Development of a Stacking-Based Ensemble Machine Learning for Detection of Depression in Parkinson’s Disease: Preliminary Research

This preliminary study used the stacking ensemble to explore the major elements (factors) which could predict depression in patients with Parkinson’s disease and presented baseline data for developing a nomogram prognostic index for predicting high-risk groups for depression among patients with Parkinson’s disease in the future. Depression, an outcome variable, was divided into “with depression” and “without depression” using the Geriatric Depression Scale-30 (GDS-30). This study developed nine machine learning models (ANN, random forest, naive bayes, CART, ANN+LR, random forest+LR, naive bayes+LR, CART+LR, and random forest+naive bayes+CART+ANN+LR). The predictive performance (e.g., REMS, IA, Ev) of each machine learning model was validated through 10-fold cross-validation. The analysis results showed that the random forest+LR had the best predictive performance: RMSE = 0.16, IA = 0.73, and Ev = 0.48. This study analyzed the normalized importance of the random forest+LR model’s variables (the final model) and confirmed that K-MMSE, K-MoCA, Global CDR, sum of boxes in CDR, total score of UPDRS, motor score of UPDRS, K-IADL, H and Y staging, Schwab and England ADL, and REM and RBD were ten major variables with high weight among predictors of Parkinson’s disease with depression in South Korea. It is necessary as well to develop interpretable machine learning to build a model for predicting depression in patients with Parkinson’s disease that can be used in the medical field.