Analysis on the machine learning model of the prognosis of acute pancreatitis based on complete blood count parameters

CHEN Tongle; HAN Baosong; WU Jingyi

doi:10.13431/j.cnki.immunol.j.20250099

Abstract：Objective To investigate the predictive efficacy of complete blood count(CBC)parameters for the prognosis of acute pancreatitis(AP),and to construct a machine learning model.Methods The clinical data of 120 patients with AP admitted from January 2021 to December 2024 were retrospectively analyzed.Based on the prognostic outcomes within 28 d of treatment,they were divided into the poor prognosis group and the good prognosis group.On the first day of admission,CBC parameters[red blood cell count,hemoglobin,hematocrit,white blood cell count,neutrophil count,lymphocyte count,monocyte count,red blood cell distribution width(RDW),platelet count,mean platelet volume,platelet distribution width,neutrophil/lymphocyte ratio(NLR),platelet/lymphocyte ratio(PLR),lymphocyte/monocyte ratio(LMR)were detected in the two groups.The general data and CBC parameters of the two groups were compared.The clinical variables were screened using the Lasso regression equation.Four models,namely Support Vector Machine(SVM),logistic regression(LR),deep neural network(DNN),and Random Forest(RF),were constructed for external validation.The predictive efficacy of the four models for the prognosis of AP was evaluated by using the receiver operating characteristic(ROC)curve,the precision-recall(PR)curve,the calibration curve and the clinical decision curve.Results The APACHE Ⅱ score,bedside severity index score,proportion of vasoactive drug use,and proportion of severe hypoproteinemia in the poor prognosis group were all higher than those in the good prognosis group(P<0.01).The levels of RDW,NLR and PLR in the poor prognosis group were all higher than those in the good prognosis group,while the level of LMR was lower than that in the good prognosis group(P<0.01).The Lasso regression equation ultimately screened out four non-zero coefficient variables:NLR,RDW,APACHE Ⅱ score,and severe hypoproteinemia.Based on the above variables,SVM,LR,RF and DNN machine learning models were constructed.The RF model had the highest area under the ROC curve(AUC),PR AUC,accuracy rate and F1 score for predicting the prognosis of AP,and had the optimal comprehensive performance.The calibration curve showed that the curve of the RF model for predicting the prognosis of AP was relatively close to the ideal curve.The decision curve showed that when the threshold probability value of the RF model was 15%to 100%,it had a significance clinical net benefit rate.External validation also showed that the RF model had the optimal predictive efficacy.The calibration curve indicated that the predictive curve of the RF model for the prognosis of AP highly coincided with the actually observed curve.The decision curve showed that the RF model provided clinical net benefits to patients across the decision threshold range of 0 to 100%.Conclusion Machine learning models constructed based on CBC parameters can make relatively accurate predictions about the prognosis of AP.Among them,the RF model has the optimal predictive efficiency,which can be used as an auxiliary tool for clinical prediction of AP prognosis,and can also provide reference for clinical treatment.