Carrying Connectivity Uncertainty into Alzheimer's Subtype and Stage Inference

Brain-scan measures of how regions coordinate from moment to moment ("dynamic functional connectivity") are thought to change as Alzheimer's disease progresses. Almost every study computes them the same way. It slides a fixed-length window along a resting-state fMRI scan and correlates regional signals inside each window. The window length is arbitrary, and it quietly decides both how noisy the result is and which fluctuations survive. When these numbers enter disease-progression models, they arrive as if exact. The noise and bias from the window choice, from short or motion-corrupted scans, and from visits spaced irregularly over years are then absorbed as apparent differences between patients. The models most at risk are the ones meant to sort patients into disease subtypes and place them at a stage along their subtype's path.

We propose to estimate each scan's connectivity without windows, using a continuous-time covariance model that returns a posterior (a full distribution of plausible values, not one number). We then pass that per-scan uncertainty directly into the likelihood of a subtype-and-stage model. In two sentences: a scan whose connectivity is poorly determined should pull less on a patient's subtype and stage assignment, and a scan with no usable fMRI should count as missing, with everything in between handled on one continuous scale. We expect this to make dynamic-connectivity subtypes of Alzheimer's disease stable across analysis choices and consistent across a patient's repeat visits. Neither property can be taken for granted today.

Core Claim

The disease-subtyping literature has already shown that how you treat imperfect measurements changes the subtypes you find. Missing-data SuStaIn extends SuStaIn, a probabilistic model that jointly infers disease subtype and stage from a sequence of biomarker "events." Instead of deleting or imputing missing biomarkers, it integrates them out with a uniform likelihood. On a synthetic benchmark it recovered subtypes and stages better than deletion, mean imputation or nearest-neighbour imputation up to 40% missingness, and on 789 ADNI participants it found five subtypes with different conversion rates.

That work treats each biomarker as either perfectly observed or entirely missing. Dynamic connectivity biomarkers sit between those extremes. Their reliability varies from scan to scan with scan length, sampling rate and how many frames were discarded for head motion.

Our claim is that SuStaIn's missing-data likelihood is the endpoint of a continuum, and the rest of the continuum is supplied by window-free connectivity posteriors. The biomarker model for biomarker $i$ in subtype $c$ at stage $k$ is Gaussian around an expected z-score $\mu_{ick}$ with population noise variance $\sigma_i^2$. Suppose the connectivity estimator gives, for subject $j$, a posterior mean $m_{ij}$ and posterior variance $s_{ij}^2$ for that biomarker. Integrating the biomarker model against this posterior gives

$$ p(m_{ij} \mid c, k) = \mathcal{N}\!\left(m_{ij};\ \mu_{ick},\ \sigma_i^2 + s_{ij}^2\right). $$

As $s_{ij}^2 \to 0$ this reduces to standard SuStaIn. As $s_{ij}^2$ grows large, the biomarker stops informing subtype and stage, which mirrors the uniform missing-data treatment. The equation is simple on purpose. The research question is whether the posteriors are calibrated enough that this weighting removes window-driven artifacts rather than hiding real signal.

Why the Three Directions Meet Here

Windowing bias is the core mechanism. Sliding-window correlation trades variance against temporal blur and acts as an implicit high-pass filter (Frequency modulation increases the specificity of time-resolved connectivity). The dynamic states it finds shift with window and clustering choices, which has produced conflicting disease findings (Tools of the trade; Aberrant dFNC in Major Depressive Disorder). Time-varying functional connectivity as Wishart processes models the covariance matrix as a smooth stochastic process over time. It produced fewer spurious fluctuations than cross-validated sliding windows or DCC-GARCH, a standard econometric time-varying covariance model, on static null data, and it natively returns a posterior. It is our method source.

Temporal sparsity shows up at two scales. Within a scan, removing motion-corrupted frames leaves an irregularly sampled signal. A continuous-time Gaussian-process model handles such gaps directly, whereas a sliding window over censored data silently changes its effective sample size. Between scans, visits are years apart and irregular.

