Multi-institutional validation of a multi-modal machine learning algorithm to predict and reduce acute care during cancer therapy

点击翻译按钮获取中文摘要

PROJECT ABSTRACT An estimated 650,000 patients with cancer receive systemic therapy or radiation therapy (RT) annually in the United States. Many of these patients undergoing outpatient cancer therapy will require acute care with an emergency department visit or hospital admission due to symptoms from treatment, disease, or comorbidities. This can impact cancer outcomes, patient treatment decisions, and costs to patients and the healthcare system. While there has been much enthusiasm for artificial intelligence and machine learning (ML) to improve healthcare delivery, high quality prospective data are lacking, especially across diverse clinical practice settings. We previously completed one of the first randomized controlled studies in healthcare ML, demonstrating that ML based on EHR data can accurately generate personalized predictions and guide supportive interventions to decrease acute care requirements and costs in patients undergoing RT and chemoradiotherapy (CRT) (NCT04277650). We have also developed a ML model for predicting hospitalizations based on prospective clinical trials of daily step counts collected in patients undergoing CRT. The research objective of this application is to leverage a geographically, racially, socioeconomically, and technically diverse network of healthcare settings and patients to assess and maximize how accurately and equitably these approaches generalize. Our team includes the University of California, San Francisco (UCSF), Duke University, Beth Israel Deaconess Medical Center, Essentia Health in Duluth, MN and Ashland, WI, Washington Hospital in Fremont, CA, Duke Regional Hospital in Durham, NC, and Duke Raleigh Hospital in Raleigh, NC. Specifically, we seek to: (1) prospectively evaluate the validity of an EHR-based acute care prediction ML algorithm across our network and establish a framework for equity, generalizability, and portability and (2) validate our existing patient-generated health data (PGHD; step count) models that predict hospitalization during CRT at a second institution and integrate with our EHR-based ML algorithm to enhance prediction of acute care needs. We hypothesize that our approaches will be accurate across institutions though require adjustments for both generalizability and fairness, and that EHR- and PGHD-based approaches will offer complementary predictive performance. The long-term goal is to develop informatics-based tools that can be broadly and equitably deployed to improve the delivery of cancer care and subsequent treatment outcomes. This research will generate data regarding the generalizability and fairness of EHR- and PGHD-based approaches and a platform for a future multi-institutional randomized controlled trial.