Serum-based assay for adnexal mass risk of ovarian malignancy.

Abstract

5551 Background: A Deep learning neural network was developed to assess ovarian cancer risk in women presenting with adnexal mass into risk categories. The algorithm shows potential to improve on the performance of CA-125 as the standard biomarker to monitor women as a clinical management metric to trace increased risk of malignancy. Methods: Serum specimens from an enriched biobank (N = 2688) were collected during previous clinical studies from women presenting with an adnexal mass at risk of ovarian cancer. Protein biomarker data from these specimens were used to develop a novel neural network for binary stratification of ovarian cancer risk assessment. These specimens were divided between training and testing data sets using 5-fold cross validation during training. A randomized, separate sample set was withheld from use in training and testing the algorithm for independent validation purposes. As algorithm inputs seven biomarkers are used in the network: cancer antigen 125, human epididymis protein, beta-2 microglobulin, apolipoprotein A-1, transferrin, transthyretin, and follicle stimulating hormone. In addition to these biomarkers, the patient’s age and menopausal status are used. Menopause was defined as the absence of menses for ≥12 months. The algorithm uses supervised learning with known histopathology diagnoses as the labels for training. The algorithm is a classification deep feedforward neural network. The neural network is regularized using node dropout to reduce overfitting. The final layer of the neural network has two nodes and uses the softmax function to assign a binary classification of low or high-risk of malignancy. Results: Algorithm performance metrics are also shown comparing predicted results from the algorithm to the known malignancy diagnoses. The performance metrics are also compared below to the standard of care biomarker test, cancer antigen 125 (CA125), reporting increased sensitivity by 26.1%, and failure to reject the null hypothesis of equivalent specificity. Conclusions: The algorithm detected 91% of malignancies in the independent validation data. This high sensitivity in malignancy detection paired with the failure to reject the null hypothesis of equivalent specificity (Pearson’s chi-squared test p-value of 0.281) and negative predictive value (NPV) suggest the algorithm could be used two-fold. First, surgical referral to gynecological oncologists for women classified in the high-risk cohort. The second as a goal with future clinical validation, is that women with a low risk of malignancy might be able to delay surgery and enter into a serial monitoring clinical management care pathway.[Table: see text]