Figure 4. ROC curves for neural network analysis compared to CA-125.

The neural network (AUC 0.93; 95% CI 0.88–0.97) significantly outperformed CA125 (AUC 0.74; 95% CI 0.65–0.83) in terms of overall operating characteristics (p=0.001).

Figure 4—figure supplement 1. Correlations between the miRNAs (vertical axes) of the neural network and CA-125 (horizontal axes) in the cancer (red markers) and benign/borderline/control (blue markers) groups.

(a) miR-23b (b) miR-29a (c) miR-32 (d) miR-320d (e) miR-1246 (f) miR-92a (g) miR-150 (h) miR-200a (i) miR305 (j) miR-1307 (k) miR-200c (l) miR-203a (m) miR-320c (n) miR-450b. None of the correlations were significant in either the training or testing set.

Figure 4—figure supplement 2. Performance of a two-tiered algorithm for ovarian cancer diagnosis incorporating both the neural network (NN) and a CA-125 cut-off of 35 U/ml.

Subjecting all negative neural network algorithm results to a second review with CA-125 would increase the probability of a false positive test result from 4.2% (5/120) to 19.2% (23/120) and a false negative rate from 5.8% (7/120) to 13.3% (16/120). If the tests were considered hierarchical so that only samples classified as negative by the neural network were then examined by CA-125, this would identify three additional cases of invasive cancer but at the expense of 19 additional false positive results. FP – false positive, TP – true positive, FN – false negative, TN – true negative.