Abstract
BACKGROUND: GBA1 variants are the major genetic risk factor for Parkinson's Disease (PD) and account for 5-30% of PD cases depending on the population and age at onset of the disease. OBJECTIVES: The aim of this study was to assess whether Artificial Intelligence (AI) could predict GBA1-mutated genotype in PD (GBA1-PD). Particularly, the main objective was to identify a Machine Learning (ML) model capable of accurately providing a pre-test estimate of GBA1-mutated status, relying on the clinical and demographic variables with the highest predictive value. METHODS: A cohort of GBA1-PD patients has been paired with non-mutated PD (NM-PD). The dataset comprised 58 GBA1-PD and 58 NM-PD, for each of whom 124 features were recorded. A Leave-One-Out cross-validation method was employed for testing and SHapley Additive exPlanations (SHAP) for examine each feature's contribution. XGBoost resulted the most effective ML model for this supervised classification task. RESULTS: Through AI, we developed a model based on four specific clinical features with significant impact in predicting GBA1-mutated genotype with an accuracy of 73%, reaching 94% in a subset of patients where the model has a SHAP confidence level greater than 80%. These variables included family history and scores for cognitive (MDS-UPDRS 1.1) and motor impairment (MDS-UPDRS 3.8a and 3.8b and rigidity subscore). CONCLUSIONS: This study underlies the potential of AI in enhancing targeted genetic screening in PD, especially in clinical settings where resources are limited. Main limitations of this study are the modest sample size and lack of external validation. Further studies on larger, independent cohorts are needed to refine the predictive model.