Fig. 3.

(A) Accuracy of diagnostic sMRI ML studies over time and sample size (circle increases with sample size). From the first study until 2015, the vast majority of studies reported accuracies ranging between 70% and 100%; from 2016, however, performances have dropped overall with accuracies ranging between chance-level and 85%. (B) Funnel plot for sMRI studies in schizophrenia and FEP showing the distribution of individual studies according to their sample size (1/√ESS) and effect size (log diagnostic odds ratio). The plot revealed statistically significant asymmetric distribution around the main effect of sMRI studies (P = .013), indicating a bias favoring higher effect sizes. sMRI, structural MRI; ML, machine learning; FEP, first episode psychosis.