Multi-Objective Optimization for Personalized Prediction of Venous Thromboembolism in Ovarian Cancer Patients

Abstract

Thrombotic events are one of the leading causes of mortality and morbidity related to cancer, with ovarian cancer having one of the highest incidence rates. The need to prevent these events through the prescription of adequate schemes of antithrombotic prophylaxis has motivated the development of models that aid the identification of patients at higher risk of thrombotic events with lethal consequences. However, antithrombotic prophylaxis increases the risk of bleeding and this risk depends on the class and intensity of the chosen antithrombotic prophylactic scheme, the clinical and personal condition of the patient and the disease characteristics. Moreover, the datasets used to obtain current models are imbalanced, i.e., they incorporate more patients who did not suffer thrombotic events than patients who experienced them what can lead to wrong predictions, especially for the clinically relevant patient group at high risk of thrombosis. Herein, predictive models based on machine learning were developed utilizing 121 high-grade serous ovarian carcinoma patients, considering the clinical variables of the patients and those typical of the disease. To properly manage the data imbalance, cost-sensitive classification together with multi-objective optimization was performed considering different combinations of metrics. In this way, five Pareto fronts and a series of optimal models with different false positive and false negative rates were obtained. With this novel approach to the development of clinical predictive models, personalized models can be developed, helping the clinician to achieve a better balance between the risk of bleeding and the risk of thrombosis.

Thrombotic events are one of the leading causes of mortality and morbidity related to cancer, with ovarian cancer having one of the highest incidence rates. The need to prevent these events through the prescription of adequate schemes of antithrombotic prophylaxis has motivated the development of models that aid the identification of patients at higher risk of thrombotic events with lethal consequences. However, antithrombotic prophylaxis increases the risk of bleeding and this risk depends on the class and intensity of the chosen antithrombotic prophylactic scheme, the clinical and personal condition of the patient and the disease characteristics. Moreover, the datasets used to obtain current models are imbalanced, i.e., they incorporate more patients who did not suffer thrombotic events than patients who experienced them what can lead to wrong predictions, especially for the clinically relevant patient group at high risk of thrombosis. Herein, predictive models based on machine learning were developed utilizing 121 high-grade serous ovarian carcinoma patients, considering the clinical variables of the patients and those typical of the disease. To properly manage the data imbalance, cost-sensitive classification together with multi-objective optimization was performed considering different combinations of metrics. In this way, five Pareto fronts and a series of optimal models with different false positive and false negative rates were obtained. With this novel approach to the development of clinical predictive models, personalized models can be developed, helping the clinician to achieve a better balance between the risk of bleeding and the risk of thrombosis.