パーキンソン病の運動症状重症度を評価する、複雑性コンセンサスを考慮した自動評価システムPARQ

Parkinson's Disease (PD) is the fastest growing neurodegenerative disease, creating an urgent need for scalable objective clinical approaches to assess and monitor cardinal PD features (bradykinesia, rigidity, tremor, and gait dysfunction). Currently, the evaluation of PD features relies on expert assessments using the Movement Disorders Society - Unified Parkinson's Disease Rating Scale part III (MDS-UPDRS-III ), which provides a standardized framework for quantifying motor disease severity. However, this approach relies on the availability and expertise of trained clinicians, limiting scalability and introducing variability related to the experience of the raters and task-specific complexity. Inter-rater variability arises from a fundamental consensus-complexity tradeoff inherent to visually rated motor tasks: tasks with less visual cues tend to elicit lower inter-rater variability (greater rater agreement) whereas tasks with more visual cues are associated with greater disagreement. To address these challenges, we introduce PARQ, a deep learning based platform to automatically quantify PD motor disease severity from routine clinical videos. Rather than predicting a single severity score, PARQ estimates both the expected motor severity and the underlying distribution of expert ratings, enabling task-specific, consensus-aware motor severity prediction. We evaluate PARQ on a clinical dataset of 48 patients across 8 tasks with three independent expert ratings per video. PARQ achieves 90% accuracy on high-consensus tasks and 80% on low-consensus tasks, demonstrating robustness to systematic rater disagreement. PARQ delivers task-specific, distribution-aware severity estimates across most visually rated MDS-UPDRS-III tasks, offering a foundation for objective PD motor disease severity assessment.