Can a connectome-constrained model discover subtypes of Parkinson's progression?

A network model that also splits patients

Most progression models solve one problem: placing each patient on a shared biomarker trajectory by estimating a personal disease time. Subtyping models solve a second problem: grouping patients whose trajectories differ. The paper argues that existing approaches rarely do both at scale, because subtyping models tend to be limited in the number of biomarkers they can handle or lack the mechanistic structure of network-based models.

The proposed model extends the COMIND framework so that neurodegeneration spreads through a system of coupled logistic-diffusion equations constrained by structural connectivity. Each brain region accumulates pathology at a rate set by its current state and its remaining capacity, and that accumulation is driven both by connected regions and by a constant regional forcing term. Subtypes are defined by different spatial patterns of those forcing terms, so a subtype is a hypothesis about where pathology originates, not a cluster found in a separate clustering step.

Subtypes differ by source pattern, not by timing

Each subject is assigned to the subtype whose predicted trajectory gives the lowest reconstruction error at that subject's estimated disease time. Because disease time is itself a free parameter, subtypes could in principle separate mainly along timing rather than along spatial pattern. To prevent that, the model adds a Jensen-Shannon divergence penalty that keeps the distribution of time-shifts similar across subtypes.

The parameter set makes the trade-off explicit: the timescale, the region-specific self-propagation rates and the biomarker scaling vector are shared globally, while only the forcing terms vary by subtype. The number of subtypes is chosen by the Bayesian Information Criterion over candidates of two, three and four, with the effective parameter count including nonzero forcing entries and nonzero self-propagation entries.

Synthetic data show the machinery can recover truth

On simulated data with 300 subjects over a 7-year disease axis, BIC correctly selected three subtypes from candidates of two, three and four. The dominant nonzero forcing entry was identified at the true biomarker index in each case, with fitted values of 0.317, 0.302 and 0.249 against a ground truth of 0.3. The sparsity prior set the true zero self-connection entry to exactly zero, and the two nonzero entries with true values of 0.651 and 0.604 were recovered as 0.726 and 0.650.

Subtype classification reached an Adjusted Rand Index of 0.96, and mean absolute time-shift error was 0.345 ± 0.317 years. These are recovery results on data generated by the model's own assumptions, so they test estimation rather than whether real Parkinson's progression follows this generative form.

Four subtypes with different driving biomarkers

In the PPMI application, BIC selected four subtypes over candidates of two, three and four. The subtypes are distinguished by which biomarkers carry the largest forcing terms. Subtype 1 is driven primarily by the TD score together with pallidal and thalamic forcing. Subtype 2 is jointly driven by the PIGD score and MCATOT, the two largest forcing terms in the model, along with strong bilateral putamen and thalamus forcing.

Subtype 3 is more cortically driven, mainly by entorhinal cortex, superior frontal cortex and anterior cingulate. Subtype 4 combines high TD and PIGD forcing with greater forcing in superior frontal and temporal pole regions. The paper presents these as morphologically distinct progression patterns, illustrated through subtype-specific cortical thickness trajectories for the ten regions with the greatest pairwise difference between subtype curves.

Only the connectome model tracks clinical and genetic groups

The benchmark is SuStaIn, a purely data-driven staging model, fitted on the same z-scored data aggregated to 13 regions and 3 clinical scores for computational tractability, with the same training and validation split. Cross-validation favored a two-subtype SuStaIn solution, resembling the cortical and subcortical staging patterns the authors cite.

The two methods were then tested against groupings neither model was trained on. On the hold-out validation cohort of 187 subjects, SuStaIn's two subtypes showed no significant association with TD/PIGD motor phenotype classification (χ2 = 1.22, df = 2, p = 0.543), while the four COMIND subtypes did (χ2 = 27.33, df = 6, p < 0.001). For genetic subgroup across the full cohort of 312 subjects, SuStaIn again showed no significant association (χ2 = 1.35, df = 3, p = 0.717), while COMIND did (χ2 = 17.80, df = 9, p = 0.038).

The comparison is not symmetric in resolution: COMIND was fitted on the full 85 biomarkers while SuStaIn used the reduced 13-region, 3-score set, and the two solutions differ in subtype count. The paper's claim is therefore about what each fitted model recovers under this protocol, not about an inherent superiority of network constraints over staging models in general.

What the validation does and does not establish

The hold-out check is cross-sectional, and the genetic comparison uses the full cohort rather than only held-out subjects. The associations are tests of independence between discovered subtype labels and external groupings, so they show that the subtypes carry clinically and genetically meaningful structure; they do not show that the model's propagation mechanism causes that structure.

The authors also note that both models recover a consistent subcortical-versus-cortical organizing structure despite their different formulations, and they list validation on independent cohorts such as ENIGMA-PD and incorporation of dynamically evolving connectome structure as future work. The connectome is treated as static, and the evidence comes from a single cohort.

The subtype count is the result to scrutinize

The most decision-relevant number here is not the fit quality but the four-subtype solution selected by BIC and its external associations. A reader weighing this model against a staging approach should focus on the matched protocol: same data, same split, and external labels held out from fitting. Under that setup, the two-subtype SuStaIn solution failed to separate motor phenotype and genetic subgroup, while the four connectome-constrained subtypes did.

The practical priority that follows is to treat the forcing-term pattern as the interpretable object: it names which biomarkers drive each subtype, and it is what makes the subtype labels comparable to motor and genetic groupings. Whether those four patterns hold in an independent cohort, and whether they survive a connectome that changes over time, are the conditions that would most change how much weight this result deserves.

Sources used in this answer