When older adults begin to decline cognitively, the way they talk changes. They use fewer nouns and more pronouns, they repeat words more, and they pack less content into each sentence. These measurable shifts are called lexical profiles. One of them, lexical density, is the share of content words (nouns, verbs, adjectives) among all words. Small studies classify Alzheimer's patients from such counts with surprising accuracy. The trouble is that word choice has many possible causes. Two of them are slow damage to the brain's small blood vessels and the neurodegeneration of Alzheimer's disease, and they call for different treatments. A third is education or dialect, which has nothing to do with disease. Today no one can tell which of these is behind a given person's drop in lexical density. The retina offers a cheap way in. It is part of the central nervous system, and its small vessels share pathology with the brain's. A thirty-second fundus photograph (a color picture of the back of the eye) measures vessel branching, caliber and tortuosity directly.
We propose to collect conversational speech and retinal imaging from the same people. We will then use retinal vascular measurements as a physical anchor that splits each person's lexical profile into a part that tracks microvascular burden and a part that does not. The claim is specific. Different lexical features carry different biological signals. We expect rate and density features to follow retinal vascular burden. We expect the noun-to-pronoun shift to follow neuroretinal thinning, which is a neural rather than vascular change. A lexical profile adjusted for vascular burden should be a more specific marker of neurodegenerative decline than the raw profile.
Core Claim and What Is New
No study in the literature connects lexical density measures to retinal microvasculature. The closest work is Detection of Mild Cognitive Impairment from Language Markers with Crossmodal Augmentation. It learns a linear mapping between conversational language markers and structural or diffusion MRI features from just 16 subjects who had both. It then uses that mapping to give language-only subjects an imaging-informed feature vector, which raised MCI detection AUC from 0.80 to 0.973. Simply substituting fundus photos for MRI in that pipeline would be an incremental modality swap. It would also inherit that paper's goal of a better classifier from an aggregate language vector.
Our goal is different. We want to explain variation in specific, interpretable lexical features using separate vascular and neural retinal measurements, and to adjust explicitly for the systemic vascular risk factors that confound both. Classification gains are a secondary check, not the point. The question we ask, how much of a person's lexical decline is vascular, is one that pure classification cannot answer.
Three findings motivate the split.
- Analysis of spontaneous, conversational speech in dementia of Alzheimer type separated 8 Alzheimer's patients from 16 controls at 87.5% cross-validated accuracy. It used noun rate, pronoun rate and Brunét's W (a vocabulary-richness index that is less sensitive to text length), but had no biological validation.
- Spoken language biomarkers for detecting cognitive impairment found text features alone reached only AUC 0.69, against 0.90 for acoustic features, in the Framingham cohort. This is the strongest argument that lexical measures are noisy. It is also consistent with their mixing several causes. - Cognition–Eye–Brain Connection in Alzheimer's Disease Spectrum showed that retinal microvascular changes appear already at the subjective-decline stage, while neural thinning couples to brain changes later. It used only static tests such as MMSE and the Boston Naming Test, so it cannot say whether everyday language tracks the same split.
The core model is a set of per-feature regressions. For each lexical feature $L_j$ we fit
$$L_j = \beta_j V + \gamma_j N + \delta_j^{\top} C + \varepsilon_j .$$
Here $V$ is a retinal vascular composite, $N$ is a neuroretinal thickness composite, $C$ collects age, sex, education, blood pressure, diabetes status and smoking, and $\varepsilon_j$ is residual variation.
The hypothesis is a dissociation in the pattern of coefficients. Density and diversity features should load on $\beta_j$. Word-class shifts (noun rate, pronoun rate) should load on $\gamma_j$. The residual $\varepsilon_j$, the vascular-adjusted lexical profile, should then separate neurodegenerative from vascular-type impairment better than the raw $L_j$.
