A statistical methodology to improve accuracy in differentiating schizophrenia patients from healthy controls

We present a methodology to statistically discriminate among univariate and multivariate indices to improve accuracy in differentiating schizophrenia patients from healthy controls. Electroencephalogram data from 71 subjects (37 controls/34 patients) were analyzed. Data included P300 event-related response amplitudes and latencies as well as amplitudes and sensory gating indices derived from the P50, N100, and P200 auditory-evoked responses resulting in 20 indices analyzed. Receiver operator characteristic (ROC) curve analyses identified significant univariate indices; these underwent principal component analysis (PCA). Logistic regression of PCA components created a multivariate composite used in the final ROC. Eleven univariate ROCs were significant with area under the curve (AUC) >0.50. PCA of these indices resulted in a three-factor solution accounting for 76.96% of the variance. The first factor was defined primarily by P200 and P300 amplitudes, the second by P50 ratio and difference scores, and the third by P300 latency. ROC analysis using the logistic regression composite resulted in an AUC of 0.793 (0.06), p<0.001 (CI=0.685-0.901). A composite score of 0.456 had a sensitivity of 0.829 (correctly identifying schizophrenia patients) and a specificity of 0.703 (correctly identifying healthy controls). Results demonstrated the usefulness of combined statistical techniques in creating a multivariate composite that improves diagnostic accuracy.