Telling Parkinson's from Essential Tremor Without Symptoms Alone: Diffusion Similarity Networks Plus a Graph Network Reach 87.5% Balanced Accuracy on a Hold-Out Set
Synopsis
Parkinson's disease and essential tremor look alike early, with misdiagnosis rates of 20% to 30% even at specialized clinics, and the white-matter tracts mapped by conventional diffusion tensor imaging do not capture microstructural change inside gray matter. The study instead builds diffusion similarity networks, linking brain regions by the statistical similarity of their diffusion profiles, then filters connections with iterative permutation sampling recursive feature elimination and classifies with a graph neural network. In five-fold cross-validation on a local cohort of 100 people, the pipeline separated Parkinson's disease from essential tremor with an AUC of 0.94, and on a separate 40-person hold-out set it reached 87.5% balanced accuracy.
Interpretation
Edges between brain regions are no longer defined only by white-matter fiber bundles but by the statistical similarity of two regions' diffusion profiles, yielding a microstructural similarity network that covers gray matter. Earlier structural connectivity matrices came from diffusion tensor imaging tractography and quantified only physical white-matter microstructure such as fiber counts or streamline densities, so they could not capture localized microstructural remodeling within gray matter or at the gray-white matter interface. In healthy controls, nodal strength of the diffusion similarity network coupled significantly with cortical thickness (r = 0.73, p < 0.001), whereas nodal strength from conventional tractography under identical conditions showed a weak, non-significant relationship (r = 0.10, p = 0.35).
The same networks serve three tasks: separating Parkinson's disease from healthy controls, tracking progression over four years, and distinguishing Parkinson's disease from essential tremor. Connectomic measures had not previously been used to differentiate Parkinson's disease from essential tremor, two conditions that both present with tremor early but require different treatments and have different long-term outcomes. The cross-sectional comparison used 41 people with Parkinson's disease and 66 age- and sex-matched healthy controls from the Parkinson's Progression Markers Initiative database, and the longitudinal analysis used the same 41 patients scanned at baseline and at roughly 1, 2 and 4 years.
On the local cohort, diffusion similarity networks filtered for discriminative connections and passed to a graph neural network produce a Parkinson's-versus-essential-tremor decision and name the brain regions driving it. Without feature selection, diffusion similarity networks plus a graph neural network reached a balanced accuracy of 0.83 ± 0.09 and an AUC of 0.88 ± 0.05 in five-fold cross-validation, and random forest with feature selection alone reached 0.79 ± 0.02 and 0.81 ± 0.02. The full pipeline reached a balanced accuracy of 0.87 ± 0.05, an F1 of 0.86 ± 0.05 and an AUC of 0.94 ± 0.03 in five-fold cross-validation on the 100-person local cohort (57 with Parkinson's disease, 43 with essential tremor), and 87.5% balanced accuracy on a separate 40-person hold-out set (20 with Parkinson's disease, 20 with essential tremor).
The longitudinal analysis places the rate of network decay and the rate of clinical motor-score accumulation in one two-dimensional velocity space and splits patients into stable and rapid-progression classes by clinical trajectory. A standard similarity metric still returns a high score when two neighboring regions degrade in parallel, so it cannot reveal the actual tissue decline. Across four years in 41 patients with Parkinson's disease, the rate of network decay correlated significantly with the rate of motor-score accumulation in the bilateral supplementary motor areas (right r = +0.3445, p = 0.0274; left r = +0.3228, p = 0.0395) and primary motor cortices (right r = +0.3100, p = 0.0486; left r = +0.3090, p = 0.0493).
Perspective
The demonstrated scope covers cross-sectional and four-year longitudinal analyses of 41 people with Parkinson's disease and 66 healthy controls from the Parkinson's Progression Markers Initiative database, plus Parkinson's-versus-essential-tremor binary classification on a local cohort of 100 people; the authors propose future validation of the longitudinal velocity models on larger, more heterogeneous datasets and extension to a multi-class model including atypical parkinsonian variants such as multiple system atrophy and progressive supranuclear palsy.
The longitudinal analysis is restricted to a highly curated sub-cohort of 41 patients, chosen to ensure age-matched controls and complete four-session follow-up, which reduces the sample available for progression modeling; the structural degradation term in the combined similarity score is measured against a group-level healthy baseline rather than each patient's own first scan; and the classification results still need validation in an external cohort.