Dynamical Reserve: Predicting MCI Conversion from How Close the Wiring Is to Collapse

People with mild cognitive impairment (MCI) are already losing white matter connections, but their thinking and test scores stay roughly normal until the brain can no longer compensate. Then decline arrives quickly. Current imaging predictors measure the amount of damage, for example how many fibres connect two regions. They do not measure how much compensation capacity remains. We propose to simulate each patient's brain on their own wiring diagram and compute a single number from it: how much further weakening of the fibre tracts between deep brain structures and the cortex the simulated network could absorb before its activity collapses into a pathologically over-synchronized state. We call this number dynamical reserve. The wiring diagram is reconstructed with diffusion tensor imaging (DTI), an MRI method that traces fibre bundles by measuring which direction water diffuses most easily. The deep structures are chiefly the thalamus and striatum.

Two lines of evidence motivate the design. First, deep structures stabilize cortical dynamics. In stroke, damage to tracts linking thalamus and basal ganglia to cortex pushes brain activity from a segregated state into a hyper-integrated one (Favaretto et al.). Second, the point at which a network tips into synchrony depends on how diverse its units are. Populations of identical units phase-lock easily and abruptly (Lengler & Steger; Dahmen et al.). A model of identical units would therefore misplace the tipping point. We want to test whether a reserve computed with diverse time constants and explicit deep-structure nodes predicts time to Alzheimer's dementia beyond what the best blood marker already provides. We focus especially on patients whose blood marker sits in the uncertain middle range.

Core Claim

Conversion from MCI to Alzheimer's disease (AD) tracks proximity to a nonlinear threshold rather than accumulated damage. Proximity to that threshold can be estimated from one baseline diffusion scan, but only under two conditions. The deep-structure loops must be modelled as active nodes, not as volume labels. Each region must be modelled as a population whose units differ in their time constants (how quickly each unit forgets its input), rather than as a single uniform unit.

The claim makes three falsifiable predictions:

  1. Reserve adds predictive value over plasma p-Tau217, a blood test for a form of the tau protein that accumulates in AD.
  2. The added value concentrates in p-Tau217's uncertain middle window.
  3. Removing either the deep-structure nodes or the within-region diversity erases the added value.

Relation to Prior Work

The closest work uses diffusion connectomes directly as classifier features. Kim et al. concatenate probabilistic connectomes with FreeSurfer morphometry, which measures regional volumes and cortical thickness from T1-weighted MRI. This works well for cross-sectional diagnosis. For predicting MCI-to-AD conversion in ADNI-2, however, it drops to 59% accuracy, below morphometry alone (69%). Zhan et al. show that nine tractography algorithms give indistinguishable MCI-versus-normal results, with area under the ROC curve (AUC) of 0.60–0.70. They also find that about 10 principal components beat the raw connectome. Together these results motivate the proposal: raw edge weights carry a weak, noisy prognostic signal. They also constrain it: we use cheap deterministic tractography and must beat a low-rank linear summary.

Oxtoby et al. order the stages of connectivity failure with event-based modelling. They find local cortical centrality losses before hippocampal atrophy, but they do not model deep-structure loops or produce individual survival times. Graph learning approaches such as BrainGB and BioBGT learn representations of brain networks. In both, each region is an aggregation step with no internal dynamics, and BioBGT reaches only 52.08% on three-way ADNI classification.

The method draws on two literatures:

  • Within-population diversity. Perez-Nieves et al. draw time constants from gamma distributions and gain 15–20% on temporally rich tasks, but negligible gains on static ones. That negative result tells us diversity should matter only when the readout depends on dynamics, which ours does by construction. Duarte et al. link diversity to the memory capacity of cortical circuits.
  • Deep structures in large-scale models. van Meegen et al. and Joglekar et al. argue that cortex-only models cannot reproduce state gating or stable propagation between areas.

What none of these do is put diverse within-region dynamics onto an individual's diffusion connectome, include deep structures, and read out a distance-to-threshold that is evaluated against conversion times.

Research Program

Network construction. For each participant we build a structural connectome from DTI. It covers about 130 cortical regions plus bilateral thalamus, caudate, putamen and pallidum. We use deterministic tractography with anatomically constrained tracking and streamline filtering, following the MRtrix pipeline of Oxtoby et al.

Node dynamics. Each region becomes a population of $N$ rate units. Their time constants are drawn from a gamma distribution with a shared mean $\bar\tau$ and coefficient of variation $c$; setting $c=0$ recovers the homogeneous model. Regions couple through the subject's connectome, scaled by a global coupling strength $G$. The deep-structure nodes close the loops cortex → striatum → pallidum → thalamus → cortex. Within each loop, the sign of each projection is set from standard anatomy (for example, the pallidum's output to the thalamus is inhibitory), not fitted.

Group-level fitting. The parameters $\bar\tau$, $c$ and the normative operating coupling $G_0$ are fitted once, at the group level. We fit them to resting-state functional connectivity from cognitively normal ADNI participants. No patient outcome enters the fit, which avoids leaking the label into the predictor.

The reserve measure. For a subject we progressively scale down only the edges between deep structures and cortex by a factor $s$. We record the smallest scaling $s^*$ at which a global synchrony index exceeds the level that defines the hyper-integrated regime. That regime is calibrated so the model reproduces the segregated-to-integrated shift reported by Favaretto et al. Reserve is

$$R = -\log s^*,$$

which is the log-fold weakening of the deep-structure–cortex tracts the subject could absorb before collapse. A larger $R$ means more reserve.

