Predictive value of machine learning for in-hospital mortality for trauma-induced acute respiratory distress syndrome patients: an analysis using the data from MIMICⅢ
10.3760/cma.j.cn121430-20211117-01741
- VernacularTitle:机器学习对创伤合并急性呼吸窘迫综合征患者院内死亡的预测价值
- Author:
Rui TANG
1
;
Wen TANG
;
Daoxin WANG
Author Information
1. 重庆医科大学附属第二医院呼吸与危重症医学科，重庆 400010
- Keywords:
Machine learning;
Acute respiratory distress syndrome;
Trauma;
Artificial intelligence
- From:
Chinese Critical Care Medicine
2022;34(3):260-264
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the value of machine learning methods for predicting in-hospital mortality in trauma patients with acute respiratory distress syndrome (ARDS).Methods:A retrospective non-intervention case-control study was performed. Trauma patients with ARDS met the Berlin definition were extracted from the the Medical Information Mart for Intensive CareⅢ (MIMICⅢ) database. The basic information [including gender, age, body mass index (BMI), pH, oxygenation index, laboratory indexes, length of stay in the intensive care unit (ICU), the proportion of mechanical ventilation (MV) or continuous renal replacement therapy (CRRT), acute physiology scoreⅢ(APSⅢ), sequential organ failure score (SOFA) and simplified acute physiology scoreⅡ(SAPSⅡ)], complications (including hypertension, diabetes, infection, acute hemorrhagic anemia, sepsis, shock, acidosis and pneumonia) and prognosis data of patients were collected. Multivariate Logistic regression analysis was used to screen meaningful variables ( P < 0.05). Logistic regression model, XGBoost model and artificial neural network model were constructed, and the receiver operator characteristic curve (ROC) was performed to evaluate the predictive value of the three models for in-hospital mortality in trauma patients with ARDS. Results:A total of 760 trauma patients with ARDS were enrolled, including 346 mild cases, 301 moderate cases and 113 severe cases; 618 cases survived and 142 cases died in hospital; 736 cases received MV and 65 cases received CRRT. Multivariate Logistic regression analysis screened out significant variables, including age [odds ratio ( OR) = 1.035, 95% confidence interval (95% CI) was 1.020-1.050, P < 0.001], BMI ( OR = 0.949, 95% CI was 0.917-0.983, P = 0.003), blood urea nitrogen (BUN; OR = 1.019, 95% CI was 1.004-1.033, P = 0.010), lactic acid (Lac; OR = 1.213, 95% CI was 1.124-1.309, P < 0.001), red cell volume distribution width (RDW; OR = 1.249, 95% CI was 1.102-1.416, P < 0.001), hematocrit (HCT, OR = 1.057, 95% CI was 1.019-1.097, P = 0.003), hypertension ( OR = 0.614, 95% CI was 0.389-0.968, P = 0.036), infection ( OR = 0.463, 95% CI was 0.289-0.741, P = 0.001), acute renal failure ( OR = 2.021, 95% CI was 1.267-3.224, P = 0.003) and sepsis ( OR = 2.105, 95% CI was 1.265-3.502, P = 0.004). All the above variables were used to construct the model. Logistic regression model, XGBoost model and artificial neural network model predicted in-hospital mortality with area under the curve (AUC) of 0.737 (95% CI was 0.659-0.815), 0.745 (95% CI was 0.672-0.819) and 0.757 (95% CI was 0.680-0.884), respectively. There was no significant difference between any two models (all P > 0.05). Conclusion:Logistic regression model, XGBoost model and artificial neural network model including age, BMI, BUN, Lac, RDW, HCT, hypertension, infection, acute renal failure and sepsis have good predictive value for in-hospital mortality of trauma patients with ARDS.
