Beyond Decodability: Do Acoustic Factors Drive Predictions in Speech-Based Alzheimer's Assessment?
Speech-based Alzheimer's assessment increasingly relies on self-supervised learning models, but their vulnerability to nuisance acoustic factors remains unclear. This paper asks whether noise and reverberation are merely encoded in SSL representations or actively drive diagnostic predictions. Using controlled input- and representation-space interventions on the ADReSSo dataset, the authors show that acoustic manipulations systematically shift classifier outputs even when no baseline group difference exists, challenging the assumption that predictive accuracy implies robustness.
The paper offers a methodologically careful contribution to trustworthy clinical speech AI. Its core argument—that decodability does not imply functional influence—is convincingly demonstrated through four complementary experiments, held-out test replication, and reversibility controls. The recommendation that intervention-based robustness tests become standard is well-supported, though the empirical scope is narrow and practical mitigation strategies are underexplored.
The experimental design is a clear strength: linear decoding (E1) establishes baseline encodability, while input-space (E2) and representation-space (E3) interventions isolate causal influence from mere correlation. The reversibility control—where flipping the steering coefficient to $\alpha=-1$ reverses logit effects ($r_{\mathrm{Pearson}}=-0.98$)—and the geometric alignment analysis (E4) provide strong mechanistic evidence that shifts are directionally structured rather than random noise.
External validity is limited by reliance on a single, small dataset—ADReSSo offers only $n=166$ train and $n=71$ test speakers—and just two acoustic factors. The authors note this and call for multilingual in-the-wild validation, but they do not calibrate whether the D1–D4 severities represent realistic clinical nuisances or synthetic extremes. In addition, while the geometric analysis shows that sensitivity is distributed across layers and streams, the paper does not propose or test concrete debiasing strategies, leaving the practical path to robustness unclear.
The evidence strongly supports the central claim that acoustic interventions causally influence predictions, replicating across three SSL backbones and a held-out test set. The extension beyond prior non-speech-only findings is valuable: whereas Gauder et al. showed that AD can be predicted from non-speech segments, this work demonstrates that acoustic manipulations shift predictions even when genuine speech content is present. However, quantitative comparisons to prior observational Clever-Hans detection methods are absent, making it difficult to assess whether the intervention framework catches subtler biases that correlation-based audits miss.
Reproducibility is moderately strong. The authors release code and use the public ADReSSo dataset with a fixed speaker-disjoint split. Methodological details—5-fold nested cross-validation, feature standardization, and regularization tuning—are clearly reported. Yet exact random seeds, full hyperparameter grids, and computational costs are omitted, and the VAD-based stream construction relies on pyannote.audio v3.4 without fully specified configuration files, which could block exact independent reproduction.
Speech-based Alzheimer's disease (AD) assessments increasingly rely on pretrained self-supervised learning (SSL) models that learn acoustic representations directly from raw audio, exposing the model to recording factors. We ask whether such factors are merely encoded in SSL representations or can systematically alter predictions. Using ADReSSo and three large SSL backbones, we apply controlled noise and reverberation interventions to participant-speech-only, non-speech, and full-recording audio. We combine layer-wise linear decoding, input- and representation-space interventions, and geometric alignment analysis to distinguish acoustic decodability from influence on AD prediction. Our results show that controlled acoustic interventions alter AD predictions across all three SSL backbones. Noise, despite showing no significant diagnostic-group difference in the original data, produces the strongest intervention effects. Importantly, these effects are systematically structured relative to the classifier's decision direction, replicate on the held-out test set and reverse when the representation-space intervention direction is reversed. Together, these findings show that high predictive performance and the absence of a significant diagnostic-group difference in a measured acoustic factor are not sufficient for robustness. We argue that intervention-based robustness tests should become standard for trustworthy clinical speech models.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.