When the eye watches a stimulus that repeats at a fixed rate, the visual cortex's electrical activity locks onto that rate, and the brain response shows up as a sharp peak at exactly that frequency in a scalp recording (electroencephalography, EEG). This is called frequency tagging. If a rarer, higher-level event, such as a face, is embedded at a slower rate inside the fast stream, a second peak appears only if the brain actually registered that event. This proposal uses the two peaks together as a staged probe of cognitive decline. The fast "carrier" peak measures how well the eye and early visual cortex are working. The slow "oddball" peak measures whether higher-level recognition happened. The proposed marker is the ratio of the oddball peak to the carrier peak, scored against a model of what healthy people of the same age produce. That marker would then be used to separate very mild dementia from healthy aging, which is the clinical distinction where current tabular tools are weakest and where patients are scarcest in training data.
Two problems make this timely, and they share a solution. Aging alone degrades the steady-state response through lens and retinal changes and a noisier brain. The large eldBETA database of 100 older adults documents a mean signal-to-noise ratio of −13.90 dB, against −13.04 dB in younger cohorts, and a flatter noise spectrum (eldBETA). A naive EEG marker of dementia would therefore largely measure eye age. Separately, the intermediate stage is rare in clinical cohorts because mild patients are managed without advanced workups. Classifiers trained on such skewed data then default to the majority class, as with the 70.8% sensitivity versus 88.2% specificity reported for a heart-sound network (Ryu et al.). Dividing by the carrier cancels much of the optical and early-cortical confound within each person. Scoring the ratio against a healthy-only reference model means the model is fit on the abundant class, and the scarce patients are needed only to set stage thresholds.
Core Claim
The claim is that a within-trial, carrier-normalized oddball index, expressed as a deviation from an age-conditioned model fit only on healthy controls, adds information about Clinical Dementia Rating (CDR) stage 0.5 beyond the Mini-Mental State Examination (MMSE). The claim also has a structural part. Probes at increasing levels of the visual hierarchy should degrade in an ordered way across CDR stages, so that the stage label can be read from which level of the probe hierarchy has been lost. Concretely, let $A_c$ be the amplitude at the carrier frequency and $A_o$ the summed amplitude at the oddball frequency and its harmonics, both after subtracting neighbouring-bin noise. The index is
$$d = \log A_o - \log A_c ,\qquad z = \frac{d - \mu(\text{age})}{\sigma(\text{age})},$$
where $\mu$ and $\sigma$ are the mean and spread of $d$ in healthy controls, estimated as smooth functions of age. The quantity $z$ is the patient-level score.
The contrast with prior work is specific. Frequency tagging has been used for fast command selection in brain-computer interfaces (eldBETA; iFuzzyTL) and for dissociating broad visual categories in healthy adults (SWIFT). None of these produced a graded clinical label. Imbalance-aware dementia staging exists only on static tabular features: SADT pairs SMOTE-ENN rebalancing with a shallow tree on MMSE and brain volume. This proposal differs from both. It treats the tagged response as a hierarchical, self-normalizing clinical measurement. It also handles imbalance by modelling normality rather than by synthesizing patients.
Why the design has to look like this
The strongest constraint comes from a negative result. Word detection in individual subjects is difficult to probe with fast periodic visual stimulation found three things with a 10 Hz carrier and a 2 Hz oddball. The carrier was significant in 10 of 10 individuals and face oddballs in 9 of 10. Word oddballs, however, reached only 4 of 40 datasets once letter-pair statistics were matched, and the authors traced earlier apparent successes to that low-level leak. This result tells us two things. Face oddballs are reliable enough at the single-person level to serve as a clinical measurement, whereas more abstract linguistic oddballs are not and should be excluded. And any high-level tag must be protected against low-level differences between oddball and base images, or it will measure image statistics. SWIFT supplies the method for that protection. Its wavelet-domain scrambling changes the semantic content of an image while holding local luminance, spatial frequency and contrast fixed, and in fMRI it left early visual cortex flat while face-selective cortex followed the tag (AUC 0.91–0.94). SWIFT was run as a slow fMRI paradigm, so adapting it to fast EEG streams is itself untested and is part of the work.
Research Program
Stimulus hierarchy. We will build a three-level EEG protocol in which every stream shares a 6–10 Hz image carrier chosen within the band where older adults still entrain (eldBETA). At the lowest level, a SWIFT-style scrambled-object oddball controls for low-level statistics and should give no oddball peak in anyone. At the middle level, intact faces appear among statistics-matched scrambled images at a slower oddball rate. At the top level, familiar faces appear among unfamiliar faces, with image statistics matched. The hypothesis is that CDR 0.5 selectively reduces the top-level index, and CDR 1 and above reduces the middle level as well. We state this as a hypothesis because the file contains no evidence on face-oddball responses in dementia.
Confound study on existing data. Before any clinical collection, we will use eldBETA's 100 older participants to quantify how much individual carrier signal-to-noise varies with age and with the flatter noise spectrum. This tells us how large the confound is that the carrier normalization must remove. In young volunteers, we will then degrade the stimuli optically with blur and contrast reduction to mimic lens aging. The prediction is that $A_c$ and $A_o$ both fall while $d$ stays within its test-retest band.
