Background: End-of-life management is a well-known challenging aspect of cancer care. In particular, timely hospice enrollment is a leading quality metric in the Oncology Care Model that has substantial room for improvement. An automated algorithmic tool that can incorporate the wealth of available EHR data and rapidly identify patients with a high risk of imminent mortality could be a valuable asset to supplement important clinical decisions and improve timely hospice care. Methods: A retrospective study cohort was formed using patients with metastatic cancer from US Oncology Network (USON) practices participating in the Oncology Care Model (OCM) between January 1, 2017 and June 30, 2019. Patients were required to have at least one record for lab values and vital signs in the EHR database. Patients were excluded from the study cohort if they were not enrolled in the OCM program or did not have a diagnosis for metastatic cancer. The patients satisfying the selection criterion were used to train and optimize the model. The training dataset was also used for internal validation and hyperparameter tuning until the final model was produced. As external validation, the final model was independently tested on 3 separate holdout datasets including OCM patients between July 1, 2019 and March 31, 2020. To avoid bias, all holdout datasets used for validation were excluded from the model. Results: A multivariable model to predict 90-day mortality was developed using a retrospective dataset derived from EHR data and Medicare claims data. A logistic regression algorithm using L1 (lasso) regularization yielded the best performance compared to other model candidates. The performance on the training cohort was given by a cross-validated AUC score of 0.85 (95% CI, 0.84 to 0.86). Further, external validation conducted using 3 independent holdout datasets demonstrated impressive generalizability marked by stable performance scores across multiple time periods (AUC between 0.84 and 0.85). Conclusions: This study builds upon previous work and further establishes the utility of machine learning to predict risk of imminent mortality for advanced cancer patients using available EHR data. A data-driven tool that estimates the probability of 90-day mortality could be leveraged as a powerful supplementary aid to clinicians managing end-of-life care at oncology practices.