Brain-scan classifiers can now flag Alzheimer's disease in people who have no symptoms yet. One reported model finds roughly two-thirds of asymptomatic cases from resting-state functional MRI, where a simpler classifier finds none. These models read dynamic functional connectivity: how strongly the activity of brain networks moves together within short sliding windows across a single scan. A resting scan, however, is not a series of independent snapshots. Over ten minutes in a scanner, people get drowsy, settle in, and drift. That drift is an order effect: a window late in the scan differs from an early one simply because it comes later. If people with early amyloid buildup drift differently from healthy people, a classifier could learn "who gets sleepy faster" and report it as early neurodegeneration.
This proposal tests that possibility directly and builds a correction for it. We first measure how much of a published classifier's preclinical signal depends on where each window falls in the scan. We then train a generator of synthetic brain signals that can hold a person's disease-related connectivity fixed while changing their within-scan drift, and use these counterfactual samples to train classifiers that ignore drift. A third ingredient tells us which signal is real. Brain function is thought to fail before brain structure. A genuine early functional marker should therefore predict structural decline measured on later scans, and a drowsiness artifact should not.
Core Claim
Preclinical sensitivity from dynamic connectivity models is partly confounded by within-session time-on-scan effects. The disease component and the drift component can be separated by two kinds of evidence. The first is intervention on window order. The second is prospective prediction of later structural atrophy. Our hypothesis is that a drift-invariant functional score keeps most of its asymptomatic sensitivity and predicts later structural loss at least as well as the original score, while the drift component carries little prognostic value.
Why This Gap Exists
The motivating result is the spatio-temporal attention model of Wei et al.. It runs convolution and temporal attention over sliding-window connectivity across 53 networks. It reports 66.10% sensitivity and 76.76% accuracy on asymptomatic individuals in the Emory Healthy Brain Study, against 0% sensitivity for a support vector machine, and 93.76% accuracy on ADNI (the Alzheimer's Disease Neuroimaging Initiative cohort). It also finds that asymptomatic profiles resemble mild cognitive impairment. That finding is exactly what a shared drowsiness pattern could produce. The model treats windows as interchangeable time slices and never asks whether evidence piles up late in the scan.
The order-effects literature says this omission is risky. Ahani et al. showed that progressive drowsiness and fatigue accumulate in EEG and respiration across a session, enough to confound comparisons between conditions. Cole et al. showed that shared task-evoked activity systematically inflates functional connectivity estimates. Time-locked shared signals therefore masquerade as connectivity, and drowsiness is such a signal.
Mathur, Baldwin and Cohn supply the key methodological lesson, from human annotators rather than brains. Randomizing one presentation order for everyone does not remove sequence bias. Only an intervention that breaks the link between sequence position and the quantity being measured turns that bias into zero-mean noise. In their data, sequence bias was strongest early in the session and then declined, which is a reminder that drift need not be linear.
We borrow that logic. A scan's time axis cannot be re-randomized, but position can be decoupled from disease status in two ways: synthetically, with generated samples, and analytically, with window-order interventions.
The functional-before-structural ordering of neurodegeneration provides the external check. Zhou et al. describe early loss of network modularity and efficiency preceding atrophy. Gonneaud et al. find functionally "older" brains in mutation carriers years before symptoms. The structural connectome study finds that white-matter integrity mediates only 21.6% of amyloid's effect on cognition. Together these papers motivate a testable prediction: true early functional disruption should foreshadow later structural change.
Research Program
The first stage is a diagnostic audit, with no new model. We reimplement the dynamic connectivity pipeline and classifier of Wei et al. on ADNI resting-state fMRI, and on the Emory cohort if access permits. For every subject we compute two quantities. One is the classifier's attribution by window position: how much of the decision comes from each third of the scan. The other is a drift trajectory, meaning how global signal amplitude and network-coupling summaries change with time-on-scan.
We then run three interventions at test time. We reverse window order, shuffle windows within subject, and truncate the scan to its first or last half. If the decision depends on temporal structure, reversal and shuffling will change it. If the evidence is concentrated late in the scan, the late-half input will drive the classifier's decisions.
We also ask a separate question: does drift trajectory alone, with no connectivity pattern, classify amyloid-positive from amyloid-negative people above chance? If it does, the confound has a path into the classifier.
The second stage is a drift-controllable generator. We build on Dual-Spectral Flow Matching (DSFM). DSFM synthesizes resting BOLD signals by moving them into a wavelet-then-cosine coefficient space and learning a flow-matching model there; flow matching is a generator that learns a smooth path from noise to data. DSFM improved accuracy on ABIDE from 64.7% to 71.5% and on REST-meta-MDD from 58.9% to 70.8%. The ABIDE result is an autism cohort and REST-meta-MDD is a depression cohort; neither involves Alzheimer's disease.
DSFM conditions only on diagnosis. We add two separate conditioning inputs: a subject-level disease embedding and a drift trajectory. The generator learns to keep them apart through a swap objective. A sample generated with subject A's disease embedding and subject B's drift should match A on drift-detrended connectivity statistics and match B on time-on-scan dynamics.
For augmentation, we generate drift-swapped counterfactuals for every training subject. Each disease label is then seen under the full range of drift profiles. We add a consistency penalty that asks the classifier to give the same output for a real subject and for that subject's drift-swapped twins.
We check plausibility with the measures DSFM already uses, hemodynamic response shape and power spectrum, rather than image-quality scores. The field's known failure mode is that synthetic neuroimaging can look right while getting connectome topology wrong.
