Ovarian mass triage without a fixed biomarker panel: adaptive plasma proteomics reaches a summary AUC of 0.835 across five validation cohorts

A plasma proteomic classifier can separate malignant from benign adnexal masses without developing targeted assays, and holds its discrimination across five independent cohorts including one from another institution. Previous triage relied on a few serum markers such as CA-125, HE4 and ROMA, plus the expert-dependent ultrasound O-RADS score, and these tools lose accuracy in early-stage and non-epithelial disease. Across 623 classifiable samples in five validation cohorts the summary AUC was 0.835, with a cohort range of 0.779 to 0.904; against 0.851 from five-fold cross-validation in the discovery cohort, the difference was +0.006 (95% CI, −0.050 to +0.062).

The classifier rebuilds its model from the proteins actually detected in each sample, so it needs no fixed protein panel and does not fail on missing values. Fixed-panel approaches require the same proteins to be measured in every sample, impute missing values, and are often optimized for a single histology. Each validation sample's model drew on a median of 121 proteins out of the 138 features seen across the validation cohorts, and none of the 623 samples was left unclassified for too few available features.

Proteomics and ultrasound O-RADS carry complementary information, and combining them discriminates better than either alone. Proteomics alone reached an AUC of 0.852, comparable to O-RADS alone at 0.863, indicating the two capture different parts of the signal. Among the 231 patients with pre-operative O-RADS, the combined model reached an AUC of 0.922; on the same matched samples proteomics reached 0.915 and serum CA-125 0.777 (n=125).

The demonstrated scope is a prospective observational study at two academic referral centers: a discovery cohort and four internal validation cohorts at the University of Chicago, plus an external validation cohort at Roswell Park Comprehensive Cancer Center that contains adnexal masses only and no healthy controls. The classification task is binary, benign versus malignant adnexal mass, and healthy controls take part in neither training nor evaluation. The authors propose prospective validation in community settings and concurrent assessment with CA-125, HE4 and standardized ultrasound models such as O-RADS and ADNEX.

The sequential classifier assumes each earlier case's pathology-confirmed outcome is available before the next case is scored, whereas in practice patients predicted benign are less likely to receive prompt pathological confirmation; the authors expect real-world performance to fall between the static basic classifier and the fully label-informed model. Comparisons with O-RADS and CA-125 were limited to subsets with available data that differed in size and case mix; HE4 was not routinely measured at either institution, so direct comparison with the clinical ROMA algorithm was not possible, and the proteome-derived ROMA-like model is only an approximation. Subgroups such as non-epithelial malignancies and BRCA carriers were small, and prediction for early FIGO stage was lower, with an AUC of 0.712 versus stages II to IV.