Automated Alzheimer's detection from voice alone is still weak. On the standard ADReSS benchmark, the best acoustic baseline reaches only 62.5% test accuracy, while text transcripts reach 75%. Every acoustic system in that literature describes a speaker with one long list of numbers, such as averages and variances of pitch, energy, and spectral shape, and treats each number as a separate signal. Speaking, however, is a coordinated motor act: pitch, loudness, and vocal-tract shape normally rise and fall together in characteristic ways. Work on depression shows that illness changes these co-movements, and that a classifier given the correlations between features outperforms one given the features themselves. We propose to test whether Alzheimer's disease leaves a similar trace in speech, a loss or rewiring of how acoustic features co-vary. We also propose to identify the specific pairs of features responsible, in a way that holds up on the small cohorts this field has.
The obstacle is counting. With 25 frame-level acoustic descriptors there are 300 feature pairs, and ADReSS has 108 training speakers. A prior attempt to feed raw pairwise covariances to a regressor overfit badly. We borrow a tool from genomics, ensemble feature selection, which keeps only features chosen by several structurally different selectors. We repair two of its known flaws, its arbitrary top-K cutoffs and its tendency to discard groups of correlated variables. The deliverable is a small, statistically calibrated set of "coordination edges" that separate Alzheimer's from control speech. The work tests whether this set adds diagnostic signal beyond marginal statistics and beyond a dense, well-regularized use of all pairs.
Core Claim
Our claim is that Alzheimer's speech differs from healthy speech in its conditional dependencies between acoustic descriptors, not just in their levels. A conditional dependency is how strongly two features move together once all the other features are accounted for. We further claim that a resampling-based consensus selection over these dependencies can recover a compact, reproducible set of them, judged against a null threshold obtained by shuffling the diagnosis labels. The claim is falsifiable in three ways. Destroying cross-feature structure while keeping each feature's distribution intact should erase the gain. A marginal-only model given the same number of inputs should do worse. The selected edge set should recur across resamples far more often than edges chosen under shuffled labels.
The scientific claim is the primary one. The accuracy claim is secondary. With 48 test speakers, one speaker equals about 2.1 accuracy points, and we will not oversell small differences.
Relation to Prior Work
Two clinical speech papers motivate the hypothesis. Modeling spectral variability for the classification of depressed speech used covariance-type descriptors of spectral variability to classify depression. The Relationship Between Speech Features Changes When You Get Depressed computed windowed correlation matrices over 32 OpenSMILE features and cut error rates by 26.6% for SVMs and 23.1% for LSTMs, reaching 88.0% accuracy on the Androids corpus. Both concern depression and read speech. Neither asks which pairs matter, and neither touches neurodegeneration. We ask both questions, on spontaneous picture-description speech.
A third paper constrains the design with a negative result. Speaker Sincerity Detection based on Covariance Feature Vectors and Ensemble Methods vectorized covariance matrices and overfit. Its development Spearman correlation was 0.562, but its test score of 0.547 fell below the 0.602 baseline, and PCA did not rescue it. Its lesson is that dumping all pairs into a model fails. Its gap is that it never applied supervised, multi-selector pruning.
Two papers supply the estimation method. Sparse Gaussian graphical models for speech recognition fit L1-penalized precision matrices, the inverse covariance whose zero entries mark conditional independence. Its sparse models beat full covariance on TIMIT, 70.3% against 69.6% phone accuracy, but performance depended on a narrowly tuned penalty. Learning Multiple Tasks with a Sparse Matrix-Normal Penalty showed that sparse precision penalties can act as structured feature selection. Neither was applied to clinical speech. We reuse the sparse-precision estimator per speaker, where data are plentiful, and move the hard selection problem to the across-speaker level.
The selection recipe comes from genomics, which is also where its flaws are documented. Identification of clinical diagnostic and immune cell infiltration characteristics of acute myocardial infarction with machine learning approach and Identification and validation of immune-related and inflammation-related genes in endometriosis intersect three selectors: LASSO, a sparse linear regression; SVM-RFE, which repeatedly drops features that contribute least to a margin classifier; and Random Forest or Boruta, tree-based importance tested against shuffled copies. Bioinformatics and experimental validation identify biomarkers for diagnosing Alzheimer's disease adds XGBoost to obtain four Alzheimer's hub genes. That paper's downstream serum marker reached only 51.2% specificity, a reminder that consensus selection does not guarantee disease specificity. These pipelines treat features as independent columns, use heuristic cutoffs, and silently break up collinear blocks. Pairwise covariance elements form exactly such blocks.
