Prospective Comparison of Medical Oncologists and a Machine Learning Model to Predict 3-Month Mortality in Patients With Metastatic Solid Tumors

• DOI: 10.1001/jamanetworkopen.2022.14514
• Publication Date: 2022-05-31T15:35:18Z
• Authors: Finly J. Zachariah; Lorenzo A. Rossi; Laura M. Roberts; Linda D. Bosserman
• Keywords: Radiation oncologist

To date, oncologist and model prognostic performance have been assessed independently mostly retrospectively; however, how compares with prospectively remains unknown.To compare a in predicting 3-month mortality for patients metastatic solid tumors an outpatient setting.This study evaluated prospective predictions cohort of seen oncology clinics at National Cancer Institute-designated cancer center associated satellites between December 6, 2019, August 2021. Oncologists (57 physicians 17 advanced practice clinicians) answered surprise question (3MSQ) within clinical pathways. A was trained electronic health record data from January 1, 2013, to April 24, identify high risk deployed silently October 2019. Analysis limited prognostications prediction the preceding 30 days.Three-month gradient-boosting binary classifier.The primary outcome comparison oncologists predict mortality. The metric positive predictive value (PPV) sensitivity achieved by medical their 3MSQ answers.A total 74 3099 3MSQs 2041 (median age, 62.6 [range, 18-96] years; 1271 women [62.3%]). In this 15% prevalence 30% both model, PPV 34.8% (95% CI, 30.1%-39.5%) 60.0% 53.6%-66.3%). Area under receiver operating characteristic curve 81.2% 79.1%-83.3%). significantly outperformed short-term mortality.In study, overall breast gastrointestinal cohorts cancer. These findings suggest that further studies may be useful examine could improve oncologists' confidence patient-centered goal-concordant care end life.