Does impulse control disorder link to dyskinesia through a cortical network in Parkinson's disease?

A longitudinal test of two linked complications

ICD in Parkinson's disease covers impulsive-compulsive behaviors such as pathological gambling, compulsive buying, compulsive sexual behavior, and binge or compulsive eating. LID is the most common motor complication of levodopa therapy. Both have been tied to dopamine treatment and to frontostriatal dysfunction, but earlier evidence linking them came mainly from cross-sectional and descriptive designs.

This study asked whether ICD present at baseline predicts later LID, and whether a cortical network explains part of that link. It followed 464 patients with early Parkinson's disease from the Parkinson's Progression Markers Initiative for more than four years, excluding anyone with LID symptoms at baseline or a disease duration over three years. ICD was defined as any score above zero on the QUIP-CS at baseline; 110 participants (23.7%) met that definition.

ICD raised LID risk beyond putaminal dopamine loss

LID developed more often in the ICD group (chi-square = 4.3111, p = 0.0379). The unadjusted hazard ratio for LID was 1.3476 (95% CI 1.0317–1.7603, p = 0.0329). After adjustment for age, sex, and putaminal dopamine transporter availability, the hazard ratio was 1.3375 (95% CI 1.0393–1.7213, p = 0.0238).

The adjustment matters because putaminal dopamine transporter availability is a marker of dorsal striatal dopaminergic loss, the territory most associated with LID. Both groups had similar caudate and putaminal transporter availability at baseline, so the excess LID risk in the ICD group was not explained by that measure. The authors read this as pointing toward cortical or network-level mechanisms beyond dopaminergic loss, while noting that striatal subregions were not finely separated.

How morphology became a connectivity measure

The cortical analysis used morphological similarity networks built from structural MRI. Each of 68 Desikan-Killiany regions became a node, and links represented how similar regions were across seven morphological measures, including cortical thickness, surface area, gray matter volume, and curvature indices. Links were weighted by Pearson correlations, so a strong link means two regions share similar architecture rather than that they are directly wired together.

Among the 464 participants, 419 had usable structural scans after 31 were excluded for preprocessing errors. The authors compared positive-link and negative-link networks separately and used threshold-free network-based statistics with permutation testing, adjusting for age and sex, to find links that differed between ICD and non-ICD groups.

A 34-node subnetwork was weaker in ICD patients

In the positive-link network, patients with ICD showed a subnetwork of 34 nodes and 41 links with significantly weaker connections. Mean connection strength across that subnetwork was lower in the ICD group (t = −6.3494, p < 0.0001). No significant subnetwork appeared for the opposite contrast, and negative-link networks showed no significant differences in either direction.

Hub regions, defined as nodes whose degree exceeded the subnetwork mean by more than one standard deviation, were the bilateral paracentral lobules and the left caudal anterior cingulate cortex. The subnetwork contained no direct links among these hubs, so their shared involvement reflects a broader disrupted network rather than a simple three-region circuit.

The left paracentral lobule carried the mediating signal

Within the hub regions, the left caudal anterior cingulate cortex and left paracentral lobule showed lower nodal strength, regional efficiency, clustering coefficient, and local efficiency in the ICD group after adjustment for age and sex. The right paracentral lobule showed no significant differences. The left caudal anterior cingulate cortex also had a lower within-module z-score, a measure of how strongly a region integrates within its own module.

The decisive result came from survival modeling of topological measures. Only the within-module z-score of the left paracentral lobule was associated with later LID (HR = 0.857, 95% CI 0.755–0.973, p = 0.017); other measures of the three hubs were not. Notably, that z-score did not differ significantly between ICD and non-ICD groups, which the authors interpret as local integration being a downstream expression of network disruption rather than a direct reflection of impulsivity.

Mediation held for the network, not for dopamine imaging

Causal mediation analysis tested whether the ICD-related network disruption influenced LID through the left paracentral lobule's within-module z-score, adjusting for age and sex. The average causal mediation effect was 0.0136 (95% CI 0.0028–0.0328), while the average direct effect (0.0132, 95% CI −0.0778 to 0.1110) and total effect (0.0268, 95% CI −0.0657 to 0.1286) were not significant. The mediation estimate held when a complementary log-log model with a 7.97-year outcome cutoff was substituted.

A parallel analysis using putaminal dopamine transporter availability as the exposure found no significant mediation through the same z-score (ACME −0.0005244, 95% CI −0.0024 to 0.0005), although its direct and total effects were significant. The contrast is the paper's central comparison: the cortical topological measure statistically carried the ICD–LID association, while the dopaminergic imaging measure did not. Because only the indirect path reached significance, the authors describe this as a statistical mediation pattern consistent with a shared cortical substrate, not as proof of a causal chain.

An early-stage cohort with a specific exposure window

The evidence comes from a retrospective observational analysis of a cohort with early, untreated Parkinson's disease, followed at 3-month intervals in year one and 6-month intervals afterward. ICD was defined by any nonzero QUIP-CS score, a broad threshold that does not distinguish gambling from compulsive eating or capture severity. LID was identified from MDS-UPDRS dyskinesia items. The authors also note that dopamine agonist doses did not differ significantly by ICD status in PPMI follow-up, and that doses may be reduced after ICD is recognized, so cumulative agonist exposure may not fully explain the excess LID risk.

The mediation result was robust to mild unmeasured confounding in sensitivity analysis, but the study cannot rule out shared genetic vulnerability; the discussion cites prior heritability estimates and candidate genes for ICD and LID as a plausible alternative or complementary explanation. The left-sided lateralization also needs systematic confirmation, given variable laterality findings in the LID literature.

ICD history belongs in LID risk assessment

The paper's actionable claim is narrower than a general call for network biomarkers. In this cohort, a baseline QUIP-CS score above zero identified patients at higher LID risk even when putaminal dopamine transporter availability was similar, and the left paracentral lobule's within-module z-score was the one topological measure that both tracked later LID and statistically mediated the ICD–LID link. That makes ICD history, rather than striatal imaging alone, the variable the authors argue should inform LID treatment planning.

The priority for readers is therefore a comparison: when assessing LID risk in early Parkinson's disease, an ICD history adds information that putaminal dopamine transporter availability did not capture here. The cortical network finding explains why that history may matter, but it remains a group-level statistical pattern from one cohort, not a validated individual prediction tool.

Sources used in this answer