PCN428 SCALABLE PREDICTION OF CANCER PATIENT CLINICAL EVENTS IN REAL WORLD CLINICAL SETTINGS

Abstract

Patient safety concerns, both related and unrelated to treatment, are a critical concern to managing patient care in oncology, abd can limit treatment adherence and achieving optimal clinical outcomes 1,2. We outline here a scalable approach leveraging common data standards for predicting patient clinical events across multiple real-world electronic medical records systems. Patients initiating treatment for Ovarian cancer with gemcitabine, platinum-based agents, taxanes or PARP inhibitors were identified in 2 nationally representative electronic medical record databases from Jan 2010- Dec 2018. ICD codes were used to identify over 20 comorbidities across multiple condition classes. Baseline descriptive characteristics and subsequent dosing changes were assessed. An ensemble-model, and stacked model machine learning approach (Logistic Regression, Gaussian Naïve Bayes, Multi-layer Perceptron, K-Neighbors, Decision Tree, Random Forest, Gradient Boosted Tree, XG Boosted Tree Classifiers) was used to develop a predictive algorithm for each underlying comorbidity, with a 3:1 training/testing split, and cross-validation scoring with a K-fold of 6. Model performance was assessed by standard metrics such as ROC curve, AUC scores, PPV, and NPV estimates. Feature ranking/impact was validated using multiple methods of interpretability and the results analyzed. The stacked ensemble model was then compared against a commercially available automated data science platform that tested over 60 additional classification methodologies. Among treated patients, comorbidity and adverse event rates were consistent with previous reports1, 4,5,6. Across both datasets, the stacked ensemble model showed up to a 20% improvement in AUC compared against a previously presented algorithm2. The automated data science platform identified a blended model of Elastic-Net (ENET), Random Forest, Gradient Boosted, Extra Trees, and XG Boosted Tree classifiers that further increased AUC score by 15%. Results indicate opportunities to improve the management of chemotherapy-associated thrombocytopenia through both active management of dose-adjustment and systematic application of point-of-care predictive algorithms.