Leveraging a Real-Time Machine Learning Platform to Personalize Quality Assurance at a Large Multi-Site Oncology Practice

Abstract

Peer review is a critical component of safe and effective oncology care and provides care teams with an opportunity to improve patient outcomes with predictive analytics. We describe a paradigm for incorporating predictive analytics into daily prospective peer review at a large multisite oncology center. Previously, we deployed a department-wide proprietary oncology analytics platform connected to the Electronic Medical Record System (outpatient and inpatient), Oncology Information System, Tumor Registry, Treatment Planning System and Radiology PACS. Variables extracted and calculated by the analytics platform on a nightly basis include patient demographics, past medical history, detailed pathology data, recent laboratory data, vitals, prior systemic cancer therapies, prior surgical treatments, prior radiation treatments, recent hospitalization, vital status, and current treatment details. Using this platform, we trained machine learning models for critical patient events, such as unplanned hospitalization, mortality, occurrence of grade 3 or higher toxicities, and treatment interruption. Importantly, these models leverage so-called interpretable machine learning techniques, allowing us to interrogate the models and obtain explanations for why a patient is predicted to be at higher or lower risk of an event. All such models were previously evaluated for performance on standalone testing sets. In this abstract, we propose a paradigm to include predictive analytics in our daily prospective peer review process by creating a list of patients at high risk of critical events (hospitalization, mortality, treatment interruption, and serious adverse events) and reviewing those lists in conjunction with our daily peer review conference. We also propose a process to review any gaps in clinical care (e.g., abnormal labs, decreases in patient weight, or missed progress visits) to recommend suitable interventions to the care team to reduce patient risk. A total of 64,286 radiation treatment cases were included in our analytics platform and used to train and validate our predictive models. Moving forward, we aim to peer review 20 patient charts per day using this paradigm. While numerous studies in the field of radiation oncology have focused on the benefits of machine learning when applied to contouring and personalized radiation dose, few have focused on how machine learning can improve supportive care. In this abstract, we propose a framework to introduce machine learning into quality assurance, with the goal of improving patient outcomes.