Intelligent Diagnosis And Prognostic Prediction Of Hepatocellular Carcinoma Based On Real-world Clinical Data

Background and aims:Microvascular invasion(MVI)of hepatocellular carcinoma(HCC)is an important cause of early recurrence after radical hepatectomy,but its diagnosis can only be obtained from resected samples by microscopically.Presently,some studies have proposed that inflammatory indicators,such as neutrophil to lymphocyte ratio(NLR),lymphocyte to monocyte ratio(LMR),platelet to lymphocyte ratio(PLR)are related to MVI,but the results are contradictory.Tumor size is closely related to MVI,but its evaluation cutoff is controversial.Preoperative accurate prediction of MVI is of important clinical significance.Most of the current MVI prediction models come from imaging omics.The research methods are heterogeneous and difficult to be generalization.The existing prognostic staging system of HCC failed to accurately predict individual survival due to the introduction of heterogeneity and the lack of key prognostic factors.Therefore,based on the clinical data of liver HCC patients in the real world,using strictly statistical analysis,this study aims to(1)Clarify the relationship between inflammatory indicators and MVI;(2)Quantify the relationship between tumor size and MVI;(3)Build a robust and clinically applicable MVI prediction model;(4)Constructed accurately survival prediction model of HCC patients with common clinical parameters,which included comprehensive factors affecting the prognosis.Methods1.Collect clinical data: Hepatocellular carcinoma diagnosed pathologically in the First Affiliated Hospital of Nanchang University from October 2015 to October 2020 as the research subjects,collect demographic and medical history,laboratory examination,imaging examination and pathological diagnosis data,and follow up the survival status of patients.2.To analyze the relationship between inflammatory index NLR,LMR,PLR,tumor size and MVI: firstly,univariate analysis and curve fitting were used to visualize the relationship between each index and MVI;Then,the relationship between MVI and covariates was screened,confounding factors were adjusted,and curve fitting was drawn;Then the relationship between each index and MVI was tested by trend test and sensitivity analysis;Finally,subgroup analysis and interaction test were used to examine the relationship between inflammatory indexes,tumor size and MVI in subgroups.3.Multiple variable screening methods were used to construct MVI prediction model:combined with medical history,imaging,laboratory and pathological parameters,multiple variable screening methods(Best subset regression BSR,Bayesian information criterion BIC,stepwise regression,Mallows CP and Lasso regression)were used to find out the independent predictors of MVI,and logistic regression was carried out respectively.The predictive value of each model was evaluated by the area under the receiver operating characteristic(AUC),calibration curve,decision curve(DCA),clinical impact curve.4.Build MVI prediction model by machine learning XGboost: compare the models built by the above different variable screening methods,calculate the net reclassification index(NRI)and Integrated Discrimination Improvement(IDI)of the model,identify important and stable characteristic factors with independent prediction value,then build MVI prediction model by machine learning XGboost method,and compare the prediction performance with Logistic regression model.5.Develop web pages of MVI prediction model online: the MVI prediction model constructed by XGboost method was uploaded on a web page based on Internet browser.The individual MVI risk can be predicted by inputting the variable value of the model,and a two-dimensional code is created for application.6.Construct the prediction model of postoperative survival of hepatocellular carcinoma based on clinical indicators: select the characteristic variables by using the best subset regression and adjusted-R2,construct the prediction model of postoperative survival of hepatocellular carcinoma patients by using Cox regression,draw the prediction nomogram of 2,3,4 and 5-year survival rate,500 times with bootstrap resampling for internal test,and draw the time-dependent ROC curve,calibration curve and decision curve to evaluate the prediction performance of the model.Results1.Relationship