4. Discussion
This study established a prediction model based on RF, GBDT, decision tree, Naïve Bayes, and KNN algorithms to predict the in-patient mortality of ICH patients, and compared the effects of the five models. The results showed that the AUROC and precision of the RF model were better than other models. RF feature importance and LASSO method were used to screen out fewer variables, so that the model was simplified and the performance is not affected. This had important implications for practical applications of the model, because researchers can predict mortality rates from just 18 data sets of patients without using more than 100 variables.
This study adds temporal information into the prediction of mortality model of ICH patients. We took the maximum and minimum values of each indicator in the first, second and third days after admission of ICH patients as the model input variables, which can make the model learn more information automatically and make the prediction effect better. Compared with previous studies[23-26], the AUROC obtained by the GBDT model in this study reached 0.93, which was higher than the previous results, although the data sources and methods used were different.
From the results, the indicators of the first, second and third days all had an important impact on the death of ICH patients. This has not been considered in previous studies. The most important influencing factor is GCS score, which is consistent with literature reports[27]. The lower the GCS score, the higher the mortality rate (Fig. 3). Another important factor affecting ICH mortality was hematoma volume [27], but this index was not included in this study, because only 405 patients (35.43% of the total number) had a record of this index, which was seriously missing. It has been reported that the larger the size of hematoma, the higher the mortality rate of patients with ICH.
This study is also the first time to use the MIMIC-III database as the data source to establish ICH patient’s mortality prediction model. The dataset spanned more than a decade and details of patient care. The integrity and normality of the data were assured, and it was the basis for our models to incorporate temporal information. Therefore, there was no problem in our data that the population age was relatively small [13] or the number of participants was small [23,26,28]. But the MIMIC-III data was based on a single medical center, which meant the generality of our results deserves consideration.
Machine learning has great potential in the field of health and even in the field of critical care [29]. Specifically, the ICH mortality prediction model constructed by using machine learning methods (such as RF, GBDT, etc.) is far better than the traditional scoring system ICH score [30]. Since ICH score involved hematoma size, ICH score was not compared with other machine learning models in this paper. However, from the existing research results, there is no doubt that the machine learning model is better than the ICH score.