Data driven insights from multiple OCM practices to optimize care for managing high-risk cancer patients across multiple cancer types.

Abstract

6576 Background: Medicare’s Oncology Care Model (OCM) aims to improve the quality and coordination of care for cancer patients. Previously, Integra Connect (IC) developed machine learning (ML) based risk prediction models called Integra Predict (IP) for In-Patient Admissions, Emergency Department Visits and End-of-Life in active chemotherapy patients within the next six months. The patients were risk-stratified with the purpose of optimizing clinical resources utilization. Our objective is to demonstrate the insights derived from an updated IP version applied to OCM patients from three IC network practices: P1 (a multi-state network of community oncology practices), P2 (a single site community oncology practice) and P3 (a regional network of oncology practices connected to academic medical institution). Methods: For training the ML models, 90,417 OCM reconciliation episodes from PP4–PP8 (Jan. 2018-June 2020) from 10 distinct IC practices were used. The models comprise 37 features: 12 adverse events (within last 6 month of episode start date), 7 comorbidities (in the past irrespective of lookback period), 7 latest vital sign measurements recorded within 30 days of date of prediction, 11 demographic and other clinical attributes. Based on LogLoss, 54 ML experiments were performed leading to three best algorithms to predict the risk for In-Patient Admissions, Emergency Department Visits and End of Life. To generalize our ML approach, we leveraged an additional holdout dataset in periods PP9 and PP10 (July 2020 - June 2021) pertaining to practices P1, P2 and P3 with 6378, 973 and 5921 episodes, respectively. Matthew Correlation Coefficient (MCC) was chosen as the metric to test the generalizability of the models. For each model, the variance of MCC was then computed. To uncover secondary factors influencing clinical care, top-3 frequently reported adverse events and top-5 comorbidities were determined. Results: For each ML model, the variance of MCC applied to the holdout dataset was low (order of 10-4). Types of cancers reported differed within P1, P2 and P3. However, across these practices, irrespective of the cancer type, the most frequently reported adverse events (pain, anemia, and cachexia) and comorbidities (cardiovascular disease, chronic pulmonary disease, dyspnea at rest, dorsopathies and renal disease or failure) observed were the same. Conclusions: Low variance in MCC implies ML models are applicable to different types of community oncology practices. A care coordination program for high-risk cancer patients involving the monitoring of pain, cachexia and anemia can lead to optimal utilization of clinical resources within these practices. Program expansion involving optimal management of comorbidities such as cardiovascular disease, chronic pulmonary disease, diabetes, dyspnea at rest and renal disease or failure may yield additional benefits.