Risk Algorithm Using Serial Biomarker Measurements Doubles the Number of Screen-Detected Cancers Compared With a Single-Threshold Rule in the United Kingdom Collaborative Trial of Ovarian Cancer Screening

Abstract

Purpose: Cancer screening strategies have commonly adopted single-biomarker thresholds to identify abnormality. We investigated the impact of serial biomarker change interpreted through a risk algorithm on cancer detection rates.

Patients and methods: In the United Kingdom Collaborative Trial of Ovarian Cancer Screening, 46,237 women, age 50 years or older underwent incidence screening by using the multimodal strategy (MMS) in which annual serum cancer antigen 125 (CA-125) was interpreted with the risk of ovarian cancer algorithm (ROCA). Women were triaged by the ROCA: normal risk, returned to annual screening; intermediate risk, repeat CA-125; and elevated risk, repeat CA-125 and transvaginal ultrasound. Women with persistently increased risk were clinically evaluated. All participants were followed through national cancer and/or death registries. Performance characteristics of a single-threshold rule and the ROCA were compared by using receiver operating characteristic curves.

Results: After 296,911 women-years of annual incidence screening, 640 women underwent surgery. Of those, 133 had primary invasive epithelial ovarian or tubal cancers (iEOCs). In all, 22 interval iEOCs occurred within 1 year of screening, of which one was detected by ROCA but was managed conservatively after clinical assessment. The sensitivity and specificity of MMS for detection of iEOCs were 85.8% (95% CI, 79.3% to 90.9%) and 99.8% (95% CI, 99.8% to 99.8%), respectively, with 4.8 surgeries per iEOC. ROCA alone detected 87.1% (135 of 155) of the iEOCs. Using fixed CA-125 cutoffs at the last annual screen of more than 35, more than 30, and more than 22 U/mL would have identified 41.3% (64 of 155), 48.4% (75 of 155), and 66.5% (103 of 155), respectively. The area under the curve for ROCA (0.915) was significantly (P = .0027) higher than that for a single-threshold rule (0.869).

Conclusion: Screening by using ROCA doubled the number of screen-detected iEOCs compared with a fixed cutoff. In the context of cancer screening, reliance on predefined single-threshold rules may result in biomarkers of value being discarded.

Fig 1.

CONSORT diagram.

Fig 2.

Multimodal screening algorithm and outcome of incidence screening. A, abnormal; CA-125, cancer antigen 125; CE, clinical evaluation; E, elevated; I, intermediate; LI, level I CA-125 test; LII, level II transvaginal scan (TVS) and CA-125 test; N, normal; ROC, risk of ovarian cancer; S, severe; SD, screening discontinued; U, unsatisfactory.

Fig 3.

Plot of all multimodal screening annual cancer antigen 125 (CA-125) measurements over time on a log scale, including truncation. Superimposed are the serial measurements for 155 invasive epithelial ovarian and/or tubal cancers with the large circles representing the final screen before diagnosis, either true positive (n = 133; gold lines and markers) or false negative (n = 22; blue lines and markers). The red line indicates one patient in whom the risk of ovarian cancer algorithm recommended surgery, but it was not performed following clinical evaluation. The black horizontal lines represent CA-125 cutoffs of 35, 30, and 22 U/mL. NOTE. 262 CA-125 values truncated above 100 U/mL and 174 CA-125 values truncated below 2 U/mL.

Fig 4.

Risk of ovarian cancer (ROC) curves based on the performance characteristics of annual cancer antigen 125 (CA125) measurement alone and annual risk of ovarian cancer algorithm (ROCA) score alone. Overlaid points represent the actual characteristics of the multimodal screening strategy, hypothetical characteristics of annual ROCA classified as normal or abnormal (intermediate/elevated) risk, hypothetical characteristics of annual CA125 using the cutoff points of more than 35 U/mL (as in the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial), more than 30 U/mL (in clinical use), and more than 22 U/mL (as suggested by other groups), respectively. P value of .0027 is test of difference.