Optimal Prescriptive Treatments for Ovarian Cancer with Genetic Data
Among the cancers affecting the female population, Ovarian Cancer (OC), while being relatively rare, is the leading cause of gynecological cancer-related deaths, with overall 5-year survival rates of approximately 50% for all stages combined. This is because of the challenges associated with the diagnosis, resulting in detection at advanced stages of OC, coupled with the slow progress in effective treatment options since the approval of platinum-based chemotherapy in the late 1970s. There has been a relative lack of sophisticated methods based on Machine Learning (ML) models that use genetic data for better prediction of Ovarian Cancer outcomes and result in more effective treatment recommendations. Therefore, there is an unmet clinical need to create models that allow physicians to make informed decisions based on all available data, including patient demographic, social, health, and genomic data. Hence, we develop new techniques for leveraging genetic information in prescribing optimal treatments for patients with OC, using a publicly available dataset from the Prostate, Lung, Colorectal and Ovarian Cancer (PLCO) trial. Our approach is able to transform genotype sequencing information into a simple tabular form that can then be used as the input to any ML model. Coupled with the recorded treatment regimen and clinical parameters of matched patients from the genetic dataset, we estimate the treatment effect in terms of mortality prediction and use it to prescribe the optimal treatment for any given patient. By including the genetic features engineered through our proposed method, our models have a higher accuracy than the models without genetic information embedded. The increase in predictive accuracy demonstrates the improved efficacy of our method in the predictive setting. Furthermore, in the prescriptive setting, the models including genetic features output different treatment choices for patients, showing the impact of their inclusion. This is further highlighted by the feature importance of the genetic features such as mutations in the FAT3, BRCA1, BRCA2, and NF1 genes, where they rank highly with a tighter aggregation of the top features, relative to the sharp drop-off in feature importance after the top feature in the models without genetic data. Taken together, in summary, our models will allow oncologists to make more informed and accurate decisions, incorporating a patients genetic data with all other available clinical information, which has the potential for improved prognosis and better long-term survival outcomes for Ovarian Cancer patients.