Disease heterogeneity is the application. The subtype-versus-stage confusion that SuStaIn resolves is exactly what current dynamic-connectivity Alzheimer's studies leave open. Linking dynamic connectivity states to cognitive decline replaces the sliding-window pipeline with an end-to-end network whose output layer is a hidden Markov model (a model of discrete hidden states that switch over time), applied to 601 ADNI-3 subjects. It relates state occupancy to diagnosis groups. Disrupted dFNC among cognitive control networks uses sliding windows and k-means on OASIS-3 and correlates state occupancy with cognition. Both stratify by clinical category and learn neither subtypes nor stages.

The closest continuous-time progression models constrain rather than supply our design. TD-HNODE and the multi-resolution trajectory model handle irregular visits and heterogeneity, but for coded EHR events or a single lung-function measure, and TD-HNODE needs expert-defined pathways. LFGP explicitly leaves between-subject heterogeneity as future work. We deliberately do not build a new neural progression model. We change what the established subtyping model is told about its inputs.

Research Program

The first component is the per-scan estimator. We fit a sparse variational Wishart process to each resting-state scan at the level of a modest number of network components, around 17 to 20 in the spirit of the OASIS-3 cognitive-control analysis. This keeps the Gaussian-process cost manageable, since LFGP and Wishart processes both scale poorly with channel count. Censored frames are simply absent time points. From posterior samples of the covariance trajectory we compute a small set of prespecified biomarkers. These include the temporal variability of selected edges, the fraction of time spent near each of a few reference connectivity patterns, and time-averaged connectivity for the same edges. Each biomarker comes with a posterior mean and variance, converted to z-scores against cognitively normal amyloid-negative participants.

The second component is estimator calibration. It must be done before any disease analysis, because the method is only as good as $s_{ij}^2$. Using HCP subjects with four resting scans each, we check whether posterior intervals from one scan cover the estimate from the other three at nominal rates. We then truncate and censor HCP scans to match the lengths and motion profiles seen in ADNI and OASIS-3, and check that coverage survives. Bayesian Structure Learning for Dynamic Brain Connectivity serves as a second window-free estimator. If the Wishart process proves miscalibrated, it gives us a fallback and a robustness check.

The third component is subtype and stage inference. We modify the missing-data SuStaIn likelihood as above and fit it to dynamic-connectivity biomarkers together with the standard ADNI biomarkers: regional volumes, amyloid and tau PET, and cognitive scores. The number of subtypes is chosen by cross-validated likelihood, as in SuStaIn practice.

The fourth component is a known limitation we will measure rather than paper over. Smoothness priors can erase abrupt within-scan changes. If temporal-variability biomarkers are systematically shrunk, that bias shows up in the null and synthetic checks below, and we report it.

Baselines

The naive reference point is missing-data SuStaIn with no functional connectivity at all, using only structural, PET, fluid and cognitive markers as in the ADNI subtyping study. Any connectivity-based model has to add something beyond this. That paper reports subtypes and conversion-rate orderings but, on Lacuna, no numeric subtype-recovery or prognostic accuracy on ADNI. We therefore rerun it ourselves on the same participants and splits.

The second baseline is static connectivity from the whole scan, entered as ordinary biomarkers. The Wishart-process paper argues dynamic methods beat static connectivity on stimulus and phenotype prediction, but not in a subtyping setting, so this comparison is open.

The third, and strongest practical competitor, is sliding-window dynamic connectivity with the window length chosen by cross-validation, followed by k-means, entered into SuStaIn as point estimates. This is what the OASIS-3 and ADNI dynamic-connectivity literature actually uses. We give it the same components, the same censoring, and a tuning budget over window lengths equal to the hyperparameter budget of our estimator.

The strongest reported evidence for window-free estimation comes from within-scan decoding. The Bayesian structure-learning model reports a task-classification AUC of 0.93 on HCP motor data, against 0.80–0.83 for regularized sliding windows. No reported result was found on Lacuna for any window-free estimator inside a disease-subtyping model.

The key ablation feeds Wishart-process posterior means into standard SuStaIn and discards the variances. It isolates whether the benefit comes from better point estimates or from carrying uncertainty.

