Artificial neural network model to predict post-hepatectomy early recurrence of hepatocellular carcinoma without macroscopic vascular invasion

Background The accurate prediction of post-hepatectomy early recurrence (PHER) of hepatocellular carcinoma (HCC) is vital in determining postoperative adjuvant treatment and monitoring. This study aimed to develop and validate an artificial neural network (ANN) model to predict PHER in HCC patients without macroscopic vascular invasion. Methods Nine hundred and three patients who underwent curative liver resection for HCC participated in this study. They were randomly divided into derivation (n = 679) and validation (n = 224) cohorts. The ANN model was developed in the derivation cohort and subsequently verified in the validation cohort. Results PHER morbidity in the derivation and validation cohorts was 34.8 and 39.2%, respectively. A multivariable analysis revealed that hepatitis B virus deoxyribonucleic acid load, γ-glutamyl transpeptidase level, α-fetoprotein level, tumor size, tumor differentiation, microvascular invasion, satellite nodules, and blood loss were significantly associated with PHER. These factors were incorporated into an ANN model, which displayed greater discriminatory abilities than a Cox’s proportional hazards model, preexisting recurrence models, and commonly used staging systems for predicting PHER. The recurrence-free survival curves were significantly different between patients that had been stratified into two risk groups. Conclusion When compared to other models and staging systems, the ANN model has a significant advantage in predicting PHER for HCC patients without macroscopic vascular invasion. Supplementary Information The online version contains supplementary material available at 10.1186/s12885-021-07969-4.


Background
Hepatocellular carcinoma (HCC) is the sixth most common form of malignancy and the fourth leading cause of cancer deaths worldwide [1]. The most common curative treatment for early HCC patients is hepatectomy [2][3][4]. However, its effectiveness is limited by the high incidence of tumor recurrence in the postoperative period (up to 60%), leading to poor long-term survival in HCC patients [2][3][4]. Therefore, accurate prognostic prediction of postoperative tumor recurrence is consequential in the screening and choice of adjuvant therapies for high-risk patients.
Post-hepatectomy tumor recurrence can be divided into early or late recurrence using 2 y as the cut-off point [5][6][7][8][9]. Numerous studies have reported that posthepatectomy early recurrence (PHER) is associated with intrahepatic metastases from primary tumors that cannot be clinically detected, while late recurrence results from tumor formation following liver cirrhosis [10]. Patients with PHER typically have worse long-term survival prognoses than those with late recurrence [11]. PHER is key in the poor prognosis of HCC patients following curative hepatectomy. Therefore, the establishment of an accurate, reliable, and specific PHER prediction model may provide a reliable means for choosing postoperative adjuvant treatments in high-risk patients, such as radiofrequency ablation (RFA), transcatheter arterial chemoembolization (TACE), or sorafenib.
Artificial neural network (ANN) models are established mathematical tools that mimic the features of neurons. They have data distributions with large-scale parallel structures and processing mechanisms similar to the biological brain [12]. ANN models have been used to process information in extremely complicated biological systems containing multiple related factors [13,14]. Recently, ANN models have been used in the prognostic assessment of various cancers [15][16][17]. However, few studies on its application in HCC have been published. This study aimed to develop an ANN model to assess PHER risk of HCC patients without macroscopic vascular invasion who underwent hepatectomy. Its predictive ability was compared with a Cox proportional hazards (CPH) model, several preexisting recurrence models, and commonly used staging systems.

Patients
Data from HCC patients who underwent liver resection from September 2013 to December 2019 in our hospital were enrolled in this study. The inclusion criteria were: preoperative Child-Pugh score of A or B, underwent a curative liver resection, and HCC diagnosis from postoperative pathology. The exclusion criteria were: had received TACE, RFA, or systemic treatment; tumor invasion in main hepatic veins, portal veins, or adjacent organs; hospital mortality after liver resection; and incomplete clinical data. The 903 HCC patients who met the criteria were randomly classified into derivation (n = 679) and validation cohorts (n = 224) in a ratio of 3:1 ( Supplementary Fig. 1). This study was approved by the Ethical Committee of the Guangxi Medical University Cancer Hospital and was performed in compliance with the Helsinki Declaration. Written informed consent was obtained from all patients.

