[Prediction and feature selection for fatal gastrointestinal bleeding recurrence in hospital via machine learning]

To propose a method of prediction for fatal gastrointestinal bleeding recurrence in hospital and a method of feature selection via machine learning models.

728 digestive tract hemorrhage samples were extracted from the first aid database of PLA General Hospital, and 343 patients among them were diagnosed as fatal gastrointestinal bleeding recurrence in hospital. A total of 64 physiological or laboratory indicators were extracted and screened. Based on the ten-fold cross-validation, Logistic regression, AdaBoost and XGBoost were used for classification prediction and comparison. XGBoost was used to search sequence features, and the key indicators for predicting fatal gastrointestinal bleeding recurrence in hospital were screened out according to the importance of the indicators during training.

Logistic regression, AdaBoost and XGBoost all get better F1.5 score under each feature input dimension, among which XGBoost had the best effect and the highest score, which was able to identify as many patients as possible who might have fatal gastrointestinal bleeding recurrence in hospital. Through XGBoost iteration results, the Top 30 indicators with high importance for predicting fatal gastrointestinal bleeding recurrence in hospital were ranked. The F1.5 scores of the first 12 key indicators peaked at iteration (0.893), including hemoglobin (Hb), calcium (CA), red blood cell count (RBC), mean platelet volume (MPV), mean erythrocyte hemoglobin concentration (MCH), systolic blood pressure (SBP), platelet count (PLT), magnesium (MG), lymphocyte (LYM), glucose (GLU, blood gas analysis), glucose (GLU, blood biochemistry) and diastolic blood pressure (DBP).

Logistic regression, AdaBoost and XGBoost could achieve the purpose of early warning for predicting fatal gastrointestinal bleeding recurrence in hospital, and XGBoost is the most suitable. The 12 most important indicators were screened out by sequential forward selection.