Leveraging artificial intelligence in the early detection of ovarian cancer: Development and validation of a risk prediction model.

OBJECTIVE: To develop and validate a predictive tool to estimate the risk of ovarian cancer in a screening setting. METHODS: Machine learning was leveraged to create a predictive model for ovarian cancer diagnosis within one year of screening. Data from the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial were used in model creation. Participants in this trial were comprised of females aged 55-74 years who underwent annual screening and prospective follow up for development of ovarian cancer. Stratified 15-fold cross validation was performed followed by external validation on a retrospective cohort. RESULTS: A random forest model was created using data from 150,907 screening events with a mean age of 62.5 ± 5.6 years. Ovarian cancer was diagnosed in 112 participants within one year of screening. The primary model was strongly predictive of ovarian cancer diagnosis, noting AUC of 0.930 (95% CI 0.886-0.974), sensitivity of 77.7% (66.4-88.9%), specificity of 98.0% (97.9-98.2%), positive predictive value of 2.9% (2.4-3.4%), negative predictive value of 99.9% (99.98-99.99%), and a positive likelihood ratio of 38.9. This model retained strong predictive value in external validation with AUC 0.910, sensitivity of 91.9%, and specificity of 72.6%, positive predictive value 29.6%, and negative predictive value of 98.6%. CONCLUSIONS: This model utilizes artificial intelligence to predict the risk of development of ovarian cancer within one year of screening demonstrating high levels of discrimination, sensitivity, and specificity. Machine learning techniques should be considered in the development of a multimodal approach to the early detection and prevention of ovarian cancer.