Research Program
Retinal side. We will run AutoMorph on fundus photographs. AutoMorph is a four-stage pipeline: image-quality grading, vessel and artery/vein segmentation, then morphometry. Its fractal dimension, vessel density and caliber show ICC above 0.9 against expert annotations. From these we will compute fractal dimension, vessel density, tortuosity, the central arteriolar and venular equivalents, and their ratio. Venular caliber matters in particular: the oculomics survey reports wider venules independently predict stroke with hazard ratio 1.28. Optical coherence tomography (OCT), a cross-sectional retinal scan, and its angiography variant (OCTA) will supply retinal nerve fiber layer and ganglion cell layer thickness for $N$, plus foveal avascular zone metrics for $V$.
The vascular composite will first be built and sanity-checked in UK Biobank. There, fundus images, brain MRI phenotypes and cognitive scores coexist. Neuron-level explainable AI for Alzheimer's Disease assessment from fundus images already showed vessel density and fractal dimension fall with cognition across the Alzheimer's continuum (p < 0.01). Calibrating there means the composite is fixed before it ever sees speech.
Language side. We will compute content-to-function word ratio, the per-100-word rates of nouns, pronouns, adjectives and verbs, and the Guiraud index (unique words divided by the square root of total words). We will also compute Brunét's W, Honoré's statistic, and the lexical diversity density measure from PsyDial. These will come from semi-structured conversation and a picture description task, transcribed automatically, with a manually transcribed subset.
We expect a measurement problem specific to automatic transcription. Speech recognizers drop and confuse short function words more than content words. That would inflate lexical density mechanically, and ADReSSo reports roughly 60% word error rate on elderly picture descriptions. We will estimate this bias on the manual subset and correct each feature through its own calibration curve, rather than assume the features survive transcription errors.
Cohort. No dataset in the literature pairs speech with retinal imaging. We will therefore recruit about 260 older adults spanning healthy, subjective decline, MCI and Alzheimer's. Each participant gets fundus photography, OCT/OCTA, recorded conversation, a standard cognitive battery and recorded vascular risk factors. A subset of about 80 also gets MRI to provide an independent measure of cerebral small vessel burden. The sample size is chosen to detect a feature's incremental $R^2$ of roughly 0.03 from $V$ at 80% power. Diagnoses and composites are fixed before speech features are analyzed.
Baselines
Naive and demographic references. The naive reference is a model using age, sex and education alone. On Framingham speech data a demographic baseline reached AUC 0.79 (Spoken language biomarkers). We reimplement it on our cohort with the same cross-validation folds. Any proposed model must beat it on held-out folds.
Strongest speech competitor. This is the elastic-net logistic regression over acoustic plus text features from the same paper. Elastic net is a linear classifier penalized to keep only a few useful features. That model reported AUC 0.92, accuracy 89%, and acoustic-only AUC 0.90. It is the most direct threat because it suggests acoustics alone may suffice. We rerun it on identical recordings with the same feature-extraction budget and the same nested cross-validation for its two penalty settings. Beating it means adding retinal vascular features to it raises held-out AUC with a bootstrap 95% interval that excludes zero.
Strongest retinal competitor. This is RETFound, a masked-autoencoder foundation model pretrained on 1.6 million retinal images, which reached AUROC 0.794 on 3-year heart failure from fundus photos. We fine-tune it on the same splits with matched tuning trials. For context, a plain VGG-16 on UK Biobank fundus images reached 71.4% average accuracy for Alzheimer's versus controls (LAVA paper). We will report whether interpretable morphometry loses meaningful accuracy relative to the foundation model's representation.
Crossmodal augmentation. This is the crossmodal augmentation method rerun with retinal features in place of MRI. That paper reported 0.973 with language plus augmented imaging and 0.80 with language alone, on I-CONECT with MRI. We have no reported result for its retinal analogue. This tests whether a linear cross-mapping already captures what our per-feature decomposition offers.
Key ablations. The ablation that isolates our key ingredient replaces $V$ with the clinical vascular risk factors alone (blood pressure, diabetes, smoking). A second ablation swaps $V$ with $N$ to test the dissociation.
Why not just…?
