OBJECTIVES: ECOG is used extensively in oncology to assess progression and determine treatment and prognosis. As it is a clinical measurement, claims databases do not capture it. Previous studies relied on logistic regression models to predict ECOG with claims information. In this study, we assessed the utility of automated machine learning (AutoML) models in enhancing prediction accuracy.

METHODS: Patients with prostate, breast, colorectal, gastric, NSCLC or pancreatic cancer were identified from the claims-based clinical-genomic database GuardantINFORM from June 2014 to December 2022. Patients with a valid ECOG value extracted from their pathology reports (gold standard) and at least 2 medical claims in the 180 days prior to date of ECOG test result were included. ECOG value was dichotomized to 0-1 (good) vs 2+ (poor). Data was split into 60% training and 40% testing. The H2O R package was used to obtain the top 20 AutoML models ranked by performance, as measured by the area under the precision-recall curve (AUPRC). In a secondary analysis, we used principal component analysis (PCA) to improve the model efficiency.

RESULTS: Data from 8673 patients encompassing 85 claims-based variables were used. 1375 (16%) had poor ECOG score. The AutoML generated best model was a gradient boosting machine model (AUPRC=0.35; AUC=0.73; sensitivity = 62%; specificity= 93%). Across all the top models, age, number of hospitalizations, outpatient visits and Charlson comorbidity score were the major contributing variables. Using PCA (first 49 principal components, representing 75% of total variability), running efficiency was improved and Lasso regression model was the best model identified (AUPRC=0.34; AUC=0.72; sensitivity=53% and specificity=78%).

CONCLUSIONS: Using AutoML, we were able to develop highly specific models to predict the ECOG score for six different cancer types using claims data. This framework can be extended to other claims databases with some linked ECOG status from medical records.