Introduction/Background Patients with ovarian cancer often exhibit abnormal blood components prior to diagnosis, including those measured in routine blood tests such as complete blood counts (CBCs). We leveraged a robust infrastructure of provincial electronic health data, when measured in healthy individuals, to predict the presence of occult (or preclinical) ovarian cancer using routine laboratory tests.

Methodology Using population-level data from Ontario, Canada, we identified a cohort of women with a history of laboratory testing between January 2007 and December 2020. We included adults aged 40 to 75 with two consecutive routine CBC records ordered 9 to 15 months apart by their family doctor. We retained CBC values as well as change in CBC values, and followed women for 5 years for an ovarian cancer diagnosis. We developed a random forest machine learning model training on 80% of our original data and testing on the remaining 20%.

Results There were 1,331,671 eligible women in the cohort, of which 2,887 developed ovarian cancer in the follow-up period. Several variables were strongly associated with an ovarian cancer diagnosis; features that were most predictive include absolute and relative changes in platelet count, lymphocyte count, and neutrophil count (Figure 1). Overall model performance decreased from time from CBC to cancer diagnosis, when looking at follow-up intervals of <1 year, 1 to <3 years, and 3 to 5 years. Future analyses will incorporate features for health services utilization patterns, comorbidities, chronic conditions, and genetic risk factors.

• Download figure
• Open in new tab
• Download powerpoint

Abstract 1131 Figure 1

Conclusion Our study demonstrates the potential of using routine blood tests, specifically platelet, lymphocyte, and neutrophil counts, as early indicators of ovarian cancer risk.

Disclosures The authors have no disclosures to declare.