Surgical procedures and follow-ups
Liver resection was performed when preoperative imaging indicated that all tumors can be resected within the hepatic functional reserve. Major hepatectomy was defined as a resection of ≥ three Couinaud's segments [26]. Additional details and indications for liver resection are described in a previous study [27].
All patients were followed-up every two to 3 months until death or withdrawal. Routine follow-up included the determination of laboratory parameters, abdominal ultrasounds, and computed tomography (CT) or magnetic resonance imaging (MRI). All serological parameters were assayed and analyzed in the early morning of the first day post-admission. Depending on liver functional reserve, general health status, and disease extent, patients with recurrent tumors underwent re-resection or treatment with RFA, TACE, or sorafenib [28]. HCC recurrence or metastasis was determined based on CT and/or MRI results, regardless of whether serum αfetoprotein (AFP) levels were elevated [29]. HCC recurrence was defined as the imaging of new lesions in the residual hepatic tissue or the occurrence of distant metastases. We defined PHER as within 2 y post-surgery from the date of the hepatectomy to the date of the first diagnosis of HCC recurrence [5].

Development of the CPH and ANN models
The CPH and ANN models were established based on the identification of significant PHER prognostic factors using univariable and multivariable Cox analyses. The CPH model is similar to multiple linear regression in that it explores the relationship between a hazard and its associated independent explanatory factors over a period of time. It describes the impact of risk factors on a patient's treatment using a parameter called the risk ratio [30]. Here, the CPH model used the sum of the relevant risks affecting the hazard function to predict PHER risk in HCC patients without macroscopic vascular invasion.
The ANN model was established via a multilayer perceptron network (MLP), which is a popular architecture and a new layer feed-forward neural network design tool consisting of input nodes, hidden layers, and an output node [13]. MLP models are always trained using back-propagation algorithms. When data is provided to neural groups through the input layers, the firstlayer neurons propagate the weighted data and randomly select the biases using the hidden layers. When the net sum at the hidden nodes is confirmed, the transfer function can then be used to provide output responses at the nodes [16,17].
In this study, eight prognostic indicators (hepatitis B virus deoxyribonucleic acid [HBV-DNA] load, γglutamyl transpeptidase [GGT] levels, AFP levels, tumor size, tumor differentiation, microvascular invasion [MVI], satellite nodules, and blood loss) were selected as the input nodes, and one indicator (with and without PHER) was used as the output node. Details of the ANN model are described in a previous study [14]. To avoid over-optimization, the results reported in this study were the optimal results following repeated randomized trials.

Statistical analysis
We randomly divided patients into derivation and validation cohorts, as detailed above (see section Patients). The cut-offs values for the included parameters were confirmed using Youden's index (i.e., sensitivity + specificity − 1) and other published reports [27,29]. These categorical variables are presented as frequencies and percentages and then compared using a chi-squared test. Recurrencefree survival (RFS) curves were assessed via the Kaplan-Meier method and compared using the log-rank test. The predictive abilities of the ANN, CPH, some preexisting recurrence models, as well as commonly used clinical staging systems were calculated using the areas under the receiver-operating characteristic (ROC) curves (AUCs) and decision curve analyses (DCA) [31]. Calibration plots were applied to test the calibration capacity of the ANN model. For its clinical application, Youden's index was calculated to determine the optimal cut-off value for predicting PHER. All patients were then classified as high-risk and low-risk groups.

Baseline characteristics
Baseline characteristics are presented in Table 1; they did not significantly differ between the derivation and validation cohorts (P > 0.05 for all comparisons).

Identification of independent prognostic factors
The univariable Cox analysis revealed that hepatitis B surface antigen level, HBV-DNA load, antiviral therapy, albumin level, aspartate aminotransferase level, GGT level, prothrombin time, Child-Pugh grade, AFP level, tumor size, tumor differentiation, MVI, satellite nodules, surgical time, blood loss, blood transfusion, and major resection were all related to PHER ( Table 2, P < 0.05 for all comparisons). Accordingly, these variables were selected for the multivariable model analysis. The Cox multivariable analysis indicated that HBV-DNA, GGT level, AFP level, tumor size, tumor differentiation, MVI, satellite nodules, and blood loss were significantly associated with PHER ( Table 2). The RFS curves of these eight prognostic factors are shown in Supplementary  Fig. 2.

Development and validation of an ANN model for predicting PHER
The CPH and ANN models were established according to the eight prognostic factors listed above (Table 2). For the ANN model ( Fig. 1), the program file (Supplementary File) was downloaded onto our computer, which then automatically calculated the PHER risk in clinical     Fig. 2b). The prediction probability plot (Fig. 2c) indicated that the ANN model can accurately identify patients without PHER. The AUCs of the ANN model for assessing PHER risk were 0.753 (95% CI: 0.715-0.792) and 0.736 (95% CI: 0.668-0.803) in the derivation and validation cohorts, respectively (Fig. 2d). Moreover, calibration plots detected a good correlation between prediction and observation of the two cohorts ( Fig. 2e  and f).

