Predicting 180-day mortality for women with ovarian cancer using machine learning and patient-reported outcome data.

Abstract

e13555 Background: Contra to national guidelines, women with ovarian cancer often receive aggressive treatment until the end-of-life. We trained machine learning algorithms to predict mortality within 180 days for women with ovarian cancer. Methods: Data were collected data from a single academic cancer institution in the United States. Women with recurrent ovarian cancer completed biopsychosocial patient-reported outcome measures (PROMs) every 90 days. We randomly partitioned our dataset into training and testing samples with a 2:1 ratio. We used synthetic minority oversampling to reduce class imbalance in the training dataset. We fitted training data to six machine learning algorithms and combined their classifications on the testing dataset into a voting ensemble. We assessed the accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) for each algorithm. Results: We recruited 245 patients who completed 1319 assessments. The final voting ensemble performed well across all performance metrics (Accuracy = .79, Sensitivity = .71, Specificity = .80, AUROC = .76). The algorithm correctly identified 25 of the 35 women in the testing dataset who died within 180 days of assessment Conclusions: Machine learning algorithms trained using PROM data offer state-of-the-art performance in predicting whether a woman with ovarian cancer will reach the end-of-life within 180 days. We highlight the importance of PROM data in ML models of mortality. Our model exhibits substantial improvements in prediction sensitivity compared to other similar models trained using electronic health record data alone. This model could inform clinical decision making and improve the uptake of appropriate end-of-life care. Further research is warranted to expand on these findings in a larger, more diverse sample.