A Powerful Nomogram Based on the Novel D-Index to Predict Prognosis After Surgical Resection of Hepatocellular Carcinoma

Introduction

Liver carcinoma is the sixth most commonly diagnosed cancer worldwide and the fourth main cause of cancer-related death. Hepatocellular carcinoma (HCC) accounts for 75–85% of primary hepatic carcinomas. In China, aflatoxin and chronic hepatitis B are the main risk factors for the high incidence of HCC.¹ Current treatments for HCC include resection, ablation, transarterial embolization, radiotherapy, systemic pharmacological treatment, and liver transplantation.² Liver transplantation is considered the best treatment for HCC, as complete tumor resection eliminates underlying disease (eg, liver cirrhosis).³ However, due to the considerable resources needed for organ transplantation, surgical resection remains the primary treatment choice in China.

Several staging/scoring systems are widely used, including the Tumor, Node, and Metastasis (TNM) staging system (eighth edition),⁴ Barcelona Clinic Liver Cancer (BCLC) staging system,⁵ Cancer of the Liver Italian Program (CLIP) scoring system,⁶ and Okuda staging system.⁷ Many indices have also been used to predict prognostic outcomes. For instance, the spleen stiffness measurement evaluated using transient elastography can reliably predict late recurrence of HCC.⁸ Malnutrition is also a prognostic factor for HCC.⁹ In patients who undergo HCC resection, the systemic immune inflammation index is a potent index of poor prognosis.¹⁰ These systems and indices play an essential role in predicting prognosis and influencing treatment choice.

However, those traditional systems have limitations. The TNM is based only on postoperative diagnosis, and the BCLC system is too complex for clinical use. The TNM and BCLC are reportedly unable to evaluate overall survival (OS) in HCC patients.¹¹ The CLIP may not be accurate when performance status is not included.¹² The Okuda was developed at a time when most patients were diagnosed with advanced-stage liver cancer. Although early diagnosis of cancer has been greatly improved, the predictive capacity of tools has become extremely inadequate.¹² Such factors may discourage the use of traditional staging/scoring systems. Several novel systems have been considered to predict prognosis in HCC patients, including the Italian Liver Cancer¹³ and the Hong Kong Liver Cancer staging systems.¹⁴ Nonetheless, those systems failed to accurately and quantitatively evaluate OS after surgical resection and are not convenient as a single index.

Nomograms are frequently used to predict prognosis in cancer patients, as they can create a single digital estimation of the possibility of a clinical event and can stratify patients in clinical tests by generating individual predictions for each patient.¹⁵ Nomograms can establish biologically and clinically integrated models by assimilating various prognostic and determinant variables, thereby personalizing treatment. Moreover, the predicted prognoses are easier to understand than conventional staging.¹⁶

Our study aimed to establish a novel index and nomogram to predict the prognosis of HCC patients and accurately quantify OS.

Patients and Methods

Patient Screening

For this retrospective study, 682 patients diagnosed with HCC between January 2007 and February 2015 were selected from our hospital database. We included patients who had a) primary tumors, b) tumors completely excised without tumor manipulation during surgery, c) complete laboratory blood tests and imaging examinations, d) pathologically diagnosed HCC, e) no severe postoperative complications, f) no other malignant tumors, g) not undertaken treatment for cancer before surgery, and h) normal brain, heart, and kidney function. Patients were excluded if they a) had incomplete data and b) received additional treatments after hepatectomy such as ablation, transarterial embolization, radiotherapy, systemic pharmacological treatment, or transplantation. Finally, we enrolled 396 patients for the study and randomly divided them into the training and validation groups at a ratio of 7:3.

The first follow-up date was considered the date of surgical resection. All patients were followed up by outpatient or telephone visits every 3 months for the first two years, every 6 months since the third year, and every 12 months after five years. Follow-up ceased on August 1, 2018. OS was defined as the duration from hepatectomy until the date of death or last visit.

