Fig. 8

Exploration of diagnostic biomarkers using machine learning methods and generation of machine learning-based classifiers. a Schematic of the learning framework for a limited sample size and unbalanced datasets. This framework consists of three steps: oversampling, parameter searches, and feature selection. b, c Tables and heatmaps showing the accuracy of the trained models according to protein composition (sEVs: b and cEVs: c). The test dataset contains an equal number of samples from controls and SALS patients (ROPI-naive). The ROPI-naive dataset consists of samples collected from SALS patients before they received ROPI, and the ROPI-exposed dataset consists of samples taken from SALS patients after they received ROPI. Accuracy is shown by %, and the average accuracy for the ROPI-naive and ROPI-exposed datasets is shown in the heatmap