Fig. 3

Model selection process and performance of predicting “cases” vs “controls”. The logistic regression model selection process with the best model fit by data-driven selection with the lowest AIC. The parsimonious model shows the model that had a similar performance (ΔAIC < 2) with as few significant predictors as possible. In subsequent models, predictors were removed in a stepwise procedure. Comparisons between AUCs were performed using DeLong statistics, *p < 0.05, **p < 0.010, ***p < 0.001 compared to the best model fit. ROC curves combining different models are plotted on the right side of the figure. AUC, area under the curve; AIC, Akaike Information Criterion; HC, healthy controls; NfL, neurofilament light; GFAP, glial fibrillary acidic protein; SICI, average short-interval intracortical inhibition (1, 2, 3 ms); ICF, average intracortical facilitation (7, 10, 15 ms); SAI, average short-latency afferent inhibition (0, + 4 ms) expressed as the ratio of the unconditioned motor evoked potential (MEP)