Introduction
•The Need for Biomarkers: Aggregation of TDP-43 in limbic structures causes amnestic cognitive impairment in the older population.
• There are currently no established biomarkers for detecting limbic predominant age-related TDP-43 encephalopathy (LATE).
•Acoustic Biomarkers From Speech: We studied the utility of acoustic features from spontaneous speech for the detection of LATE neuropathological changes. We used machine learning models.
Methods
Participants from The 90+ Study:
•Participants were video-recorded during clinical interviews to collect spontaneous speech samples.
•Audio from baseline visit samples were extracted and processed using Audacity to reduce background noise and remove interviewer's speech.
Acoustic Features:
•Mel-frequency cepstral coefficients (MFCC), spectral centroid (SC), zero-crossing rate (ZCR), root mean square (RMS), and chroma.
•These features were extracted and scaled from the audio files to analyze speech brightness, quality, energy, and pattern changes.
Machine Learning Models:
•Logistic Regression, Support Vector Classifier (SVC; part of Support Vector Machine (SVM)), K-Nearest Neighbor (KNN), Extra Trees, and Gradient Boosting.
•The acoustic features were analyzed for detection of TDP-43.
•These models were selected for performance comparison and trained on 80% of the dataset. Each model's performance was evaluated using a test data consisting of unseen participants who were not included in the training set (Figure 1).
Results
Participant Demographics (Table 1):
•112 participants
•81 participants had no TDP-43 at autopsy (LATE-NC stage 0-1)
•31 had TDP-43 present (LATE-NC stage 2-3)
•73 participants (65.1%) were female
•Mean age of 97.64 years at death
Performance Results (Table 2):
•Logistic Regression showed the best detection results:
•F1 score = 71.4%
•Accuracy score = 82.6%
•AUC = 75.8%
•Followed by KNN and Gradient Boosting
Conclusion
•Analysis of acoustic features of speech holds promise as a novel and noninvasive approach for the detection of LATE.
•Availability of a reliable biomarker for LATE could substantially improve the clinical management by offering clinicians an opportunity to diagnose the disease in its initial stages as opposed to post-mortem detection.
•Integrating both acoustic and linguistic features of speech and using a larger sample size might further enhance the accuracy of detection models.
•Speech biomarkers could also be useful for detection of Alzheimer's Disease Neuropathologic Change (ADNC).
Aknowledgment
This work was supported in part by the National Institutes of Health (R01 AG062706), R01 AG021055, the UCI Institute for Memory Impairments and Neurological Disorders (IRPA-01-2024), the Institute for Precision Health at UCI, UCI ADRC (P30 AG066519), and NIH TITAN T32 Award at UCI MIND.
