(Sunday, 7/14/2019) 4:00 PM - 4:30 PM
Room: Exhibit Hall | Forum 5
Purpose: Decision support for cancer patients at end-of-life is needed to optimize patient-centered outcomes, including quality of life and choosing interventions that align with goals of care. Currently there are a limited number of mortality prediction tools such as the TEACHH, Chow, and NEAT models. Here we build a model incorporating multiple factors to predict 30-day mortality in patients receiving palliative radiotherapy for advanced cancer.
Methods: A cohort (n=518) treated with external beam radiotherapy to a metastatic site between 2012-2016 was included. Factors associated with 30-day mortality, including demographics and clinical and laboratory data, were retrospectively collected. Generalized linear models were built with no regularization (logistic regression (LR)), and with regularization (Lasso and Elastic-Net). Random forest (RF) and gradient boosting machine (GBM) models were also built to assess non-linearity in the response. Missing data was imputed using the mean or mode, and a variable was added to indicate which data was imputed. Models were built in R using cross validation withholding a 25% test set, and stratified subsampling due to unbalanced outcome. A seed was used for pairwise t-test comparison of the test balanced accuracy distributions. The area under the ROC curve (AUC) was also reported.
Results: The LR model has a mean-balanced-accuracy (MBA) of 0.72 (AUC=0.79) compared to Lasso and Elastic-Net (MBA=0.71, p=0.11, AUC=0.79; MBA=0.71, p=0.77, AUC=0.80). Non-linear models proved worse with RF and GBM (MBA=0.70, p=0.02, AUC=0.78; MBA=0.70, p=0.01, AUC=0.79). TEACHH, Chow, and NEAT models alone had inferior results (MBA=0.67, p<0.01, AUC=0.68, MBA=0.61, p<0.01, AUC=0.63, MBA=0.62, p<0.01, AUC=0.65) to the new developed model.
Conclusion: Individual scoring methods (TEACHH, Chow, NEAT) have lower predictive power when used alone. By incorporating multiple patient indicators together with these scores we have improved AUC by 17%. Our model will be publicly released together with a nomogram for different institutions to validate performance.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
