Fig. 2

Models developed from multi-site data suggest peripheral biomarkers for LBD. A The model performance suggested good separation for HC vs. LBD, but not for HC vs. ADD. Note that a random guess baseline would yield an AUROC of 0.50 and an AUPRC equivalent to the prevalence of the positive class in the sample group, which are shown as patterned gray bars. B Performance using cross-site splitting instead of random cross-validation suggests the generalizability of the biomarkers. C Transferring the HC vs. LBD model (without retraining) to classify LBD from disease controls, including ADD, NDC, and IDC, yielded similarly high performance. D The predicted values from the HC vs. LBD model for all diagnosis groups show that the model is LBD-specific. E Model residual (errors from each prediction) did not significantly (M.W.U. P < 0.05) vary with sex, age, Levodopa dosage, APOE e4 status, or PD vs. PDD. This indicates that the model performed equally well across these variables. In contrast, the model’s residual varied for the DLB vs. PD/PDD group, suggesting that the performance of the DLB group differed from the PD/PDD group. F The required number of top immune features needed to achieve similar performance as all 1,184 features. G Correlation network highlighting the top features and the immune features with which they are correlated