Learning Covariance-Based Multi-Scale Representation of Neuroimaging Measures for Alzheimer Classification is the closest Alzheimer's precedent. It used the covariance across brain regions to build learned filters and reached 85.8% accuracy for normal cognition versus early mild cognitive impairment on ADNI. It works on static imaging and population-level covariance. We work on within-speaker dynamic coordination and seek an explicit, selected edge set rather than a learned spectral filter.
Research Program
The first step builds a coordination profile for each speaker. For each ADReSS speaker we extract frame-level eGeMAPS low-level descriptors, about 25 per frame, separately for voiced and unvoiced regions. Pooling the speaker's segments gives thousands of frames. From these we estimate a sparse precision matrix $\Theta_s$ by graphical lasso, choosing the penalty per speaker by held-out frame likelihood rather than by hand. This addresses the tuning fragility noted by Bell and King. The partial correlation between descriptors $i$ and $j$ for speaker $s$ is
$$\rho^{(s)}_{ij} = -\frac{\Theta_{s,ij}}{\sqrt{\Theta_{s,ii}\,\Theta_{s,jj}}},$$
which measures direct co-movement after removing the influence of all other descriptors. We Fisher-transform each $\rho^{(s)}_{ij}$ to stabilize its variance. Plain correlations are kept as a comparison arm.
The second step groups edges before selecting them. Across training speakers, many edges move together, for example all edges involving one MFCC band. We cluster edges by their across-speaker correlation and represent each cluster by its first principal score. Selection then operates on clusters, and individual edges are reported within the clusters that survive. This targets the collinearity failure of flat ensemble selection, where LASSO picks one arbitrary member of a group and Random Forest spreads importance across the group.
The third step is calibrated consensus. On each of $B$ bootstrap resamples of the training speakers we run LASSO, SVM-RFE, and Boruta. We record $\pi_c$, the fraction of the $3B$ runs in which edge cluster $c$ is selected. We then repeat the whole procedure under $R$ label permutations to get the null distribution of the maximum selection frequency, $\max_c \pi_c^{\text{null}}$. A cluster is retained if
$$\pi_c > q_{0.95}\left(\max_c \pi_c^{\text{null}}\right),$$
where $q_{0.95}$ is the 95th percentile across permutations. This replaces arbitrary top-K intersections with a family-wise threshold. All selection runs strictly inside each outer cross-validation fold, because Tackling the ADReSS Challenge fell from 92.7% cross-validation accuracy to 77.8% on test, consistent with feature choices leaking into evaluation.
The final step is classification and regression. Retained clusters, optionally joined by standard eGeMAPS marginal statistics, feed a regularized logistic regression for the Alzheimer's-versus-control decision and a ridge regression for Mini-Mental State Examination (MMSE) scores.
Baselines
The naive reference point is the ADReSS challenge's own acoustic system. Alzheimer's Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge reports 0.625 test accuracy with linear discriminant analysis on ComParE features, and a test RMSE of 6.14 for acoustic MMSE regression with decision trees on MRCG features. These are the strongest acoustic-only test results reported in the file on this benchmark. The INESC-ID Multi-Modal System for the ADReSS 2020 Challenge reports a weaker 54.17% test accuracy for x-vectors alone. We rerun the challenge baseline under our evaluation protocol and also quote the published figures. Beating it means exceeding 0.625 on the official test split, with a paired test on the same 48 speakers that indicates the gain is not one or two speakers.
The strongest competitor on the actual question is a dense, regularized use of all pairs. Here every Fisher-transformed correlation and partial correlation goes into an L2-regularized logistic regression, with the penalty tuned by nested cross-validation. This is the obvious port of the depression approach in The Relationship Between Speech Features Changes When You Get Depressed, with its SVM swapped for a heavily regularized linear model to guard against the Senoussaoui failure. No reported result exists on Lacuna for this baseline on ADReSS. It receives the same frames, folds, and hyperparameter search budget as our method.
Two ablations isolate the key ingredients. The first is single-selector stability selection: LASSO alone on bootstrap resamples, with the same permutation threshold. It tests whether combining three different selectors adds anything beyond resampling. The second is ungrouped consensus, the same pipeline without edge clustering, which tests the collinearity fix. A matched-marginal control gives a model the same number of inputs drawn from marginal statistics such as means, variances, and percentiles. Finally, a surrogate control repeats everything after independently circularly shifting each descriptor's frame sequence within a speaker. This preserves each feature's distribution and temporal autocorrelation but destroys cross-feature alignment. If the surrogate performs as well as the real data, coordination is not the source of the signal.
Why not just…?