Clinical Information and Laboratory Results

Clinical information included age, sex, weight, height, body mass index, and alcohol consumption. Laboratory reports included levels of alpha-fetoprotein, albumin, fibrinogen, total bilirubin (TBIL), total cholesterol, alanine transaminase, aspartate transaminase (AST), γ-glutamyl transpeptidase, neutrophil, monocyte, lymphocyte, platelet, and prothrombin time (PT). The pathological diagnosis and other information were acquired from surgical records, including the presence of ascites fluid, liver cirrhosis, tumor size (TS), tumor capsule, tumor stage, satellite nodules, peri-cancerous invasion, single/multiple tumors, invasion of the biliary duct, metastasis of lymph nodes, vascular invasion, tumor thrombus in the portal vein (PVT), intrahepatic metastasis, degree of tumor differentiation, and adjacent invasion (AI).

Radiographic Results and Imaging Analysis

All patients underwent preoperative abdominal noncontrast computed tomography (CT) to evaluate the general abdominal cavity. Two experienced radiologists (with 5 years of imaging experience), who were blinded to the patients’ clinical data, analyzed the CT images using a postprocessing station (GE Healthcare Advantage Workstation, version 4.6) with the axial image at the level of the L3 vertebra. A cross-sectional CT image at the inferior aspect of the third lumbar vertebra (L3) was selected for estimating muscle area (TAMA), subcutaneous fat area (SFA), and visceral fat area (VFA) as described previously.^17–21 Adipose tissue was distinguished from other tissue by using the Hounsfield scale, and the boundaries were outlined manually as needed. The predetermined Hounsfield unit (HU) thresholds were −29 to −150 HU for TAMA, −30 to −190 HU for SFA, and −50 to −150 HU for VFA (Supplemental Figure 1). Fat area was defined as the sum of visceral and subcutaneous fat areas. We also calculated the ratio of fat-to-muscle area (RFM).

Statistical Analysis

To improve the robustness and reliability of this study, the enrolled 397 participants were randomly split into a training set and another separate validation set at a ratio of 7:3 without replacement. The comparability of the two sets was then evaluated (Table 1). Continuous variables with normal distribution are presented as means ± standard deviation (x ± s). For variables with a skewed distribution, the median (1st quartile, 3rd quartile) was utilized. Categorical variables are represented as frequencies (proportion). The independent-sample t-test or Mann–Whitney U-test was used for continuous variables with normal or skewed distributions, respectively. The chi-square test or Fisher’s exact test was used to compare categorical variable.

Table 1 Baseline Characteristics of the Study Participants in the Training and Validation Cohorts

Univariate and multivariable Cox analyses were utilized to filter possible indicators and estimate their weights in the training set. Significant risk factors (P < 0.05) of the univariate analysis were included in the multivariate analysis. Factors with p-values less than 0.05 in the multivariate Cox analysis were retained in the associated models. After then, a candidate nomogram model was built depending on the five most significant risk factors, including D-index. Survival curves were plotted using the X-tile software and Kaplan-Meier method from the diagnosis date to the death or last follow-up date in the training and validation sets, respectively. Survival curves were assessed using Log rank tests. The curve (AUC) of the receiver operating characteristic (ROC) analysis and the consistency index (C-index) were used to compare the predictive performance of the nomogram and other models. The clinical net benefit of the D-index and all systems mentioned in this study were evaluated using decision curve analysis (DCA). R version 3.6.3 and SPSS version 25.0 (SPSS, Chicago, IL, USA) were used for the statistical analyses.

Study of the Novel D-Index

To select the most important indicators from the clinical diagnostic variables and generate a model for the D-index, the laboratory and radiographic indices of all 396 patients were included in the logistic regression analysis. Important variables linked to Cox regression were subsequently chosen. Related indicators that would determine participant mortality (P < 0.1) were chosen using univariate analysis. The variance inflation factor was used to evaluate potential collinearity between the screened variables. Based on the partial regression coefficients of multivariate analysis, five indicators, including preoperative PT, TS, neutrophil counts, TBIL levels, and RFM, were added as candidate variables (Supplemental Table 1). Considering the partial regression coefficients, all potential combinations of candidate indicators were modeled, and the area under the ROC curve (AUC) of each model was compared. According to Occam’s razor, we selected three indicators and built a formula based on their mathematical relations.²² We obtained the D-index, which has a correspondingly large diagnostic value. The larger the D-index, the greater the mortality risk. The D-index was calculated as follows:

The D-index cutoff with the best sensitivity and specificity was obtained using Youden’s index. The best D-index cutoff was 37.9 (AUC 0.71, 95% confidence interval [CI] 0.66–0.76) (Supplemental Table 2).

