Longitudinal vs Cross-Sectional Speech Biomarkers in Dementia

Peter S. Pressman¹, Francesca Dino², Chelsea Chandler³, Kevin Bretonnel Cohen⁴, Steven Bedrick⁵, Elliott Ross², Peter Foltz³, Lawrence E. Hunter⁶

¹Layton Aging and Alzheimer's Disease Research Center, Department of Neurology, Oregon Health & Science University, Portland, OR; ²Department of Neurology, University of Colorado Anschutz Medical Campus, Aurora, CO; ³Institute of Cognitive Science, University of Colorado Boulder, Boulder, CO; ⁴Third Battalion, Ukrainian Volunteer Army, Ukraine; ⁵Medical Informatics and Clinical Epidemiology, Oregon Health & Science University, Portland, OR; ⁶Department of Pediatrics, University of Chicago, Chicago, IL

Introduction

• Speech reflects cognition; dementia changes speech.
• Most speech-biomarker studies have been single-visit, patient versus control.
• But Alzheimer's disease and related disorders are diseases of trajectory.
• Q1: Does rate of change add diagnostic signal beyond one visit?
• Q2: Do linguistic and acoustic features differ in approximation of diagnosis versus progression?

Methods

• Corpus: DementiaBank Pitt Corpus, Cookie Theft; Control vs probable AD, two or more visits.
• Features: 95 linguistic (lexical diversity, psycholinguistic norms, syntax); 528 acoustic (eGeMAPS via OpenSMILE).
• Design: per-individual slopes; three analyses under nested 5-fold CV, first visit, last visit, trajectory.
• Selection: top features by importance within each training fold; five classifiers tested, highest-AUC reported.
• Cohort: 149 with two or more visits, 119 Control or AD; matched to 82 to 86 per analysis, 32 acoustic last visit.
• AUC with 1000-iteration bootstrap 95% CIs.

Results
Linguistic features classified well at every visit; acoustic separation was lower and trended upward, but not significantly.

Modality | Analysis | AUC [95% CI] | Best model | Matched N

• Linguistic (95 NLP) — First visit: 0.751 [0.656, 0.854], SVM, 86 · Longitudinal slopes: 0.766 [0.652, 0.870], SVM, 84 · Last visit: 0.819 [0.720, 0.904], SVM, 86
• Acoustic (528 eGeMAPS) — First visit: 0.534 [0.410, 0.666], AdaBoost, 86 · Longitudinal slopes: 0.618 [0.490, 0.739], SVM, 82 · Last visit: 0.719 [0.524, 0.889], LogReg, 32*
• Multimodal (623) — Longitudinal slopes: 0.626 [0.501, 0.754], SVM, 82

Group | N | Age±SD | Female | Visits | Span

• Control: 60 | 62.7±7.0 | 63% | 3.0 | 2.3
• ProbableAD: 59 | 70.3±8.5 | 71% | 2.4 | 1.6

Linguistic slope features | Acoustic slope features (↑/↓ = direction of change over time in probable AD)

• Type-token ratio variability ↑ | Alpha-ratio variability ↓
• Sentiment (minimum) ↑ | Spectral flux variability ↓
• Pronoun frequency ↓ | Jitter (local) ↑
• Word meaningfulness ↓ | F2 amplitude ↑
• Content density ↑ | MFCC1 variability ↑

Discussion

• A good diagnostic marker is not always the best progression marker.
• Linguistic features index accumulated vocabulary and syntax damage: strong at baseline, largely saturated (0.751 to 0.766).
• Acoustic features weaker at baseline; point estimates higher at later stages (0.534 to 0.719) but not significant, may suit monitoring.
• Multimodal fusion did not beat the best single modality.
• Cookie Theft is language-loaded, favoring linguistic content.

Limitations

• One single-site, English corpus; clinical probable AD, no biomarker confirmation.
• Differences descriptive, CIs overlap; only linguistic over acoustic at baseline is significant.
• Best-of-five reporting optimistic; wide CIs, especially acoustic (528 features).
• Acoustic baseline and slopes straddle chance; last-visit cell n = 32.
• Short follow-up, shorter in AD; groups differ in age, so aging and disease are confounded.

Future directions

• Test the modality-by-timepoint interaction directly.
• Paired within-subject design; separate disease change from age drift; match on follow-up.
• Stronger selection: elastic net with stability selection, pre-reduce correlated acoustic features, fold model choice into CV.
• Replicate in biomarker-confirmed, multi-site cohorts with longer follow-up.

Differences are descriptive and confidence intervals overlap; only linguistic over acoustic at first visit reaches significance (approximate two-sided z-test, p < 0.01). Acoustic last-visit cohort reduced by matching; lower CI bound approaches chance. Five classifiers tested per analysis; highest-AUC reported.

References

1. Becker JT, Boller F, Lopez OL, Saxton J, McGonigle KL. The natural history of Alzheimer disease. Arch Neurol. 1994;51(6):585-594.
2. Eyben F, Scherer KR, Schuller BW, et al. The Geneva Minimalistic Acoustic Parameter Set (eGeMAPS) for voice research and affective computing. IEEE Trans Affect Comput. 2016;7(2):190-202.

Acknowledgements

• Work was funded via NIH K23 AG 063900.
• Thanks to staff and faculty of OHSU, CU, and CSU.

www.csandlab.org

Language best distinguishes diagnoses. Acoustics may track change. Different speech signals may serve detection and monitoring.