Clinical cohort. We will collect a memory-clinic cohort deliberately left at a realistic skew. The target is roughly 80 CDR 0, 35 CDR 0.5 and 20 CDR 1 or above, each with MMSE and, where available, whole-brain volume, matching the SADT features. Staging will be ordinal. The healthy-only normative model provides $z$ scores for each probe level, and a monotone cumulative-link model maps the vector of $z$ scores plus MMSE to CDR stage. Its thresholds are the only parameters the scarce patients inform. All evaluation is at the subject level, with nested cross-validation. Trials are never resampled across folds, because many trials per person would otherwise inflate apparent sample size.
Baselines
The naive reference is the majority-class predictor, together with a single MMSE threshold at the value SADT found as its primary split (≈27.57). This shows how much of the staging problem a one-number bedside test already solves.
The strongest existing staging method is SADT itself: Gini-ranked feature selection, SMOTE-ENN rebalancing and a shallow decision tree. It reports balanced accuracy 0.667 and macro F1 0.638 for CDR staging on OASIS. That figure comes from a different cohort and cannot be compared directly, so we will re-run the full SADT pipeline on our cohort's tabular features with the same folds and the same hyperparameter search budget as our model. The proposal beats it if adding the tagging $z$ scores raises subject-level balanced accuracy, and CDR 0.5 recall specifically, beyond the fold-to-fold spread of SADT. The confidence interval of the paired difference across repeated splits must exclude zero.
The strongest EEG competitor is a discriminative decoder trained directly on patient and control EEG with class-weighted loss. We will run two versions. One uses task-related spatial filtering followed by regularized logistic regression, the supervised family that reaches 150 bits per minute on eldBETA (eldBETA). The other is a fuzzy transfer network in the style of iFuzzyTL, which reports 76.50% one-second command accuracy on eldBETA. Command accuracy is not a diagnostic number: no reported result was found on Lacuna for any frequency-tagging decoder on dementia staging. These decoders get the same subject-level folds and the same tuning budget.
Two ablations isolate the proposal's ingredients. The first replaces $d$ with the raw oddball amplitude $\log A_o$, which tests whether carrier normalization matters. The second replaces the healthy-only normative model with a discriminative classifier on the same features rebalanced by SMOTE, which tests whether modelling normality handles imbalance better than synthesis.
Why not just…?
Why not just add EEG spectral power to MMSE and use class weights? This is the cheapest serious alternative, and it is a real threat if dementia mainly slows resting rhythms. It is covered by the class-weighted discriminative baseline and by feeding resting-state band power into the SADT pipeline. The reason to expect it to fall short at CDR 0.5 is that generic power is exactly the kind of feature eldBETA shows is shifted by normal aging. If it matches the tagging index, we will report that the probe hierarchy adds nothing.
Why not just use the carrier response, which already changes with age? Because that change is what we cannot interpret: eldBETA attributes it partly to lens, retina and broadband noise (eldBETA). The optical-degradation experiment decides this directly. If the raw carrier predicts CDR as well as $d$ but also moves under pure blur in young eyes, it is a measure of eye health, not cognition.
Why not just train a deep network on raw EEG with oversampling? Oversampling few positive subjects invites memorization, and Red-GAN-style conditional synthesis was built for images with segmentation masks rather than for subject-level EEG. With roughly 20 CDR 1 patients, any flexible model will have wide error bars. The SMOTE ablation and the iFuzzyTL-style baseline show whether this concern holds on our data.
Why not use a hierarchical classifier like BiLT for the staging hierarchy? BiLT addresses coarse labels crowding out fine ones in large natural-image label trees. With three or four ordinal stages there is little such competition to fix, so an ordinal model is the simpler fit. We use BiLT's idea only by placing the hierarchy in the stimuli rather than in the classifier.
Proposed Experiments
The first decisive check is on eldBETA and optical blur. The carrier-normalized index should stay stable under simulated lens degradation while raw amplitudes fall, and age should explain far less of the variance in $d$ than in carrier signal-to-noise. If this fails, the clinical study is not worth running.
The second is individual reliability in healthy older controls. The middle-level face oddball must reach per-person significance at a rate comparable to the 9 of 10 reported in young readers (FPVS). The scrambled-statistics level must stay near the false-positive rate, which confirms that no low-level leak remains.
The third is clinical. On the skewed cohort we will report subject-level balanced accuracy, macro F1, per-stage recall, ordinal mean absolute error, and the AUC for CDR 0.5 versus CDR 0 with MMSE alone and with MMSE plus $z$. All of these are compared against the baselines above under identical folds. The structural claim is tested by checking whether the top-level $z$ separates CDR 0.5 from CDR 0 while the middle-level $z$ does not, with both separating CDR 1. An unordered pattern would refute the claim that the probe hierarchy mirrors the diagnostic hierarchy, even if overall accuracy improved.
References
- eldBETA: A Large Eldercare-oriented Benchmark Database of SSVEP-BCI for the Aging Population
- Classification of Heart Sound Recordings Using Convolution Neural Network
- iFuzzyTL: Interpretable Fuzzy Transfer Learning for SSVEP BCI System
- Semantic Wavelet-Induced Frequency-Tagging (SWIFT) Periodically Activates Category Selective Areas While Steadily Activating Early Visual Areas
- Selectively Augmented Decision Tree for Explainable Dementia Detection
- Word Detection in Individual Subjects Is Difficult to Probe With Fast Periodic Visual Stimulation
- Red-GAN: Attacking class imbalance via conditioned generation. Yet another medical imaging perspective
- Bidirectional Logits Tree: Pursuing Granularity Reconcilement in Fine-Grained Classification