The predictive ability of the ANN model when compared to other models and staging systems
We compared the predictive ability of the ANN model with other models and staging systems using their AUC values and net benefits. As shown in Table 3, the ROC analysis revealed that the AUC of the ANN model (0.753) was larger than all other models and staging systems (corresponding AUCs: 0.489-0.733, P < 0.05 for all comparisons; Fig. 3a). Moreover, the DCA plot indicated that the ANN model demonstrates better net benefit with a wider threshold probability range (Fig. 4a). Accordingly, the ANN model was superior in predicting PHER in the derivation cohort. Similar results were obtained in the validation cohort. Although the AUC values between the ANN and some models (i.e., CNLC stage and CPH model) did not significantly differ, the ANN model still had the largest AUC value and net benefit (Figs. 3 and 4b, b).

Risk stratification performance based on the ANN model
According to the Youden's index, the optimal cut-off value for the ANN model to predict PHER was 0.37, and its sensitivity and specificity were 72.0 and 68.6%, respectively. Thus, we obtained a risk stratification of two groups, low risk (≤0.37) and high risk (> 0.37). Furthermore, in the derivation and validation cohorts, the RFS curves for all patients were stratified by these risk groups (Fig. 5). These results suggest that the high-risk group is closely related to poorer prognosis (P < 0.001 for all comparisons).

