Using artificial intelligence-powered evidence-based molecular decision-making for improved outcomes in ovarian cancer.

person Brian Christopher Orr Medical University of South Carolina, Charleston, SC info_outline Brian Christopher Orr, Arlette Harper Uihlein, Robert Montgomery, Mackenzy Radolec, Robert P. Edwards

Authors person Brian Christopher Orr Medical University of South Carolina, Charleston, SC info_outline Brian Christopher Orr, Arlette Harper Uihlein, Robert Montgomery, Mackenzy Radolec, Robert P. Edwards Organizations Medical University of South Carolina, Charleston, SC, Predictive Oncology, Pittsburgh, PA, University of Pittsburgh Medical Center (UPMC), Pittsburgh, PA, UPMC Magee-Womens Hospital, Pittsburgh, PA Abstract Disclosures Research Funding No funding sources reported Background: High grade serous ovarian cancer (HGSC) is a complex deadly disease. Due to the lack of symptoms at early stages, vast majority patients present at stage III/IV. Frontline chemotherapy and surgery is effective with nearly all patients experiencing a remission, though over 80% patients will recur in the first 1-2 years and ultimately die of the disease in the next few years. Approximately 20% will be long-term survivors, but no current clinical data can predict who may be a short- or long-term survivor. Clinical management, monitoring, and systemic treatment decisions could be optimally individualized to patients if able to better define the prognosis. The goal of this study is to employ artificial intelligence multi-omic machine learning models to predict survival outcomes. Methods: Clinical data and tumor specimens from University of Pittsburgh Center from 2010-2016 were analyzed. Patient data, whole exome sequencing (WES), whole transcriptome sequencing (WTS), drug response profile, and digital pathology profile were used as input feature sets for training the multi-omic machine learning (ML) models. Hypothesis-free training of the ML models was utilized to classify patient survival at 2yr and 5yr threshold. Model performance was estimated using AUROC (area under the receiver operating characteristic curve) metric, scores >0.5 having higher prediction potential. Validation of the prediction results of each ML model with top performing models meeting statistical significance threshold of α=0.01. Results: 160 HGSC model builds were completed. ML models achieved high prediction accuracy for both short-term (2y) and long-term (5y) patient cohorts, identifying 7 models with ≥ 0.7 AUROC model performance for 2yr OS threshold and 13 models with ≥ 0.7 AUROC model performance for 5yr OS threshold. Addition of multi-omic feature sets to clinical profile improved the model’s ability to predict OS for both short-term and long-term thresholds. Multi-omic feature set inputs led to superior prediction and improved performance over clinical profile information alone, and top performing multi-omic models predicted better than any feature set in isolation. Comparison of top features identified by 2y and 5y OS threshold models, we found that molecular features (WTS followed by WES feature sets) predominantly drove 2y cohort while digital pathology imaging features were the driver the 5y cohort top performing models. Conclusions: Current clinical data does not accurately predict prognosis. Utilizing multi-omic machine learning models, superior prediction of short- and long-term survivors was achieved. The specific drivers of the top performing models were different for the short- and long-term cohorts, identifying future research opportunities as well as development potential of a clinical decision tool.