The third stage is structural anchoring. ADNI contains people with a baseline resting scan and later T1 or diffusion scans. For them we regress later change in cortical thickness and in fractional anisotropy (a diffusion measure of white-matter integrity) on three baseline scores: the original functional score, the drift-invariant score, and the pure-drift score. We use networks with prior support, namely default mode, frontoparietal, visual and cognitive-control, and control for age and baseline structure. This step is what makes "real" distinguishable from "artifact" without ground-truth vigilance recordings, which ADNI lacks.
Paired longitudinal function-plus-structure data are scarce, so this analysis will be underpowered. We will pre-register effect sizes and report confidence intervals rather than claim significance we cannot reach.
OCR noise enters as a narrow, secondary stress test, not a pillar. Zhang et al. combine diffusion-based connectomes with clinical reports and reach 92.94% accuracy on ADNI mild impairment. Reports containing terms like "MMSE" (the Mini-Mental State Examination) can carry the diagnosis themselves, so the imaging contribution of such models is unclear. In a fusion variant, we degrade the reports with character-level OCR-style corruption following Cheap Character Noise. We then measure how much accuracy survives on imaging alone, with and without our drift correction. Real scanned clinical records are noisy: Berge et al. attribute 32.3% of extraction failures to OCR. The question is whether a fusion model's imaging branch is robust enough to carry diagnosis when its text is unreliable.
Baselines
The naive reference is the Wei et al. classifier trained and tested as published. Its reported 66.10% asymptomatic sensitivity on the Emory cohort and 93.76% ADNI accuracy are the targets our reimplementation must approach before any comparison counts. A static, whole-scan connectivity classifier serves as a floor; Wei et al. report 0% asymptomatic sensitivity for a support vector machine.
The strongest competitor is the same architecture with conventional nuisance control. Before computing windows we regress out window index, global signal, motion, and a smooth time-on-scan basis, and we add window position as an explicit input. This is what a careful practitioner would do, and it may remove much of the drift. All variants share data splits, preprocessing, architecture, and a fixed hyperparameter-search budget.
The ablations isolate our key ingredient, drift-swap conditioning. One ablation is DSFM-style augmentation with diagnosis-only conditioning at the same number of synthetic samples. Another is our generator with the swap objective removed.
Beating the baselines means three things together. First, asymptomatic sensitivity stays within a few points of the original. Second, the gap between early-half and late-half evidence shrinks substantially. Third, baseline-to-follow-up structural prediction is no worse than nuisance regression achieves.
Why not just…?
Why not just regress time-on-scan out of the signal? This is the real threat. If nuisance regression matches the generator on all three criteria, the generator is unnecessary, and we will report the audit and the regression fix as the contribution. Our reason to expect a gap is that drift is not a single additive trend. Mathur et al. found that sequence bias changes shape across a session. Drowsiness also alters coupling between networks, not only signal amplitude, and linear regression cannot remove coupling changes without also removing disease-related coupling. The head-to-head comparison in the baselines settles the question.
Why not just use only the first few minutes of each scan? Truncation trades away the windows that dynamic models need and lowers the signal-to-noise ratio. We include first-half-only training as a baseline. If it keeps sensitivity and predicts structural change as well as our method, it wins on simplicity.
Why not just use plain generative augmentation, as DSFM does? Diagnosis-only augmentation can copy the confound, because samples generated for a diagnosis inherit that group's typical drift. The diagnosis-only ablation measures exactly this. If its late-window attribution matches the uncorrected model, that is direct evidence.
Decisive Experiments
The pivotal result comes from the audit. If drift trajectories alone classify amyloid status above chance, and the original model's evidence is concentrated in late windows, the confound is real and the rest of the program follows. If neither holds, we stop and report a useful negative result: current preclinical dynamic-connectivity signal is not a drowsiness artifact.
The second decisive check is structural anchoring. Suppose the drift-invariant score predicts later thinning and fractional-anisotropy loss in default mode, frontoparietal and visual networks, and the pure-drift score does not. That would validate the separation using the expected ordering of neurodegeneration, without vigilance recordings.
The third is generalization from ADNI to the Emory cohort, the asymptomatic setting where the original claim was made.
The OCR fusion test is reported separately. It is secondary and should not be read as the main result.
References
- A deep spatio-temporal attention model of dynamic functional network connectivity shows sensitivity to Alzheimer's in asymptomatic individuals
- Quantitative change of EEG and respiration signals during mindfulness meditation
- Task activations produce spurious but systematic inflation of task functional connectivity estimates
- Sequence Effects in Crowdsourced Annotations
- The Biomarkers for Identifying Preclinical Alzheimer's Disease via Structural and Functional Magnetic Resonance Imaging
- Accelerated functional brain aging in pre-clinical familial Alzheimer’s disease
- Mapping Structural Connectome Patterns in Alzheimer's Disease
- Functional MRI Time Series Generation via Wavelet-Based Image Transform and Spectral Flow Matching for Brain Disorder Identification
- Bridging Brain Connectomes and Clinical Reports for Early Alzheimer's Disease Diagnosis
- Cheap Character Noise for OCR-Robust Multilingual Embeddings
- Combining Unsupervised, Supervised, and Rule-based Algorithms for Text Mining of Electronic Health Records - A Clinical Decision Support System for Identifying and Classifying Allergies of Concern for Anesthesia During Surgery