Levodopa-induced Dyskinesia in Parkinson’s Disease: Insights from cross-cohort prognostic analysis using machine learning#

Authors#

Rebecca Ting Jiin Loo, Olena Tsurkalenko, Jochen Klucken, Enrico Glaab

Abstract#

Background Prolonged levodopa treatment, commonly prescribed to alleviate motor symptoms of Parkinson’s disease (PD), often leads to adverse effects, including involuntary adventitious movements known as levodopa-induced dyskinesia (LID). While most PD patients develop LID during later disease stages, some patients remain free from this symptom despite receiving continuous levodopa treatment. The key factors distinguishing affected from unaffected treated patients are still largely unknown. Objective This study explores the application of machine learning in predicting LID development in PD patients over a 4-year period using baseline clinical data to discover the relevant clinical biomarkers. Methods We use data from three large-scale, longitudinal PD cohorts to build models for cross-cohort prognosis of LID to discover the potential risk and protective factors for LID development using baseline clinical assessments. Apart from investigating predictive factors for LID in each cohort separately, we examine the differences and similarities between the cohorts to identify the most robust shared predictors and assess cohort-specific variability and confounding factors. Furthermore, we critically assess the benefits and limitations of cross-study normalization approaches. With a focus on interpretable tree-based machine learning models, we compare multiple modeling approaches for LID prognosis and time-to-LID. Results Cross-validation results using multiple performance metrics indicate that individual clinical features tend to have only a moderate association with dyskinesia development, but their combination can provide significant prognostic information. Conclusion Overall, this study provides robust explainable machine learning models for LID prognosis and time-to-LID to define the risk factors and the association in multi-cohorts. It may help to lay the ground for earlier and more effective interventions against LID development in PD.