Artificial Neural Network Model For Predicting Survival Of Patients With Early Hepatocellular Carcinoma Following Hepatectomy

Background and aimHepatocellular carcinoma (HCC) is one of the most common malignancies worldwideand the major component of tumor-related mortality. However, as routine screening ofhigh-risk patients (such as patients with hepatitis B or C) has been implemented, HCC isbeing diagnosed at earlier stages in many patients. Even in the patients with early HCC, theprognosis varies distinctively under different clinicopathological characteristics. Theinteraction among these factors is complicated and nonlinear mode, thus, the conventionallinear models and the presented prognostic staging system we often used cannot predictaccurately. Artificial Neural Network (ANN) is the abstraction and simulation of a numberof basic characteristics of the human brain or natural neural network, also, it is nonlinearand has the function of learning, cross-validation and testing. The aim of our study isdeveloping the ANN model which is used to predict the prognosis of the early HCC on thebasis of patients’clinicopathological characteristics and comparing of the predictive abilityof ANN model and the linear models (logistic regression model and Cox proportionalhazard model) and other prognostic staging system (TNM stage, BCLC stage, IHPBAstage, and Vauthey stage) which are widely used. In this study, the early HCC is defined aspatients with HCC meeting the Milan Criteria.Method1. Patient informationWe retrospectively collected a base population of patients who had early HCC andunderwent partial hepatectomy between February2000and March2008at the EasternHepatobiliary Surgery Hospital (EHBH). After screening the exclusion criteria,725casesof early HCC meeting the Milan Criteria are included in the study.104patients werecollected as external cohort at the First Affiliated Hospital of Fu Jian medical school.2. Surgical approachThe criteria of radical resection is that all tumors are completely removed; no residualtumor confirmed by intraoperative B ultrasound; negative cutting edge by histologicalexamination; the high preoperative AFP level declines to normal in2months; nointrahepatic new lesions2months after the operation by imaging examination.3. Follow-upAll patients underwent standardized follow-up. The end of follow-up time was May 1st2011, the median follow-up time is42.5(2.1-135.6) months.4. Statistical methodsUsing NeuroSolution Version6.0(Neurodimension, Florida, USA) to develop theANN model and process the data. The cut-offs of tumor size and AFP level, the predictivevalue of ANN, logistic and Cox models is evaluated by Receiver Operative Curves (ROC).Results1. Clinicopathological characteristics of patientsIn our study of725patients,631males and94females and the proportion is87%and13%respectively. The range of age is from10to78years old. As the development of ANNmodel, all patients from EHBH are randomly divided into the training cohort (543patients75%) and internal testing cohort (182patients25%). Adding with the external cohort with104patients, the clinicopathological characteristics of the three groups have no significantdifference.2. Univariate and multivariate analysisMultivariate analysis in logistic regression showed that tumor size (OR:2.2644, P=0.0007), tumor number (OR：1.7654，P=0.0066), AFP level (OR:1.7618, P=0.0059), MVI(OR:2.6937, P <0.0001) and tumor capsule (OR：1.7407，P=0.0069) were the independentprognostic risk factors to the5-years survival. The Cox proportional hazard regressionshowed similar results.3. The development of Artificial Neural Network modelThe artificial neural network consists of a set of highly interconnected processingunits (neurons) tied together with weighted connections. The network itself consists of aninput layer, an output layer, and one or more hidden layers. The input layer include the riskfactors in the univariate analysis of logistic regression, when comparing with Cox model,we need add the survival time as covariate to develop the ANN model.4. Comparing of ANN model with logistic and Cox modelsUsing Area Under Curves (AUC) of ROC to compare the predictive ability of all themodels. The results showed: ANN（0.784）was larger than Logistic（0.758）and there wassignificant difference （P=0.0009）, also, ANN（0.855）was larger than Cox（0.826）andthere was significant difference（P=0.0115）. The testing cohort showed the similar results.5. Comparing of ANN model with other HCC staging systemsThe AUC results showed that ANN （0.784） larger than IHPBA （0.711），TNM（0.639），Vauthey （0.639）and BCLC （0.612）. It also proved in the testing cohort. 6. Testing the performance of ANN modelWhether in the internal testing cohort or in the external testing cohort, ANN modelwas superior to the logistic and Cox models, meanwhile, comparing with the widely usedprognostic staging systems, the ANN model also had a better performance.ConclusionHaving been proven by the testing cohorts, Artificial Neural Network modelperformed significantly better than the conventional linear models and HCC stagingsystems.