Void Fraction and Neighbor Loss as Triggers of Accelerating Atrophy in ADNI

In crystals, an ordered arrangement resists heat for two measurable reasons. The first is how much of its local cavity each unit fills. The second is how many intact neighbors hold it in place. Structural phase transition drives outright photoluminescence quenching and dielectric duple bistable switching shows the first effect cleanly. Filling the voids around an organic cation with bulkier halogenated groups raises its order-disorder transition from 293 K to 393 K and 454 K. The second effect, the role of coordination number, is the lesson drawn from high-throughput ferromagnet screens such as High-throughput computation and structure prototype analysis for two-dimensional ferromagnetic materials. This note borrows those two descriptors, not the thermodynamics, and asks a sharp question about Alzheimer's disease. On the brain surface, does a patch of tissue start shrinking faster once the fluid-filled space around it grows past a threshold and once enough of its neighboring patches have already shrunk? If so, atrophy is a cooperative process that spreads outward from damaged neighbors. If not, it is a sum of independent local declines.

The practical payoff is better forecasting of where and when atrophy will accelerate in people with mild cognitive impairment (MCI), using longitudinal scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The scientific payoff is a falsifiable test. The dominant ADNI models treat deformation either as a smooth field with no neighbor interaction or as flattened feature vectors. Feature selective temporal prediction of Alzheimer's disease progression using hippocampus surface morphometry explicitly names that as a limitation, noting that neighborhood information on the hippocampal surface is ignored.

Core Claim

We claim that the future atrophy rate at a point on the medial-temporal and periventricular surface depends nonlinearly on two local quantities. The first is the local void fraction $\phi_i$: the share of cerebrospinal fluid (CSF) within a fixed-radius shell around surface vertex $i$. The second is effective coordination $c_i$: the fraction of vertex $i$'s mesh neighbors whose own shrinkage remains below an "intact" cutoff. These two quantities should interact, and the effect should hold beyond the vertex's own past atrophy. The model we fit is

$$ r_i(t{+}1) = \alpha_i + \beta\, a_i(t) + \gamma\,\sigma\!\big(w_\phi \phi_i(t) + w_c\,(1-c_i(t)) - \theta\big), $$

The terms are as follows. $r_i$ is the next-interval atrophy rate, measured as the log-Jacobian contraction between visits. $a_i(t)$ is cumulative atrophy so far. $\alpha_i$ is a vertex- and subject-level baseline. $\sigma$ is a logistic function. $\gamma$, $w_\phi$, $w_c$ and $\theta$ are learned. The lattice analogy predicts three things: $\gamma>0$, a sigmoid sharper than a linear fit can mimic, and $w_c$ that stays significant after we control for spatially smoothed own-atrophy. Mean-field theory is the source of the coordination prediction. There, the ordering threshold scales with the number of coupled neighbors, so thin or boundary regions with low intact coordination should tip earlier at the same void fraction.

This is not a claim that the brain has a Hamiltonian. $\phi$ and $c$ are measurable geometric descriptors, and the sigmoid coupling is a hypothesis the data can reject.

Relation to Prior Work

Three ADNI lines come closest.

Disentangling normal aging from Alzheimer's disease in structural magnetic resonance images models deformation as a continuous velocity field that is linear in time. It uses field divergence to separate disease-specific atrophy from aging, reaching 91% sensitivity and 84% specificity for AD versus controls. We reuse its idea of removing a normal-aging component before measuring atrophy. We differ by modeling rate changes that depend on the state of neighbors, which a stationary linear field cannot express.

The mTBM work above supplies our surface representation. It extracts 2,100 tensor features per hippocampus and predicts ADAS-Cog, the Alzheimer's Disease Assessment Scale cognitive subscale, with a fused sparse group lasso. It motivates the proposal, since it names the neighborhood gap. It does not model interactions between vertices.

LSOR: Longitudinally-Consistent Self-Organized Representation Learning projects scans onto a learned 2D grid. That grid is not anatomical, so the word "lattice" there means something different from ours.

Two further works constrain the design.

GSplit LBI: Taming the Procedural Bias in Neuroimaging for Disease Prediction shows that preprocessing and scanner effects correlate with diagnosis and can inflate accuracy. Registration errors concentrated at ventricle boundaries are exactly what could fake a neighbor-coupling effect near CSF, so we build that control in rather than assume it away.

Multimodal subtypes identified in Alzheimer's Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference shows that atrophy follows distinct subtype sequences. A spatial coupling effect could simply reflect subtype membership, so we stratify by subtype.

On the materials side, the two papers already cited play only the role of supplying descriptors. Antiferromagnetic order in the pyrochlores R2Ge2O7 is a caution about carrying the analogy further. There, compressing the lattice does not raise ordering temperatures monotonically, and it moves the system between competing ground states. We therefore do not assume that the void-fraction effect is monotone. The logistic form is tested against a non-monotone spline.

Research Program

The data are ADNI-1 longitudinal T1-weighted scans, the same population used by LSOR (2,389 scans, 632 subjects) and by the mTBM study (616 subjects). We restrict to subjects with at least three visits so that rates can be computed on two intervals.