Results

Clinical Variables

In total, 277 participants comprised the training cohort, and 119 comprised the validation cohort (ratio 7:3). Comparisons between both cohorts and baseline variables are shown in Table 1. There were no significant differences in baseline variables between both cohorts (P ≥ 0.05). The OS rates at 1, 3, and 5 years in the training and validation cohorts were 88.4%, 70.8%, and 56.3%, and 83.2%, 67.2%, and 58.8%, respectively; 63.2% and 64.7% of participants in the training and validation cohorts, respectively, had a D-index of ≥ 37.9. The mean survival time and median follow-up time were 3.77 and 5.33 years, respectively, in the training cohort and 3.65 and 5.12 years, respectively, in the validation cohort.

Predictive Indicators of the Training Cohort

Results of the Cox proportional hazard analyses in the training cohort are summarized in Table 2. Because the D-index was a combination of three risk indicators and every scoring system was independent, the staging systems and the three factors of TBIL, TS, and RFM were not included in this process. Furthermore, the predictive capability of the D-index and the three variables that informed it was quantified using the AUC (Supplemental Table 3). According to the results, we believe that D-index is clearly advantageous over TBIL, TS, and RFM in terms of AUC and Youden’s index. It may be a powerful, optional indicator for the clinicians to consider. Both univariate and multivariate Cox proportional hazards analyses included criteria with P < 0.05. The indicators with multivariate Cox analysis revealed five important independent risk indicators: D-index, PT, AST, PVT, and AI.

Table 2 Univariate and Multivariable Cox Proportional Hazard Analyses of the Training Cohort

Generation and Validation of the Prognostic Nomogram

We generated a nomogram based on the five indicators (Figure 1). Every chosen factor was assigned a correlated score based on its value in the nomogram. A sample vertical line generated by the total score helped predict 1-year, 3-year, or 5-year survival after the total score was calculated. Ascertaining the interval where the total score was located enabled obtaining the estimated survival possibility easily. Concordance indices (C-indices) were calculated at three time points in both cohorts to evaluate the nomogram’s predictive performance. The C-indices of the training and validation sets at 1, 3, and 5 years were 0.780 (95% CI 0.705–0.854), 0.733 (95% CI 0.680–0.786), and 0.727 (95% CI 0.683–0.772), and 0.815 (95% CI 0.726–0.903), 0.758 (95% CI 0.686–0.830), and 0.739 (95% CI 0.670–0.808), respectively (Supplemental Table 4). At each time point, the C-indices of both cohorts performed similarly, indicating the robustness of the nomogram in predicting OS. The cutoff point of the nomogram was detected by X-tile analysis and adopted to estimate incompatible HCC-related death risk in the training set (Figure 2). Participants were separated into three groups using the X-tile software. These three groups followed the cutoff nodes described above in the validation set (low risk: < 69.2; medium risk: 69.2119.9; high risk: ≥ 119.9). The cumulative survival rates according to risk are shown in Figure 3. Compared to participants in the medium- or low-hazard group, individuals in the high or medium risk group had hazard ratios of 3.67 (95% CI 1.54–8.76) and 2.14 (95% CI 1.05–4.35), respectively, in the validation set. The calibration (Figure 4) and Kaplan-Meier curves are also displayed in the training and validation cohorts (Figures 2 and 3). The calibration curves showed high consistency in predicting survival in patients with primary hepatocellular carcinoma (PHCC) in both cohorts. The two sets of calibration curves were compatible with the prediction outcomes of the Kaplan-Meier curves.

Figure 1 Nomogram of the D-index for predicting the overall survival after curative resection of PHCC. The scores of each variable were added to obtain the total score, and a vertical line was drawn on the total score to obtain the corresponding survival probability. Abbreviations: PHCC, primary hepatocellular carcinoma; PVT, portal vein tumor thrombus; AI, adjacent infiltration; PT prothrombin time; AST, aspartate transaminase.

