Results:
A total of 1143 ICH patients were obtained from the MIMIC-III database, including 760 survivors and 383 died. Table 1 shows the clinical characteristics of patients who survived and died during hospitalization. Table A.2 in Appendices shows the changes of physiological characteristics and laboratory parameters of dead and alive patients over time.
First, we used all 122 variables to construct five models, and used the learning curve and grid search to determine the optimal parameters. Prediction performance comparison results after 5-fold cross validation are shown in Table 2.
We can find that the GBDT model have the best accuracy (0.87) and the best F1 score (0.80). In terms of accuracy and AUROC, GBDT model had better values than other models, 0.87 and 0.93, respectively. Naïve Bayes had the best recall rate (0.85), but its accuracy and precision were the lowest. KNN model had the lowest recall rate (0.60), F1 score (0.70) and AUROC (0.87). Then we used the feature importance of RF and LASSO regression to select the most important feature variables. The first 39 most important variables were selected by two methods respectively (Table A.3 in Appendices), and the intersection of the two methods was taken as the screened variables, a total of 18. The importance order of the intersection variables is shown in Fig. 1. The importance score is normalized value, distributed between 0 and 1, and the closer to 1, the more important the variable is.
We reconstructed and trained five models with 18 variables obtained, and observed the changes of each indicator as shown in Table 3. The ROC curves of these predictive models are presented in Fig. 2.
Compared with the model constructed with all variables, it was found that although GBDT model has a small decline in precision, recall and F1 value, it can be seen from the AUROC index that GBDT model was the best among the five models. From the results, we found that the prediction effect of all the five models had not decreased significantly. Therefore, the input variables of our models were reduced from 122 to 18 successfully, greatly improving the practicability.