2.2. Patient selection
Patients admitted to ICU who were diagnosed with ICH (ICD-9 = 431) were eligible for inclusion. In the MIMIC-III database, each patient has its ID (SUBJECT_ID), and each admission is reassigned with an ID (HADM_ID). We screened out the variables of each ICH patient at the time of the last admission in the database record period (that is, each SUBJECT_ID corresponds to a HADM_ID). After extracting, the data of 1143 ICH patients older than 18 years were included and divided into the dead (383 people) and the alive group (760 people).
2.3. Predictors and outcomevariables
For each ICH patient, we collected variables including, patient physiological characteristics, laboratory parameters, and CT findings. For variables with temporal information, the maximum and minimum values of daily variables within three days after admission were extracted, and we took them as the parallel input of the model, because we believed that these variables with time information could more directly reflect the limiting condition of the patient’s body and had a more obvious influence on whether the patient died or not. After consulting with experts in this area, we selected variables: Gender, marital status, ethnicity, age, complication (high blood pressure, diabetes, ischemic heart disease, heart failure, pneumonia), and ICH location (basal ganglia, lobe, infratentorial). In addition, variables with temporal information (the maximum and minimum values of the first, second and third days of admission) included: systolic pressure, diastolic pressure, mean blood pressure, heart rate, body temperature, GCS score, white blood cells, lymphocytes, neutrophils, eosinophils, basophils, red blood cells, hemoglobin, platelet count, hematocrit, glucose, potassium, sodium, creatinine, and urea nitrogen. The outcome of this study was in-hospital mortality of ICH patients. If the variable’s missing values were more than 20% of the total, we removed it.