Calibration check before any prognosis. Prognostic analysis is gated by a validity check. The model must reproduce the direction of the segregated-to-integrated shift when lesion-derived disconnection is applied. We test this in the Washington University stroke cohort if access allows, or otherwise in simulated tract lesions on ADNI controls. If the homogeneous and diverse models fail this check, or pass it equally, we report that and stop before the prognostic analysis.

Baselines

Do-nothing reference. A Cox survival model with age, sex, education, APOE genotype and baseline MMSE score. This shows how much any biomarker adds at all.

Strongest existing competitor: plasma p-Tau217. We fit it in a random survival forest exactly as in Chen et al., who report a 4-year AUROC of 0.787 for p-Tau217 alone on 201 ADNI MCI subjects. Amyloid PET reached 0.758 in the same study. Our analysed sample is MCI participants with both diffusion MRI and plasma p-Tau217, which will likely be smaller and different. We therefore refit Chen's model on our exact sample with the same cross-validation folds. We report the published 0.787 only as context. Beating it means that p-Tau217 plus reserve improves the 4-year time-dependent AUROC and Harrell's concordance index (C-index, the share of patient pairs whose predicted order of conversion matches the observed order) over p-Tau217 alone. The improvement must have a bootstrap confidence interval excluding zero and a significant likelihood-ratio test.

Morphometry baseline. FreeSurfer volumes and thickness, as in Kim et al. They report 69% conversion accuracy in ADNI-2; we found no reported survival C-index for it in the file.

Connectome baselines. These represent "damage, not threshold":

  • a Cox model on the top ~10 principal components of the connectome, following Zhan et al.;
  • a random survival forest on mean fractional anisotropy (FA, a 0-to-1 measure of how strongly water diffusion is aligned along the fibres) of each deep-structure–cortex tract;
  • a BrainGB-style graph neural network with connection-profile node features and a survival head.

Fair comparison. All baselines get the same participants, folds and hyperparameter search budget. For each baseline we match the number of tuned settings to the three group-level parameters the simulation fits.

Ablations that isolate the key ingredients. We compute reserve three more ways: with $c=0$ (homogeneous populations), with deep-structure nodes removed and reserve computed by weakening long-range cortical edges, and with diverse time constants placed between regions instead of within them. The proposal succeeds only if the full model beats all three on the incremental C-index.

Why not just…?

Why not just use p-Tau217, which is cheaper and already strong? This is the main threat, and we do not claim to replace it. Chen et al. show that p-Tau217 conversion risk rises from 10% to 45% across 0.08–0.56 pg/mL. Their two-cutoff rule confidently classifies only 68.6% of patients, leaving about a third in an uncertain middle range. Our primary endpoint is therefore the concordance of reserve within that window. If reserve adds nothing there, a diffusion scan has no case at all.

Why not just use FreeSurfer morphometry? Kim et al. show it beats their connectome model for prognosis. However, atrophy is a late, cumulative signal, while our hypothesis concerns the margin before decompensation. We test both directions: reserve's added value over morphometry, and whether reserve predicts conversion in participants with normal hippocampal volume, where morphometry should have little to work with.

Why not just feed deep-structure tract FA into a flexible survival model? This is the most serious technical threat. If $R$ were just a monotone transform of total thalamo- and striato-cortical FA, a random survival forest on those tract values would match it. Our decisive analysis therefore checks whether $R$ keeps its predictive value after adjusting for those tract FA values and the connectome principal components. We also examine whether two subjects with matched tract FA but different reserve, due to how damage is distributed, actually convert at different rates. If the forest matches the simulation, we will report that the simulation adds interpretability but no prognostic value.

Why not just skip the diversity and use a standard homogeneous population model? Perez-Nieves et al. found diversity irrelevant for static tasks, so this objection could be right. The $c=0$ ablation settles it directly. Our specific prediction is that homogeneous models produce abruptly synchronizing networks. That would compress $R$ into a narrow range and reduce its ability to rank patients, which we can check from the variance of $R$ alongside the C-index.

Proposed Experiments and Decisive Checks

Data. The main cohort is ADNI-2/3 MCI participants with baseline diffusion MRI (ADNI-2 uses 41 gradient directions), plasma p-Tau217 and longitudinal diagnoses. Cognitively normal participants serve for normative fitting.

Primary metrics. The 4-year time-dependent AUROC and C-index under repeated nested cross-validation. We also report net reclassification of patients in the uncertain p-Tau217 window.

Robustness checks.

  • Repeat the analysis with probabilistic tractography; Zhan et al. predict little change.
  • Harmonize across scanner sites and report results per site, since the ADNI study of Kim et al. shows site effects matter.
  • Vary the hyper-integration threshold within the range that still reproduces the stroke result.

Decisive outcome. The full model must beat p-Tau217 inside the uncertain window. It must survive adjustment for tract FA. It must lose its advantage when either diversity or deep-structure nodes are removed. If all three hold, compensation capacity is measurable and adds prognostic information. If only the tract-FA forest matches it, the threshold story is unnecessary. If nothing beats p-Tau217, diffusion MRI should not be added to the MCI workup on prognostic grounds alone.