Predicting 30-day mortality with routine blood tests in patients undergoing palliative radiation therapy: A comparison of logistic regression and gradient boosting models.

Abstract

This study aimed to estimate the 30-day mortality (30D_M) and compare models for 30D_M prediction in patients undergoing palliative radiation therapy (RT). Data from 3,756 patients who underwent palliative RT between 2018 and 2020 at two institutions were retrospectively reviewed. From one institution, 3,315 patients were randomly assigned to the training (N = 2,652) and internal validation (N = 663) cohorts. The remaining 441 patients from the other institution constituted the external validation cohort. Nineteen features, including seven blood test features, were extracted from medical records. For 30D_M prediction, 4 models were constructed: logistic regression comprising all features (LRM-A) and 7 blood test features (LRM-B) and gradient boosting using all features (GBM-A) and 7 blood test features (GBM-B). The 30D_M rates were 10.6 %, 11.2 %, and 17.5 % in the training, internal validation, and external validation cohorts, respectively. GBM-B demonstrated a good value for the area under the receiver operating characteristic curve (AUC) (0.830-0.863). Among the four models, GBM-A exhibited the highest AUC values, although GBM-B still generally outperformed LRM-A and LRM-B. The 30D_M rates significantly differed across the four prognostic groups according to the quantile values of predictive probability of GBM-B: 0-0.8 % (1st quantile), 1.2-3.4 % (2nd quantile), 8.7-12.9 % (3rd quantile), and 31.1-36.6 % (4th quantile), respectively. The 30D_M rates were successfully stratified into distinct prognostic groups by using the GBM-B model. The model could serve as a straightforward and objective tool for predicting mortality in patients undergoing palliative RT.