Statistical learning and prediction of diseases associated with the occurrence of death in the emergency department: An observational study on administrative data from Saint Joseph Hospital in Kinshasa-DRC

Abstract Objective The determinants, timing and predictive tools of mortality in emergency departments for specific diseases have already been evaluated in several studies. The diseases that are most likely to cause death in emergency departments have been studied in limited numbers. This study aimed to identify the diseases which are possibly predictive of the occurrence of death by using statistical learning methods. Methods Administrative data of patients admitted to the emergency departments of Saint Joseph Hospital in Kinshasa between January 1st 2018 and December 31st 2019 were reviewed retrospectively. The occurrence of death, defined as a patient who had a "deceased" outcome reported in the emergency registers, was predicted. Nineteen predictive conditions were modalities of the 'diagnosis' variable and were adjusted for age, gender and number of diagnoses. Six models were run on this data set, 70%. The AUC and its confidence interval were used to evaluate the performance of each model on the test data set, 30%. Results A total of five thousand one hundred and seventy-two (5172) subjects were incorporated into the analyses. The average age of the complete dataset was 47.9 ± 20.0 years. In the training and test datasets, the mean ages were 48.0 ± 20.2 and 47.4 ± 20.4 years, respectively. The training dataset comprised 50% male patients, while the test dataset included 48.8% male patients. Among these individuals, 7.5% in the training group and 7.4% in the test group did not survive. For each model, the area under the curve (AUC) values were as follows: ElasticNet 0.746 [0.689–0.784], logistic regression 0.735 [0.687–0.783], neural networks 0.735 [0.687–0.783], xGboost 0.734 [0.687–0.783], decision trees 0.50 [0.50–0.50], and random forest 0.428 [0.391–0.464]. The six primary predictors of mortality, listed in descending order, were shock states, sepsis, tuberculosis or pneumonia, coma, meningoencephalitis, and cancer. Conclusions Four of the six models used, performed well. ElasticNet was a network model that maximized the area under the ROC curve. Statistical learning methods applied to the twelve potential diseases in the prediction of the occurrence of death could help policy makers and health professionals to approach optimal decisions in the management of patients.