Elizabeth C. Mormino, Lauren Sidelinger, Sherral Devine, Vijaya B. Kolachalama, Honghuang Lin, Rhoda AuAlzheimer 2024 · DOI

TL;DR

This study explores whether acoustic features extracted from speech can serve as a more sensitive biomarker for preclinical Alzheimer's Disease (AD) than traditional memory tests. Researchers found that specific speech-based acoustic features, particularly those related to Mel-frequency cepstral coefficients and audio spectral features from delayed recall, were significantly associated with tau PET imaging markers in individuals asymptomatic for dementia, suggesting their potential as an early, non-intrusive AD biomarker.

What Problem Does This Paper Address?

Detecting Alzheimer's Disease (AD) in its early, preclinical stages is crucial for timely intervention and management. Current preclinical markers include amyloid and tau accumulation detected via PET scans, and declines in verbal memory assessed through neuropsychological tests. However, PET scans are expensive and invasive, and traditional cognitive tests might not be sensitive enough to pick up subtle changes in the very early stages of AD. The authors propose that analyzing acoustic features from recorded speech during cognitive assessments could offer a more accessible, non-intrusive, and scalable alternative for monitoring preclinical AD. The core problem is to determine if these digital voice biomarkers are more strongly associated with AD pathology (amyloid and tau PET markers) than traditional verbal memory scores.

Methods

The researchers utilized data from 196 dementia-free participants in the Framingham Heart Study who had both speech recordings and PET imaging data for amyloid and tau. Speech samples were collected during the Wechsler's Memory Scale Logical Memory immediate (LMI) and delayed (LMD) recall tests. From these recordings, 65 acoustic features were extracted using the OpenSMILE toolkit.

The core analytical approach involved multiple regression analyses. They examined the association between:

  1. PET imaging markers (global amyloid composite score, and tau in five specific brain regions: entorhinal, amygdala, parahippocampal, inferiortemporal, and putamen) as outcomes.
  2. Traditional memory measures (LMI and LMD scores).
  3. Acoustic features derived from both LMI and LMD speech segments.

To account for potential confounding factors, age, sex, education level, and APOE status were included as covariates in all models. For models where tau was the outcome, global amyloid levels were also included as a covariate. Multiple comparisons were adjusted using the False Discovery Rate (FDR) method, with a significance threshold set at 0.5.

Datasets

The study used data from the Framingham Heart Study. This is an existing, well-established longitudinal cohort study that provided the participants, their speech recordings, and their PET imaging data (amyloid and tau). No new datasets were introduced or created for this study.

Key Insights and Ideas

  • Speech as a Sensitive Preclinical AD Indicator: The most striking insight is that acoustic features from speech, particularly from delayed recall, showed a significant association with tau pathology in the inferiortemporal region, even in individuals who were clinically asymptomatic for dementia. This suggests that subtle changes in speech patterns might precede or co-occur with early AD-related brain changes.
  • Digital Biomarkers Outperform Traditional Tests: The study highlights a potential advantage of digital voice analysis over traditional neuropsychological tests. While acoustic features showed significant associations with tau biomarkers, the traditional LMI and LMD scores did not. This implies that acoustic analysis might be more sensitive in detecting early AD-related changes.
  • Tau Pathology is Key: The associations were specifically found with tau biomarkers, particularly in the inferiortemporal region. This aligns with current understanding of AD pathology, where tau accumulation is a significant driver of cognitive decline.
  • Delayed Recall is More Informative: The significant acoustic features were derived from the delayed recall test (LMD), suggesting that the ability to retrieve information after a delay, and the way this retrieval is vocalized, might be more indicative of early AD pathology than immediate recall.

Key Results

  • After adjusting for multiple comparisons (FDR < 0.5), two acoustic features were significantly associated with inferiotemporal tau:
  • Mel-frequency cepstral coefficient (MFCC) derived from LMD speech: p = 0.01
    • Audio spectral features derived from LMD speech: p = 0.02
  • No traditional measures of LMI or LMD recall were significantly related to any AD PET biomarker (amyloid or tau).
  • The study controlled for age, sex, education, and APOE status, and adjusted tau models for global amyloid.

Limitations and Future Work

  • Study Population: The study was conducted on a group that was clinically asymptomatic for dementia/AD. The findings need to be validated in more diverse populations to assess the robustness of these digital voice features as biomarkers.
  • Generalizability: While the findings suggest increased sensitivity of digital voice features, future studies are needed to confirm their utility across different demographic groups and in individuals with varying degrees of cognitive impairment.
  • Specific Brain Regions: The significant associations were found with inferiotemporal tau. Further research could explore the relationship of acoustic features with tau and amyloid in other brain regions implicated in AD.
  • Causality: The study demonstrates associations, not causation. Future longitudinal studies could track how these acoustic changes evolve over time in relation to AD progression.

Summary updated

Abstract

Abstract Background Preclinical measures of Alzheimer’s disease (AD) risk include markers of amyloid and tau measured via PET scans and decline in verbal memory. Audio features from recorded speech collected during neuropsychological testing offer an alternative for monitoring preclinical AD symptoms that is nonintrusive, economical, and scalable but warrants further validation. This study aims to examine whether there is a stronger association between acoustic metrics and PET imaging biomarkers compared to a traditional verbal memory test. Methods This study involved 196 Framingham Heart Study dementia‐free participants with available acoustic, as well as PET amyloid and tau imaging data (refer to Table 1 for demographics). For each participant, 65 acoustic features were extracted from speech samples recorded during the Wechsler's Memory Scale Logical Memory immediate (LMI) and delayed (LMD) recall tests, utilizing OpenSMILE for feature extraction. Multiple regression analyses were performed with a global composite score for PET amyloid, or PET tau for five regions (entorhinal, amygdala, parahippocampal, inferiortemporal, putamen) that were examined individually as outcomes. LMI, LMD, acoustic features derived from LMI and from LMD recorded voice segments were related to AD PET biomarkers in each regression model. Age, sex, education level, and APOE status were used as covariates for all analysis. Models using tau as outcome were adjusted for global amyloid. Results After adjusting for multiple comparisons using FDR (significance set at 0.5), among the 65 acoustic features, the Mel‐frequency cepstral coefficient (p=0.01) and audio spectral features of the acoustic features derived from LMD (p=0.02) were most strongly associated with inferiotemporal tau region (see Table 2). No traditional measure of LMI or LMD was significantly related to any AD PET biomarker. Conclusions In a group that was clinically asymptomatic for dementia/AD, only digital measures were associated with AD tau biomarker. These findings suggest that digital voice features have increased sensitivity to detect those at AD risk compared to traditional cognitive tests and have potential to serve as a biomarker of AD. Future studies in diverse populations are needed to test the robustness of these findings.
BibTeX
@inproceedings{sunderaraman2024contrasting,
  author = {Preeti Sunderaraman and Chenglin Lyu and Christina B. Young and Cody Karjadi and Elizabeth C. Mormino and Lauren Sidelinger and Sherral Devine and Vijaya B. Kolachalama and Honghuang Lin and Rhoda Au},
  title = {{Contrasting the Association of episodic memory using traditional neuropsychological test and acoustic features with PET imaging markers}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.091608},
}
Figure