Figure 2 Results of using X-tile analysis by total risk score calculated by the nomogram scoring system in the training cohort.

Figure 3 Survival curves stratified by total risk score (low risk: < 69.2; medium risk: 69.2–119.9; and high risk: ≥ 119.9) in the validation cohort.

Figure 4 Calibration curves for predicting the overall survival rate by nomogram scoring system in the training and validation cohorts. Calibration curves of the prognostic nomogram for (A) 1-year overall survival, (B) 3-year overall survival, and (C) 5-year overall survival in the training set, and calibration curves for (D) 1-year overall survival, (E) 3-year overall survival, and (F) 5-year overall survival in the validation set.

Predictive Ability of the Nomogram Model

To further confirm the predominance of the nomogram (D-index, PT, AST, PVT, and AI) for assessing prediction in PHCC patients, we depicted ROC curves to compare predictive ability among the D-index, prognostic nomogram, CLIP, TNM, Child–Pugh, BCLC, Okuda, and the combined group (CLIP, TNM, Child–Pugh, BCLC, and Okuda) (Figure 5). The performance of all methods at 1-, 3-, and 5-year OS was completed using AUC values; the D-index AUC values were 0.8103, 0.7479, and 0.7338, respectively. The D-index performed better than all ordinary systems (CLIP, TNM, Child–Pugh, BCLC, and Okuda) for 1-, 3-, and 5-year OS. Furthermore, the Child–Pugh was the worst predictor (0.6038) of 3-year OS, and the Okuda was the worst predictor of 1- and 5-year OS (0.6315 and 0.5576, respectively). Overall, the nomogram showed the greatest accomplishment. For 1-year OS, the nomogram’s C-index was > 0.8 in the validation cohort. The nomogram was the only one that performed better than the combined group (1-year OS, 0.8251; 3-year OS, 0.7462; 5-year OS, 0.6776).

Figure 5 ROC curve of the D-index, prognostic nomogram, CLIP, TNM, Child–Pugh, BCLC, Okuda system, and combined group in the training cohort. ROC curve for (A) 1-year survival, (B) 3-year survival, and (C) 5-year survival. The combined group included the CLIP, TNM, Child–Pugh, BCLC, and Okuda. Abbreviations: CLIP, Cancer of the Liver Italian Program; TNM, Tumor, Node, Metastasis classification; BCLC, Barcelona Clinic Liver Cancer; ROC, receiver operating characteristic.

Decision Curve Analysis

As shown in the DCA (Figure 6), disregarding how large the threshold possibility was (excluding the 0.35–0.45 range wherein the nomogram was concurrent with the D-index), the nomogram was the best predictor in most cases, achieving the highest clinical net profit. In contrast to the traditional staging systems, the D-index is clinically important. When the threshold was within 0.2–0.5, the net profit of the nomogram including the D-index as well as the D-index itself was larger than that of the other systems. These values could provide a more practical and beneficial predictive ability. As the boundary was > 0.5, the clinical net profits to patients decreased significantly.

Figure 6 DCA curve of the D-index, TNM, BCLC, Okuda, CLIP, and prognostic nomogram in the training cohort. The horizontal axis represents the threshold value, which is the reference probability of whether a patient receives treatment, and the vertical axis represents the net benefit rate after the advantages minus the disadvantages. Under the same threshold probability, the larger net benefit implies that patients can obtain the maximum benefit using the diagnosis of this model. The closer the curve in the DCA graph to the top, the higher the value of the model diagnosis. Abbreviations: DCA, decision curve analysis; TNM, Tumor, Node, Metastasis classification; BCLC, Barcelona Clinic Liver Cancer; CLIP, Cancer of the Liver Italian Program.