Why not just use acoustic features and drop lexical analysis? This is a real threat on pure detection. The 0.90 versus 0.69 gap is large. We do not claim lexical features detect impairment better. We claim they carry interpretable, separable vascular and neural signal that acoustics may not. Two outcomes would favor the acoustic objection. First, the vascular-adjusted lexical residual might add nothing to the elastic-net acoustic model for distinguishing Alzheimer's-type from vascular-type impairment. Second, acoustic features might show the same coupling to $V$. In either case we would report lexical features as redundant for this purpose. ADReSSo's observation that acoustic models overfit on unseen test sets while fused linguistic models did best (78.87%) suggests lexical features may generalize better across sites.
Why not just adjust for blood pressure, diabetes and age instead of imaging the retina? If clinical risk factors explain as much lexical variance as $V$, the retina is an expensive detour. This is the central ablation. The survey reports retinal microvascular features forecast progression better than blood glucose, which is suggestive but not about language. The experiment settles it directly. We compare incremental $R^2$ of $V$ over $C$ against $C$ over $V$. On the MRI subset we also test which better predicts imaging-measured small vessel burden. If $V$ adds nothing, the proposal reduces to a negative result worth publishing.
Why not just reuse the crossmodal augmentation model with retinal images? That model yields a better classifier but no account of which lexical features reflect which pathology. Its own authors flag its linear-combination design as a limitation. We run it as a baseline. If it matches our classification numbers, that is acceptable, because our contribution is the dissociation pattern in $\beta_j$ and $\gamma_j$, which it cannot produce.
Why not just use RETFound embeddings instead of hand-measured morphometry? Foundation embeddings may predict better. But they mix vascular and parenchymal signal: Kim et al. show background texture alone encodes age and sex even with vessels masked. That mixing defeats the vascular-versus-neural split. We report both. Morphometry is preferred only if its accuracy loss is small.
Proposed Experiments and Decisive Checks
Pipeline validation. We first reproduce reference points. On ADReSSo we check that our automatic lexical features plus the late-fusion design approach the reported 78.87% classification accuracy and 5.28 MMSE root-mean-square error. On the AutoMorph benchmark suite we confirm segmentation quality near the reported vessel F1 of 0.73 on AV-WIDE and 0.78 on DR-HAGIS before trusting morphometry. In UK Biobank, we test whether the vascular composite tracks cognitive scores after adjustment for clinical covariates.
Dissociation test. This is the decisive experiment in the paired cohort, with hypotheses pre-registered. Density and diversity features should show significant $\beta_j$ and small $\gamma_j$. Noun and pronoun rates should show the reverse. The difference is tested by a permutation test on the vector of coefficients. The hypothesis fails if every feature loads equally on $V$ and $N$, or if $V$'s contribution vanishes once $C$ is included.
Utility checks. Two secondary checks test whether the decomposition is useful. First, on the MRI subset, does the vascular-adjusted residual separate participants with high versus low small vessel burden worse, and Alzheimer's-type impairment better, than the raw profile? Second, does adding $V$ to the strongest elastic-net speech model raise held-out MCI AUC?
Robustness. All analyses are repeated on transcripts corrected for automatic-transcription bias and on uncorrected transcripts, and stratified by education. This tests whether the lexical-vascular link is an artifact of transcription errors or of schooling.
References
- Detection of Mild Cognitive Impairment from Language Markers with Crossmodal Augmentation
- Analysis of spontaneous, conversational speech in dementia of Alzheimer type: Evaluation of an objective technique for analysing lexical performance
- Spoken Language Biomarkers for Detecting Cognitive Impairment
- Cognition–Eye–Brain Connection in Alzheimer's Disease Spectrum Revealed by Multimodal Imaging
- AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline
- The Eye as a Window to Systemic Health: A Survey of Retinal Imaging from Classical Techniques to Oculomics
- Neuron-level explainable AI for Alzheimer’s Disease assessment from fundus images
- PsyDial: A Large-scale Long-term Conversational Dataset for Mental Health Support
- Detecting Cognitive Decline Using Speech Only: The ADReSSo Challenge
- A foundation model for generalizable disease detection from retinal images
- Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images