Preprocessing runs in four steps. We segment subcortical structures and ventricles. We build hippocampal and periventricular surface meshes in template space. We compute per-vertex log-Jacobians between consecutive visits using within-subject registration. We regress out a normal-aging trajectory fitted on amyloid-negative controls, following Pennec et al. From each vertex's CSF shell we compute $\phi_i$, and from its one-ring and two-ring mesh neighbors we compute $c_i$. The intact cutoff and the shell radius are chosen on a validation split only.

We then fit the model hierarchically. Subject random effects enter through $\alpha_i$. Region-specific $\gamma$ is estimated separately for the hippocampus, entorhinal border and periventricular white matter. The model is trained on subject-level splits. Field strength is held out: we fit on 1.5T and test on 3T.

The fitted per-vertex tipping score, $\sigma(\cdot)$, is then summed into a subject-level index. That index serves as a candidate predictor of MCI conversion and of 48-month ADAS-Cog.

Baselines

The naive reference carries forward each vertex's last observed atrophy rate, together with scalar bilateral hippocampal volume for the clinical tasks. Baseline MRI Predictors of Conversion from MCI to Probable AD in the ADNI Cohort reports hippocampal volume as the strongest single conversion predictor, with Cohen's $d = 0.603$. Any added machinery has to beat that. On clinical tasks, beating it means a paired improvement in conversion AUC with a bootstrap confidence interval excluding zero.

The strongest competitor for the core question is a spatially smoothed independent model. This is the same hierarchical rate model, but its interaction term is replaced by a graph-Laplacian penalty or a graph convolution over the mesh. It represents the reviewer's view that "neighborhood information" just means smoothing. It gets the same vertices, the same splits, and a matched hyperparameter search budget, with the same number of configurations and the same validation data. Our model beats it only if held-out vertex-level rate prediction improves and the coupling term keeps a significant $w_c$ when the smoothed own-atrophy is included as a covariate. The Lacuna file reports no result for vertex-level atrophy-rate forecasting on ADNI, so this comparison has to be established here.

For 48-month ADAS-Cog, the strongest reported method is mTBM features with cFSGL, the convex fused sparse group lasso, at rMSE $3.164 \pm 1.091$. That compares with $4.367 \pm 1.179$ for standard volumes plus demographics, per the mTBM paper. We rerun cFSGL on our split, then add the tipping index. Success is a paired rMSE reduction across repeated splits.

For stable versus progressive MCI, LSOR reports AUC 72.4. We rerun it with the released LSOR code on identical splits. Sparse learning and stability selection for predicting MCI to AD conversion using baseline ADNI data reports a multimodal biosignature AUC of 0.8587 against 0.7877 for MRI alone. That MRI-only figure is the relevant ceiling to beat with imaging-only inputs.

Three ablations isolate the ingredients. The first replaces the sigmoid with a linear term, which tests whether there is a threshold. The second drops $\phi$ or drops $c$, which tests which descriptor matters. The third uses a degree-preserving shuffle of the neighbor graph, which tests whether true anatomical adjacency matters.

Why not just…?

Why not just use hippocampal and ventricular volumes? Scalar volumes are strong for diagnosis, as the $d=0.603$ result shows. But they cannot say where on a structure atrophy will accelerate. The mTBM result, where rMSE falls from 4.367 to 3.164, shows that local surface information adds forecasting value. Scalars remain the naive baseline for every clinical endpoint. If the tipping index adds nothing over volume for conversion, we will report the index as a mechanistic finding only.

Why not just smooth features over the mesh or train a graph neural network? This is the real threat. A flexible graph network could absorb any neighbor effect without naming it. Our answer is empirical. The proposal's value is not higher accuracy than a graph network. It is a specific, interpretable interaction between void fraction and intact-neighbor count, together with a threshold, that a smoothed additive model lacks. If the Laplacian or graph-convolution baseline matches held-out error and $w_c$ vanishes once smoothed own-atrophy is controlled, the cooperative hypothesis is rejected. We would then say the lattice lens reduces to smoothing.

Why not just stage patients with SuStaIn or LSOR? Both describe where a patient sits in a population trajectory. Neither predicts local rate changes from the local state. Subtype-driven spatial clustering could mimic coupling, so we refit within each SuStaIn subtype. The coupling effect has to hold inside subtypes to count.

Why not just fit a separate nonlinear time curve per vertex? Independent per-vertex acceleration can produce apparent tipping without any neighbor influence. The ablation that keeps a flexible own-history spline but removes the neighbor terms tests exactly this.

Decisive Checks

The central test is whether $w_c$ and the $\phi \times c$ interaction survive four checks together: out-of-sample prediction on held-out subjects, transfer from 1.5T training to 3T testing, a degree-preserving neighbor shuffle, and adjustment for spatially smoothed own-atrophy.

Guarding against procedural artifacts gets its own check. We re-estimate the coupling on vertices more than a fixed distance from any CSF boundary. If the effect exists only at ventricle edges and disappears in interior tissue, registration bias of the kind GSplit LBI documents becomes the leading explanation.

A positive outcome would also leave a direct signature: at matched void fraction, vertices with fewer intact neighbors should show earlier onset of acceleration. A null result on these checks is equally publishable. It would show that ADNI atrophy is well described by independent, smoothly varying local decline.