Location
https://www.kennesaw.edu/ccse/events/computing-showcase/sp26-cday-program.php
Document Type
Event
Start Date
22-4-2026 4:00 PM
Description
Automated dementia detection from speech offers a scalable approach to cognitive screening, but its reliability depends on rigorous experimental design. In this work, we reconstructed a Wav2Vec2-based dementia classification pipeline and identified critical methodological flaws, including speaker leakage, nondeterministic preprocessing, and invalid test partitions. We rebuilt the dataset using strict speaker-level separation, deterministic segmentation, and validation checks to ensure reproducibility. The corrected baseline achieved an accuracy of 44.74 percent and macro F1 score of 0.4439, reflecting a more realistic performance estimate than prior inflated results. This work establishes a scientifically valid foundation for evaluating augmentation strategies such as SpecAugment in future phases.
Included in
GRM-175-233 From Leakage to Reliability in Dementia Detection
https://www.kennesaw.edu/ccse/events/computing-showcase/sp26-cday-program.php
Automated dementia detection from speech offers a scalable approach to cognitive screening, but its reliability depends on rigorous experimental design. In this work, we reconstructed a Wav2Vec2-based dementia classification pipeline and identified critical methodological flaws, including speaker leakage, nondeterministic preprocessing, and invalid test partitions. We rebuilt the dataset using strict speaker-level separation, deterministic segmentation, and validation checks to ensure reproducibility. The corrected baseline achieved an accuracy of 44.74 percent and macro F1 score of 0.4439, reflecting a more realistic performance estimate than prior inflated results. This work establishes a scientifically valid foundation for evaluating augmentation strategies such as SpecAugment in future phases.