Discussion
This retrospective study developed an ANN model to predict PHER in HCC patients without macroscopic vascular invasion. Our ANN model achieved satisfactory discriminatory and calibration capacities in both the derivation and validation cohorts. In addition, it demonstrated greater prediction capacity than a CPH model, some preexisting recurrence models, and commonly used staging systems. Finally, the ANN model stratified patients into two risk groups, highlighting the significant differences in RFS between different risk groups.
The high incidence of PHER remains a major challenge for clinicians [2][3][4]. Early tumor recurrence is associated with intra-hepatic metastasis, while late tumor recurrence is closely related to multi-centric oncogenesis [6,7]. Despite similar adjuvant treatments, patients with PHER have significantly worse outcomes than patients with late recurrence [10,11]. To prevent PHER and prolong the longterm survival of HCC patients, the effectiveness of different postoperative adjuvant treatments including RFA, interferon, TACE, and immunotherapy have been evaluated [16,32,33]. Ueno et al. [33] reported that TACE adjuvant can reduce PHER risk, but it cannot reduce the risk of late recurrence. In contrast, Jiang et al. [34] found that postoperative adjunctive TACE did not improve RFS and overall survival in HCC patients after curative liver resection. In fact, Ahmed et al. [35] have suggested that TACE may worsen patients' quality of life. Systematic reviews and meta-analyses have failed to find strong evidence to support the use of these adjuvant treatments [36,37]. These discrepancies can be attributed to the heterogeneity of patient groups in randomized and non-randomized controlled studies and their consequent differences in adjuvant treatment outcomes. Indeed, if these studies focused only on the impact of adjuvant treatments on postoperative tumor recurrence in high-risk patients, the results may have been different. Therefore, the identification of high-risk patients with PHER is clinically significant; these patients can become the focus of future clinical studies on adjuvant therapies. An accurate, reliable, and specific prediction tool based on ready-made prognostic variables can accurately identify patients at high-risk for PHER and therefore be an effective method to address this clinical problem. The prediction model may also closely relate to the monitoring of postoperative tumor recurrence. In theory, high-risk patients should adopt more aggressive and effective monitoring programs such as the use of more accurate radiological assays to detect PHER and thereby allow more timely remedial measures to be taken. For example, some centers perform preventive remedial transplants after hepatic resection in high-risk patients to prolong their long-term survival [38].
Numerous clinical and pathological prognostic factors have been identified to cause tumor recurrence following curative hepatectomy. However, few studies have been published on the establishment of accurate and effective models to predict PHER. To date, most proposed models do not focus specifically on it. These models include: the recurrence clinical risk score from Lee et al. [39], the Shanghai score from Sun et al. [40], the recurrence score for hepatitis B virus-related HCC proposed by Qin et al. [41], a RFS nomogram for AFP-negative patients created by Gan et al. [42], early recurrence after Table 3 The performance of the ANN model and other models and staging systems in predicting early recurrence in derivation cohort and training cohort  surgery for liver tumor models built by Chan et al. [5], the Hong Kong recurrent model developed by Ng et al. [8], the recurrence after curative hepatectomy score constructed by Tokumitsu [43], and some radiomics-based prognostic prediction models [44][45][46][47]. Although these models are accurate, they were created using traditional linear models such as CPH model and survival analysis. The correlation between different risk factors is multidimensional, complex, and non-linear. Therefore, correlational analyses between these factors are limited when they are applied using only traditional linear methods. An ANN model is probably more effective when multiple risk factors are involved in multidimensional and complex functions that interact with each other [12][13][14][15][16].
In this study, based on the eight most important prognostic factors, we developed an ANN model that performed better in predicting PHER in HCC patients without macroscopic vascular invasion than a CPH model as well as some preexisting recurrence models (Figs. 3 and 4, Table 3).  Our data revealed that HBV-DNA, GGT levels, AFP levels, tumor size, tumor differentiation, MVI, satellite nodules, and blood loss were associated with PHER (Table 2). These factors are common and readily assayed in clinical practice. Past studies have reported that the higher the HBV-DNA load, the greater the risk of low survival and tumor recurrence after hepatectomy [48][49][50]. High GGT levels may lead to liver dysfunction by inducing deoxyribonucleic acid instability, tumorigenesis, and cancer progression. Additionally, high GGT levels predict poorer HCC prognosis [51,52]. High AFP levels indicate that the tumor is highly aggressive; the probability of intrahepatic metastasis is usually greater in patients with high AFP levels [53]. Others have reported that larger tumor sizes are significantly associated with PHER [54] and that tumor differentiation, MVI, and satellite nodules relate to more severe PHER [55][56][57]. Finally, excessive blood loss often leads to systemic inflammatory reactions and reduces immunity, leading to increased risk of serious complications and tumor recurrence after surgery. All these parameters are easily found in medical records, thereby facilitating routine use of the ANN model relative to other models that use complex radiological variables [44][45][46][47].
The ANN model built in this paper is more accurate in predicting PHER than commonly used staging systems (Figs. 3 and 4, Table 3). The increased accuracy is likely because these systems and models contain very few parameters and try to balance risk variables by summarizing them. Accordingly, simplified models may limit PHER prediction accuracy in HCC patients. Moreover, these systems and models are linearly additive forms based on prognostic variables, and the interactions between prognostic variables cannot be accurately delineated. ANN models, in contrast, contain a wide range of predictors and manage the interactions between all prognostic factors, consequently improving their predictive power.
Currently, no clear consensus or guidelines exist on the ideal adjuvant therapy after hepatectomy. Moreover, the criteria for identifying patients at high risk for PHER remain unclear. Notably, the ANN model can be stored in a computer as a program (Supplementary File). After the clinician enters the eight prognostic factors (HBV-DNA, GGT level, AFP level, tumor size, tumor differentiation, MVI, satellite nodules, and blood loss) into the program, it will automatically and accurately calculate PHER risk ( Supplementary Fig. 3). Here, our ANN model had a cut-off value of 0.37, and its sensitivity and specificity, respectively, in assessing PHER risk was 72.0 and 68.6%. All patients were then divided into high-risk and low-risk groups. The risk stratification analysis detected significant differences in RFS curves between the two risk groups ( Fig. 5; P < 0.05 for all comparisons). As expected, patients in the high-risk group had poor RFS; however, the screening of these patients can greatly positively impact adjuvant therapy strategies. For instance, in low-risk patients, the appropriate adjuvant therapy should reduce side effects, especially in elderly patients. In contrast, high-risk patients may need to combine adjuvant treatments to obtain optimal prognoses, particularly in younger patients. A recent meta-analysis reported that in HCC patients, TACE + RFA can provide therapeutic outcomes comparable with hepatectomy but with the advantage of reduced morbidity [58]. This finding is important for clinical decision making in high-risk patients, as it is worth considering whether a patient deemed to be at high risk of PHER should undergo a hepatectomy. Accordingly, high-risk patients must be closely monitored, and appropriate treatment options must be explored. Finally, the ANN model allows the stratification of PHER risk in patients in a way more appropriate for the design of clinical trials.
Despite the promising data presented here, this study does have some limitations. First, clinicians have no mathematical formula to use directly in the ANN model, which may limit its wide-spread use. Therefore, more convenient algorithms for ANN models should be developed. Also, most of the patients in our study had HBV infections; therefore, further validation in other etiological populations is necessary. Finally, this study is retrospective and was conducted with patients from a single medical center. Thus, prospective studies using patients from several medical centers are required to verify the results obtained here.

Conclusion
An accurate, reliable, and specific ANN model that predicts PHER risk in HCC patients without macroscopic vascular invasion was established and validated in this paper. The ANN model provides valuable data to help identify high-risk patients for future adjuvant therapy and active surveillance studies.