Fig. 2

Logistic regression model output of multiple sclerosis versus non-multiple sclerosis. Illustration of the model expressed according to best combination of A CSF, B serum, and C combined CSF and serum biomarkers. Top left: violin plot illustrating the predicted probability of multiple sclerosis (y-axis) by cohort (x-axis), according to the optimum biomarker model adjusted for sex and age. The width of the violin plots indicates the relative number of probabilities at the probability. Horizontal dashed line illustrates an arbitrary cut-off value that could be applied to distinguish MS from non-MS cases, where red dots above the line indicate false positives and blue dots below the line false negatives. Top right: receiver operator curve illustrating performance of the optimum model. Bottom right: regression table indicating the magnitude and direction of effect of each variable. Note: a cross-validation approach was taken whereby the cohort was randomly divided into four roughly equally sized and distinct groups (see Methods), which were separately used as “training datasets” to produce a model using roughly 75% of the data, which was then tested on the remaining approximately 25% of the data (the “test dataset”). This was repeated 4 times to generate a mean test AUC to determine the optimum combination of biomarkers. The ROC curve, regression table and AUCs shown here were produced using one of the train/test sub-cohorts