Defining the Clinical Phenotypic Ground Truth: Expert Interrater Agreement in Hyperkinetic Movement Disorders.

Abstract

Background: Reliable clinical phenotyping is crucial for diagnosis, treatment, and research in hyperkinetic movement disorders, yet its interrater reliability remains uncertain.

Objectives: To quantify interrater agreement in clinical phenotyping of hyperkinetic movement disorders among experts in a large, phenotypically diverse cohort.

Methods: Within the NEMO project, 161 patients with hyperkinetic movement disorders underwent standardized video assessments. Each patient was independently classified by three international movement disorders experts in a structured, stepwise procedure. Experts first assessed videos alone, then received clinical history and examination, and finally diagnostic test results. At each step, they assigned primary and secondary phenotypes and rated their diagnostic confidence. Interrater agreement was assessed using Fleiss' and Cohen's kappa.

Results: Across all phenotypes, interrater agreement on the primary phenotype increased with added information (κ = 0.457 to κ = 0.607). Agreement with the NEMO reference classification similarly improved (κ = 0.478 to κ = 0.656). Among patients with a single dominant phenotype, agreement reached κ = 0.78. In 93% of cases, at least two out of three experts reached consensus. Agreement was substantially lower in patients with mixed or complex presentations, such as myoclonus-dystonia and functional movement disorders. Diagnostic certainty rose throughout, though secondary phenotype agreement remained poor. Experts rarely revised initial impressions, suggesting robust pattern recognition but also potential confirmation bias.

Conclusions: Movement disorders specialists can reliably identify hyperkinetic phenotypes, even without in-person assessment. While complex, mixed and functional presentations remain diagnostically challenging, our findings reinforce the value of expert phenotyping in clinical care, training, and research, and provide a benchmark for future studies and computational approaches.