Improving ovarian cancer risk assessment using a machine learning model developed on data from the Prostate Lung Colorectal Ovarian Cancer Screening Trial.

Authors

Sarayi SMMJ, Tammemägi M, Meyer LA, Toumazis I

Affiliations

• The University of Texas MD Anderson Cancer Center, Department of Health Services Research, Houston, TX, USA.
• Brock University, Department of Health Sciences, St Catharines, ON, Canada.
• The University of Texas MD Anderson Cancer Center, Department of Gynecologic Oncology and Reproductive Medicine, Houston, TX, USA.
• The University of Texas MD Anderson Cancer Center, Department of Health Services Research, Houston, TX, USA. Electronic address: IToumazis@mdanderson.org.

Abstract

OBJECTIVE: Ovarian cancer screening for the general population is not effective at reducing mortality primarily due to the low incidence of ovarian cancer. Risk assessment could improve ovarian cancer prevention by targeting available and emerging interventions to high-risk individuals. This study aimed to develop an ovarian cancer risk prediction tool with sufficient predictive performance to guide ovarian cancer prevention.

METHODS: We used data from the PLCO screening trial to develop and validate a 10-year ovarian cancer risk model. We used PLCO's control arm to train the model and the intervention arm for validation. Potential predictors included sociodemographic factors, medical history, and female reproductive history, among others. We compared alternative machine learning algorithms for model development and assessed their performance using the area under the curve, sensitivities, specificities, and positive predictive values.

RESULTS: Extreme Gradient Boosting algorithm produced the best performing model, which consisted of 14 trees, each with a maximum of 3 layers (area under the curve 0.80, 95% confidence interval [CI] 0.78 to 0.81 on the PLCO control arm-training set). Seventeen readily available features were included in the final model, among which body mass index, age, and duration of female hormone use increased risk, whereas bilateral oophorectomy and number of live births decreased risk. The model's performance on the PLCO's intervention arm was satisfactory (area under the curve 0.66, 95% CI 0.64 to 0.68), outperforming previously published models when tested on the same validation dataset. At a 58% risk threshold the model yielded sensitivity of 75% (95% CI 72 to 79) and specificity of 70% (95% CI 69 to 70) on the training set. We assessed model's performance on the validation set at various risk thresholds.

CONCLUSIONS: The developed ovarian cancer incidence risk model demonstrated superior performance over existing models. Nevertheless, more research is warranted to further improve the predictive performance of the models and ensure feasibility of risk-based programs for ovarian cancer prevention.