Fig. 1. Illustration of workflows of this study.

Fecal samples and clinical features were obtained at year 0 in 224 PD patients. At year 2, clinical features were evaluated in 165 PD patients, and fecal samples were obtained in 104 PD patients. Although clinical features were evaluated at year 2 in 17 additional PD patients (a total of 182 patients), they were excluded from making prediction models (see Materials and Methods). Similarly, fecal samples were obtained at year 2 in nine additional PD patients (a total of 113 patients), but they were excluded from analysis of temporal profiles of gut microbiota. Bacterial and clinical features at year 0, as well as clinical features at year 2, in 165 PD patients were used for Figs. 2 and 3. Gut microbiota at years 0 and 2 in 104 PD patients were used for Fig. 4. Construction of prediction models was constituted of two steps: (1) nested cross-validation has no leakage between the training and test datasets, and is for evaluation of the modeling strategy, and (2) cross-validation has marginal leakage between the training and test datasets, and is for determination of essential features to predict the progression of PD. Recursive feature elimination (FRE) was employed in both steps. AUROCs were calculated in both steps, but AUROC of the nested cross-validation should be dependable because of lack of potential leakage.