Discussion

Nomograms are more practical and accurate than other staging systems in building a predictive model for certain tumors.^23–25 Here, we generated and validated a novel nomogram to accurately forecast survival in HCC patients. The nomogram considered five risk factors: D-index, PT, AST, PVT, and AI. Participants were divided into three series (low, moderate, and high risk) using X-tile analysis. Calibration curves showed high consistency between predicted and actual values in the training and validation cohorts, confirming the reliability of our nomogram for reuse. The nomogram better predicted OS than other ordinary systems at 1 (AUC: 0.8486), 3 (AUC: 0.7785), and 5 years (AUC: 0.752). Moreover, the DCA, generally utilized to obtain the maximum net profit and considered to be more accurate than the ROC curve,²⁶ was used to examine the nomogram’s clinical performance. DCA can be utilized to combine clinical effects and compare prognostic models.^27,28

Compared with other nomograms, our nomogram had relatively higher sensitivity and accuracy owing to the D-index. Liao et al²⁹ built a nomogram that included TS, tumor number, microvascular invasion, and NMLR to predict postoperative OS in HCC patients; the nomogram demonstrated great predictive capability at 3 years (AUC: 0.821) but a general one at 5 years (AUC: 0.664). Chen et al³⁰ reported that tumors were associated with diverse and complex factors; however, their S-index only combined blood indicators, suggesting that nomograms based on multiple factors have better predictive performance [1-year (AUC: 0.738), 3-year (AUC: 0.7293), 5-year (AUC: 0.752)]. Our D-index not only combined blood indicators but also included tumor characteristics and size, and we considered body composition by calculating the RFM. Thus, our nomogram has superior performance than other nomograms and better calibration ability than that created by Chen et al.

Our D-index was associated with TBIL levels, TS, and the RFM. Elevated TBIL levels can appear when hepatic cells are damaged, increasing the risk of hepatic fibrosis among HBV-infected patients³¹ and increasing the invasiveness of HCC.³² Therefore, HCC patients with high bilirubin are likely to have a poor prognosis, and TS influences the 5-year all-cause mortality.³³ Preoperative sarcopenia and obesity are considered risk factors for poor prognosis in patients who underwent HCC resection.^34–37 This suggests that higher fat and lower muscle contents are related to a poor prognosis. In our D-index formula, the RFM can be regarded as the quantification of fat and muscle tissue content. However, the formula infers that higher fat and lower muscle contents predict a favorable prognosis, which is contrary to the abovementioned studies. Zhang et al³⁸ reported that sarcopenia may be triggered by cachexia and malnutrition caused by cancer; they found that patients with advanced cancer had a higher risk of skeletal muscle consumption. One possible explanation is that our patients were at an earlier stage of cancer, without severe muscle decomposition. Furthermore, in the early stages of cancer, the body mainly mobilizes fat against the high consumption of malignant tumors. Alternatively, a lower RFM may imply a more serious consumption condition and, therefore, a poorer prognosis. However, these are mere speculations and require further investigation.

Our nomogram involved five independent risk factors. Among them, PVT, AI, AST, and PT have already been shown to correlate with poor prognosis after HCC resection.^39–43 In terms of these risk factors, our research is consistent with previous studies.

Our nomogram has several advantages. The D-index had better predictive power than the conventional systems, so our nomogram had favorable stability and accuracy. The variables were easy to obtain from clinical practice and were objective, avoiding the effects of surgeon subjectivity. Moreover, implementing the scoring system was simple and convenient. By producing an accurate prognosis, this nomogram could help to select specific treatment regimens for patients under variable conditions. Furthermore, the nomogram could help surgeons to stratify patients and provide more tailored treatment.

However, our study has some limitations. First, it was restricted to a single center; thus, our results need to be verified by multicenter investigations. Second, the nomogram was mainly based on baseline levels of the D-index, PT, AST, PVT, and AI, whose levels might have not remained steady during the whole phase, affecting accuracy, especially for 3- and 5-year survival. Finally, this model is not suitable for patients who receive treatments other than PHCC resection. There are far more factors to be investigated and used to predict the prognosis of HCC.

Conclusion

Our team managed to use a novel index and nomogram to predict OS in PHCC patients undergoing liver resection. Our nomogram showed a powerful predictive ability and could help clinicians make better therapeutic decisions. We hope our results can be verified further and widely used.