We also include the CNN-HMM state occupancies from the ADNI-3 study as biomarkers. This tests whether a learned discrete-state summary suffices.

Beating a baseline means three things at once on held-out participants. First, subtype assignments must agree more across analysis choices. Second, stages must stay more consistent across follow-up visits. Third, conversion prediction must be at least as good. Better cross-validated likelihood alone does not count.

Why not just…?

Why not just use static connectivity and avoid windows entirely? Static connectivity has no window choice and can be very reliable, so it is a real competitor for subtyping, where slow disease changes might be carried by averages. The within-scan decoding evidence against it (Wishart processes; Bayesian structure learning) does not directly transfer to progression. Our design includes time-averaged connectivity among the posterior biomarkers and a static-only baseline. If adding the variability and occupancy biomarkers produces no gain over static connectivity, we report that dynamic features carry no additional subtype information. That would be a useful negative result.

Why not just run sliding windows at many lengths and average the resulting subtypes? A window-length ensemble is cheap and directly targets the sensitivity problem, so we treat it as a serious threat and run it. It stabilizes answers across windows, but every window still has sampling variance that shrinks with scan length. Averaging does not tell the model that a six-minute censored scan is less informative than a long clean one. The decisive test is the stratified analysis described below. If the ensemble matches our model among short, heavily censored scans, the uncertainty machinery is not earning its cost.

Why not just plug window-free point estimates into SuStaIn? This is the posterior-mean ablation, and it could plausibly match the full model if posterior variances are nearly constant across scans. In that case the variance term only rescales the noise, and the contribution collapses to "use a Wishart process." We will first measure how much $s_{ij}^2$ actually varies across ADNI and OASIS-3 scans. If the spread is small, we will say the simpler pipeline suffices.

Why not just learn subtypes end to end with a deep continuous-time model? Models like TD-HNODE rely on predefined pathways and much larger event datasets. Neural ODEs and flow models also struggle to beat carry-forward heuristics on slow longitudinal change unless carefully designed, as CRONOS reports for LCI (last context image). With hundreds of fMRI participants, an interpretable event model is the more defensible starting point.

Proposed Experiments

The first experiment is synthetic and has ground truth. We extend the 500-subject, 10-biomarker synthetic SuStaIn benchmark by replacing several biomarkers with simulated fMRI. We generate network time series whose within-scan covariance dynamics shift by subtype and stage, then vary scan length and censoring across subjects. We measure subtype-recovery accuracy and stage error for all baselines. We expect the uncertainty-aware model's advantage to grow as scan quality becomes more heterogeneous across subjects, not merely as it becomes worse on average. That prediction follows from the equation: uniform noise only rescales $\sigma_i^2$.

The second experiment is a null check. We use static null scans, built by randomizing time series so that true connectivity is constant, as in the Wishart-process paper. Then we insert null-derived dynamic biomarkers into the ADNI analysis. The subtype model should assign them near-zero event contributions. Sliding-window features should produce spurious events at a measurable rate.

The third experiment tests window sensitivity on real data in ADNI-3 and OASIS-3. We refit every method under several window lengths, or the equivalent estimator hyperparameters, and report the adjusted Rand index of subtype assignments across settings.

The fourth experiment is longitudinal and replaces the missing ground truth. We fit on baseline visits only. For participants with irregularly spaced follow-up scans, we check two things. Does the follow-up scan receive the same subtype? Is its stage non-decreasing, with the gain scaled by the elapsed time? We report both separately for short or heavily censored scans versus long clean ones.

The fifth experiment is prognostic. We predict time to MCI-to-AD conversion from the inferred subtype and stage using a survival model. For comparison we use the conversion ordering reported by missing-data SuStaIn, in which the Cognitive subtype converts fastest and the Subcortical subtype slowest. We check whether dynamic-connectivity biomarkers sharpen or reorganize that ordering.

The proposal succeeds if the uncertainty-aware model wins on window stability and longitudinal consistency and matches or beats the static-only baseline on conversion, with the posterior-mean ablation falling measurably short among low-quality scans. It fails informatively if the ablation or the window ensemble ties it.