Prediction of 30-day unplanned hospital readmission through survival analysis

Figures

• 

  Figure 1. Rules for removing episodes based on exitus date. The boxes represent hospitalization periods (A) exitus date and date of discharge are the same, (B) exitus happens in less than 90 days after discharge, and (C), same as B, but there is a following episode, making Episode C suitable for the study.

• 

  Figure 2. Absolute Pearson Correlation Coefficient (blue) and matching p-value (orange) between the target feature (days from discharge to readmission) and each continuous predictive feature. Although lower p-values exist, the scale is limited to 10−4. Only non-censored data was used in this test.

• 

  Figure 3. F-Statistic from a One-Way ANOVA test (blue) and matching p-value (orange) between the target feature (days from discharge to readmission) and each pharmacy feature (categorised as 0 or ≥1 treatments). Although lower p-values exist, the scale is limited to 10−4. Only non-censored data was used in this test.

• 

  Figure 4. F-Statistic from a One-Way ANOVA test (blue) and matching p-value (orange) between the target feature (days from discharge to readmission) and each binary predictive feature. Although lower p-values exist, the scale is limited to 10−4. Only non-censored data was used in this test.

• 

  Figure 5. Performance evaluation and comparison of the trained survival models by 10-fold cross validation.

• 

  Figure 6. Comparison of models: ROC-AUC when predicting readmission risks at 30 days since discharge. Results in boxplots are computed with a 10-fold cross-validation.

• 

  Figure 7. 30-day ROC-AUC comparison depending on the censoring criteria. Left uses raw observed time periods, while right censors 90-day or later readmissions. Results obtained with XGBcox, 10-fold CV.

• 

  Figure 8. Effect of filtering episodes followed by exitus. Comparison of ROC-AUC at 30 days when censoring time-to-event to exitus date versus filtering out exitus happening in less than 90 days since last discharge, 10-fold CV with XGBcox.

• 

  Figure 9. Performance at 30 days since discharge when including (left) or excluding (right) post-discharge information. XGBcox model, 10-fold CV.

• 

  Figure 10. Top 20 most important features obtained with permutation importance of XGBcox. For each feature, a boxplot shows de ROC-AUC score variation when evaluating the model with shuffled values of said feature across many iterations.

• 

  Figure 11. Partial dependence (PD) plot of the most important features according to permutation importance. For each partial dependence line (the average of the individual conditional expectation or ICE), the 95% confidence interval of the ICE is displayed.

• 

  Figure 12. Survival function when splitting test data by median value of main risk factors into low and high risk groups. Each panel compares the mean survival function obtained by a non-parametric Kaplan-Meier estimator with our model's predicted survival function, for all individuals in test set. Our model prediction combines the baseline survival function of a Cox Proportional Hazards model obtained with Breslow's method and the risks predicted by an XGBoost regression model. (dashed line) Kaplan-Meier estimation, (solid line) own model prediction, (red) population with a higher expected risk of readmission or lower readmission-free time, (blue) low expected readmission risk population.