关键词: 极低出生体重儿, 机器学习, 红细胞输血, 贫血, 危险因素

Abstract: Objective To investigate the influence factors of blood transfusion and transfusion amount and the construction of prediction model in very low birth weight infants (VLBWIs). Methods A total of 102 VLBWIs admitted to our hospital from February 2017 to January 2021 were selected and divided into the training set (76 cases) and the test set (26 cases) with a ratio of 3∶1 by using computer-generated random number method. The medical data, diseases and treatments were compared between the transfusion and non-transfusion groups. The risk factors for transfusion in VLBWIs were analyzed by Logistic regression. Four machine learning methods, including Logistic regression, CatBoost, XGBoost and LightGBM, were used to construct transfusion prediction models to compare the prediction effects. Multiple linear stepwise regression was used to analyze the independent factors affecting the blood transfusion volume to fit the prediction model in VLBWIs. Results Low birth weight, small for gestational age infant, large volume blood collection within two weeks after birth, long-term parenteral nutrition and cesarean section were independent risk factors for blood transfusion (P<0.05). The AUC of Logistic regression, XGBoost, CatBoost and LightGBM models were 0.836 (95%CI: 0.745~0.889), 0.801 (95%CI: 0.734~0.862), 0.738 (95%CI: 0.658~0.800), 0.700 (95%CI: 0.609~0.785), respectively. There were no statistically significant differences compared with the test set (P>0.05). Stepwise multiple linear regression analysis was used to determine that birth weight, gestational age, Hb value at birth and Hct at birth were the independent influencing factors for transfusion. The prediction model for VLBWIs blood transfusion volume was: Y= 24.175-0.731×birth weight-0.538×gestational age- 0.431 × Hb value at birth- 0.569 × Hct, F=33.321, P<0.001, D-W (Durbin-Watson) = 1.725, R²=0.671. Conclusion Birth weight, gestational age, blood collection volume within two weeks after birth, long-term parenteral nutrition and cesarean section were independent risk factors for blood transfusion in VLBWIs. Birth weight, gestational age, Hb value at birth and Hct at birth were independent factors affecting transfusion. The prediction effect of Logistic curve is more accurate by using four machine learning methods of Logistic regression, CatBoost, XGBoost and LightGBM.

Key words: Very low birth weight infants, Machine learning, Red blood cell transfusion, Anemia, Risk factor