Why not just feed the full correlation matrix to a classifier, as the depression work did? This is the most direct alternative and could plausibly match us on accuracy. The dense L2 baseline above is that experiment. If it ties our method on test accuracy and RMSE, we will report that sparse selection buys no accuracy. The contribution then narrows to the calibrated, reproducible edge set. We measure that set's reproducibility directly by overlap across bootstrap resamples and across the two independent fits described below. A dense model cannot deliver an interpretable motor-coordination signature, but we will not claim it is less accurate unless the data show it.
Why not just reduce the pairs with PCA or covariance shrinkage? In Speaker Sincerity Detection based on Covariance Feature Vectors and Ensemble Methods, PCA on vectorized covariances did not prevent test collapse. PCA is unsupervised, so it keeps directions of large variance across speakers, which may reflect channel or speaker identity rather than diagnosis. We include PCA followed by logistic regression, with the number of components tuned by nested cross-validation, as an additional arm so the comparison is explicit rather than assumed.
Why not just use pretrained deep speech embeddings? On ADReSS, x-vectors reached 72.73% on development but 54.17% on test, according to The INESC-ID Multi-Modal System for the ADReSS 2020 Challenge. That gap suggests out-of-domain speaker embeddings overfit or miss Alzheimer-relevant cues. No reported ADReSS result exists on Lacuna for newer self-supervised encoders. We test the two approaches as complements, not rivals: coordination edges computed over the channels of such embeddings form an extension arm, and we check whether edges add to pooled embeddings when combined.
Why not just use a single stability-selection run instead of an ensemble of selectors? This is a genuine threat. A single well-calibrated selector may be just as stable. The LASSO-only ablation settles it. If it matches the three-selector consensus in both accuracy and edge reproducibility, we will drop the multi-selector component and say so.
Proposed Experiments and Decisive Checks
The main experiment uses the ADReSS train/test partition: 108 training speakers and 48 test speakers, age- and gender-matched. We also report leave-one-subject-out cross-validation on the training set. Metrics are accuracy, F1, and AUC for classification, and RMSE for MMSE regression. Each comes with bootstrap confidence intervals and paired tests on the 48 test speakers.
The first decisive check is the surrogate test. The real-data edge model must beat its circularly shifted surrogate, otherwise the coordination hypothesis fails for Alzheimer's speech. The second is the matched-marginal comparison. Edges must add to, or outperform, an equal number of marginal statistics. The third is calibration. Under label permutation, the full pipeline must select no clusters at the stated rate, and test accuracy must fall to chance. This guards against the leakage pattern visible in the Syed et al. results.
For a reproducibility check beyond ADReSS's small test set, we will identify DementiaBank Pitt speakers who do not overlap with the ADReSS sample, if such speakers can be cleanly separated. We will refit the consensus on them independently and measure overlap with the ADReSS edge clusters. Because ADReSS was drawn from Pitt, this overlap measures stability, not true external validity, and we will report it as such.
The specificity check addresses the lesson of the c-Myc result, where a marker sensitive to disease was not specific to it. We apply the same pipeline to the Androids depression corpus and compare the selected clusters. Shared edges would point to generic psychomotor slowing. Edges unique to Alzheimer's would be candidates for disease-specific signal. Language and task differ between the corpora (Italian read speech versus English picture description), which confounds this comparison, so the result is hypothesis-generating rather than conclusive.
Partial-correlation and plain-correlation arms are compared throughout, which tests whether conditional dependencies, not just pairwise correlations, carry the signal. A clear positive outcome would require four things together: test accuracy above 0.625, a win over the surrogate and the marginal-matched control, a selected cluster set recurring in well over the permutation threshold of resamples, and accuracy at least on par with the dense L2 competitor.
References
- Modeling spectral variability for the classification of depressed speech
- The Relationship Between Speech Features Changes When You Get Depressed: Feature Correlations for Improving Speed and Performance of Depression Detection
- Speaker Sincerity Detection based on Covariance Feature Vectors and Ensemble Methods
- Sparse Gaussian graphical models for speech recognition
- Learning Multiple Tasks with a Sparse Matrix-Normal Penalty
- Identification of clinical diagnostic and immune cell infiltration characteristics of acute myocardial infarction with machine learning approach
- Identification and validation of immune-related and inflammation-related genes in endometriosis
- Bioinformatics and experimental validation identify biomarkers for diagnosing Alzheimer’s disease
- Learning Covariance-Based Multi-Scale Representation of Neuroimaging Measures for Alzheimer Classification
- Tackling the ADReSS Challenge: A Multimodal Approach to the Automated Recognition of Alzheimer’s Dementia
- Alzheimer’s Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge
- The INESC-ID Multi-Modal System for the ADReSS 2020 Challenge