A machine learning approach to predicting short-term mortality risk for patients starting chemotherapy.

Abstract

6538 Background: Patients who die soon after starting chemotherapy incur symptoms and financial costs without survival benefit. Prognostic uncertainty may contribute to increasing chemotherapy use near the end of life, but few prognostic aids exist to guide physicians and patients in the decision to initiate chemotherapy. Methods: We obtained all electronic health record (EHR) data from 2004-14 from a large national cancer center, linked to Social Security data to determine date of death. Using EHR data before treatment initiation, we created a machine learning (ML) model to predict 180-day mortality from the start of chemotherapy. We derived the model using data from 2004-11 and report predictive performance on data from 2012-14. Results: 26,946 patients initiated 51,774 discrete chemotherapy regimens over the study period; 49% received multiple lines of chemotherapy. The most common cancers were breast (23.6%), colorectal (17.6%), and lung (16.6%). 18.4% of patients died within 180 days after chemotherapy initiation. Model predictions were used to rank patients in the validation cohort by predicted risk. Patients in the highest decile of predicted risk had a 180-day mortality of 74.8%, vs. 0.2% in the lowest decile (area under the receiver-operating characteristic curve [AUC] 0.87). Predictions were accurate for patients with metastatic disease (AUC 0.85) and for individual primary cancers and chemotherapy regimens—including experimental regimens not present in the derivation sample. Model predictions were valid for 30- and 90-day mortality (AUC 0.94 and 0.89, respectively). ML predictions outperformed regimen-based mortality estimates from randomized trials (RT) (AUC 0.77 [ML] vs. 0.56 [RT]), and National Cancer Institute Surveillance, Epidemiology, and End Results Program (SEER) estimates (AUC 0.81 [ML] vs. 0.40 [SEER]). Conclusions: Using EHR data from a single cancer center, we derived a machine learning algorithm that accurately predicted short-term mortality after chemotherapy initiation. Further research is necessary to determine applications of this algorithm in clinical settings and whether this tool can improve shared decision making leading up to chemotherapy initiation.