Prognostic Model Of Liver Cancer Based On Machine Learning Method

Objective: Primary hepatocellular carcinoma(HCC)has attracted more and more attention.Comparing the prognostic differences and clinical characteristics of patients with HCC based on the clinical data before hepatectomy,analyzing the adverse prognostic factors of patients with HCC,and establishing the prognostic prediction model of patients with hepatocellular carcinoma based on machine learning method will be helpful to provide some reference for individualized adjuvant therapy after operation.Methods:Principal component analysis and cluster analysis were performed on 34 baseline clinical data of 386 patients with primary hepatocellular carcinoma,And then Kaplan-Meier method and Cox proportional hazard model were used to compare the clinical information characteristics and prognosis differences of each subgroup.Recursive Feature Elimination based on Multiple Support Vector Machine method was used to rank the clinical variables of patients with primary hepatocellular carcinoma who underwent hepatectomy.Five fold cross validation support vector machine is used to determine the optimal feature subset.Then we construct the Nomograms of 1-year and 3-year disease-free survival and overall survival of patients.Logistic regression,random forest,support vector machine,C5.0 decision tree,neural network,Bagging algorithm and Adaboost algorithm were used to construct the prognostic prediction model of 3-year disease-free survival time and3-year overall survival time of patients after hepatectomy.Results: Four subgroups of clinical phenotypes were identified by cluster analysis.There were statistical differences in some clinical phenotypes and prognostic effects among the four groups.The prognostic effect of subgroup 1 was the best,while that of subgroup 3 was the worst.Compared with subgroup 1,the second subgroup had a recurrence risk ratio of 1.32(95% CI: 1.03-1.70),and the third subgroup had a recurrence risk ratio of 3.60(95% CI: 1.97-6.58).When considering the risk of death,the risk of death in the second subgroup increased by 1.43times(95%ci: 1.10~1.86),and the risk of death in the third subgroup was 4.11 times(95%ci: 2.11~8.00).After communication with the clinician,the results of featuresequencing were confirmed to be reasonable.The consistency index of Nomograms of1-year and 3-year disease-free survival and overall survival of patients.were 0.701 and0.706,respectively.The AUC of the model was 0.7229 and 0.7216,respectively,followed by logistic regression model,C5.0 decision tree,Bagging algorithm,random forest and Adabo.The AUC of OST algorithm is 0.7194,0.7091,0.6978,0.6967 and 0.6571,respectively.Conclusion: Cluster analysis can determine the clinical phenotype subgroup of primary hepatocellular carcinoma patients and provide some reference for individualized adjuvant therapy after operation.The accuracy of prediction model after MSVM-RFE has been improved,and the Nomograms can provide information of patients’ survival risk in clinical practice,which can simply and clearly reflect the survival risk of patients.Support vector machine and neural network have good accuracy in predicting survival and recurrence of patients,and help doctors to judge prognosis and therapeutic effect of patients.