Objective: To estimate medication effect, as the score difference between MDS-UPDRS III in OFF and ON-state, using a minimal set of clinical features.

Background: The MDS-UPDRS III is the most employed scale to assess motor progression in People with Parkinson’s Disease (PwPD) [1]. In this context, accurately separating the contribution of treatment on natural disease progression is crucial [2]. Estimating MDS-UPDRS III in OFF-state could allow for the quantification of treatment effect, as well as avoiding a medication washout during the clinical assessment [3].  ​

Method: This study employed data from the Parkinson’s Progression Markers Initiative [4]. We included PwPD followed up longitudinally that had MDS-UPDRS III scores in ON and OFF-state after at least a 6-hour medication washout. A random forest classifier was trained to identify and filter patients with abnormal treatment effect (e.g., cases where MDS-UPDRS III ON was higher than OFF). Remaining data was used to cross-validate a robust linear mixed-effect model (rLMM) to estimate MDS-UPDRS III OFF, with clinical features selected through a stepwise procedure. Post-modeling principal component analysis (PCA), associated with Kruskal–Wallis and Chi-square tests, investigated the difference across clinically significant under and overestimated groups (better or worse medication response than expected, respectively).

Results: Data from 275 PwPD (121 men, mean age: 60.3 ± 10.1 y, LEDD: 632.60. ± 495.84 mg/day, MDS-UPDRS III ON: 21.3 ± 12.3, MDS-UPDRS III OFF: 30.6 ± 13.5, Medication washout: 14.05 ± 9.55 hrs) were analyzed. The classifier’s testing accuracy was 0.94, and F1-score was 0.79. The rLMM employed four parameters: disease duration, ON-state score, L-DOPA equivalent daily dose and hours since last dosage. The model had a test R² of 0.65 and RMSE of 7.85 [Fig.1]. PCA revealed that two components explained 63% of the variance between over and underestimated groups, including 9 MDS-UPDRS III items driving separation between the groups [Tab.1].

Conclusion: This hierarchical, two-step approach quantifies treatment effect using readily available clinical data. The method may be employed to accurately estimate the natural progression or medication response in PwPD. Finally, clinical profiles associated with an enhanced or decreased effect of medication could provide valuable insights for treatment decision support.

Figure 1

Table 1

References: [1] Holden SK, Finseth T, Sillau SH, Berman BD. Progression of MDS-UPDRS Scores Over Five Years in De Novo Parkinson Disease from the Parkinson’s Progression Markers Initiative Cohort. Movement Disorders Clinical Practice. 2018;5(1):47–53.

[2] Sánchez‐Ferro Á, Matarazzo M, Martínez‐Martín P, Martínez‐Ávila JC, Gómez De La Cámara A, Giancardo L, et al. Minimal Clinically Important Difference for UPDRS‐III in Daily Practice. Movement Disord Clin Pract. 2018 Jul;5(4):448–50.

[3] Comparison of Motor Scores between OFF and ON States in Tremor-Dominant Parkinson’s Disease after MRgFUS Treatment [Internet]. Available from: https://doi.org/10.3390/jcm1115450

[4] Marek K, Chowdhury S, Siderowf A, Lasch S, Coffey CS, Caspell‐Garcia C, et al. The Parkinson’s progression markers initiative (PPMI) – establishing a PD biomarker cohort. Ann Clin Transl Neurol. 2018 Dec;5(12):1460–77.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/combining-machine-learning-and-linear-mixed-models-to-estimate-medication-effect-in-people-with-parkinsons-disease/