between inflammatory indexes NLR,LMR,PLR,tumor size and MVI:univariate analysis and curve fitting showed that NLR,LMR,PLR and tumor size were all correlated with MVI,and NLR was correlated with MVI as a curve.But after adjusting for confounding factors,it was found that:(1)NLR was not independently correlated with MVI(OR: 0.99,95% CI: 0.84 ～ 1.16,P = 0.886);(2)LMR was not independently correlated with MVI(OR: 1.01(0.89～1.14)0.933);(3)PLR was not independently associated with MVI(OR: 1.06,95% CI: 0.85 ～ 1.32,P = 0.593;(4)Tumor size was independently associated with MVI(OR: 1.34,95% CI: 1.22 ～ 1.46,P < 0.001).The tumor increased by 1 cm and the risk of MVI increased by 34%.2.MVI prediction models constructed by identified variables according to different methods: the four MVI prediction models(models 1～4)constructed based on different variable selection methods contain 2,8,9 and 11 variables respectively,and their AUC(95% CI)are 74.4%(70.2% ～ 78.6%),76.4%(72.4% ～ 80.5%),76.7%(72.7% ～ 80.7%)and 76.9%(72.9% ～ 80.9%)respectively.Only the AUC of model4 and model 1 was statistically different(AUC difference 0.025,95% CI: 0.002,0.048,P = 0.034).NRI showed that model 4 was improved by 49.8%(95% CI: 0.331～ 0.666,P < 0.01),and IDI was 0.045(95% CI: 0.026-0.063,P < 0.01).Considering that model 4 contains 11 variables and the improvement of IDI is not much higher,the two variables contained in model 1(alpha fetoprotein AFP and tumor size)are selected as the most stable independent predictors.3.Results and comparison of MVI prediction model constructed by machine learning XGboost: Based on the identification and verification results of the above variables,the MVI prediction model with AFP and size variables is constructed by using XGboost with e Xtreme gradient lifting,and the nomogram is drawn.The AUC and95% CI of the predicted MVI after 500 times of bootstrap resampling were 0.831(0.796 ～ 0.867).Taking 0.468 as the best threshold,the sensitivity and specificity of the model were 0.843 and 0.695.The AUC of the model constructed by XGboost is significantly better than that of model 1 constructed by logistic(P < 0.001).The calibration curve shows that the predicted results are in good agreement with the actual results,and the DCA shows that the model has an obvious net benefit in clinical application.4.Online intelligent use of MVI prediction model based on XGboost: Integrate the XGboost model into a web calculator of Internet browser,scan the code and input the AFP and tumor size(cm)index values of patients to obtain the MVI risk value.5.Construct the survival prediction model of patients with HCC postoperative:according to the adjusted R2 and clinical experience,construct the survival prediction model of HCC including the following variables: tumor size,number of tumors and satellite nodules,portal vein invasion,Child-Pugh grade,cirrhosis,age,CA125,AFP and postoperative preventive TACE.The model was presented as a nomogram,and the ROC curves of 2,3,4 and 5 years were drawn by bootstrap resampling 500 times.The AUC and its 95% CI were 76.6(71.3,81.8),73.1(67.8,78.4),68.8(62.9,74.8)and 67.4(60.4,74.5)respectively.The calibration curve showed good consistency of the model,and the decision curve showed obvious benefits in clinical application.The risk score chart shows that the death events increase with the increase of risk score.ConclusionsThis study conclude the following conclusions:1.Through rigorous and detailed analysis of the relationship between inflammatory indexes and MVI of HCC,it is found that NLR,LMR and PLR are not independent related factors of MVI.2.Tumor size was independently associated with MVI of HCC.With the increase of tumor,the risk of MVI increased significantly.3.Through multiple variable screening methods and constructing MVI prediction model,it is finally determined that AFP and tumor size are stable and independent MVI prediction factors.The MVI prediction model constructed by XGboost is better than Logistic regression model.Using web calculator and QR code can be easily and effectively applied to MVI prediction of hepatocellular carcinoma.4.The survival prediction model of HCC constructed according to common clinical indicators performed well in predicting the 2,3,4 and 5-year overall survival of patients,which is convenient for the hierarchical evaluation of patients’ prognosis.