Prognostic nomogram model for overall survival in patients with hepatocellular carcinoma

Abstract

BACKGROUND

Hepatocellular carcinoma (HCC) poses a significant health challenge in China because of its high incidence and mortality rate.

AIM

To determine the prognostic factors for HCC patients in southwestern China and develop a novel model to predict overall survival.

METHODS

A total of 958 primary HCC patients were included in this study. Multivariate Cox regression analysis revealed independent prognostic factors, which were subsequently used to construct a nomogram. The performance of the nomogram was evaluated using the concordance index (C-index), area under the receiver operating characteristic curve, time-dependent C-index and area under the receiver operating characteristic curve, calibration curves, and decision curve analysis.

RESULTS

Multivariate Cox regression revealed the following independent prognostic factors: (1) Karnofsky performance status; (2) Sex; (3) Diabetes status; (4) Tumor node metastasis stage; (5) Chemotherapy, surgery; (6) Fibrinogen degradation products; (7) β2-microglobulin level; (8) Albumin; and (9) Neutrophil-to-lymphocyte ratio. For the training set, the C-index was 0.752 (95%CI: 0.731-0.774), whereas for the validation set, it was 0.716 (95%CI: 0.680-0.752). The calibration curves for both the training set and the validation set revealed a high level of consistency between the observed values and the model-predicted survival probabilities. Decision curve analysis validated the potential clinical utility of the model.

CONCLUSION

This study developed a novel prognostic model for predicting overall survival in HCC patients, providing a valuable tool to support clinical decision-making and optimize patient management.

Key Words: Hepatocellular carcinoma; Predictive model; Nomogram; Overall survival; Prospective study

Core Tip: A novel prognostic nomogram for overall survival was developed in this study of 958 Chinese patients with hepatocellular carcinoma, using routine clinical variables. This study revealed that Karnofsky performance status, sex, diabetes status, tumor node metastasis stage, treatment, fibrinogen degradation products, β2-microglobulin, albumin, and neutrophil-to-lymphocyte ratio play significant roles in predicting patient outcomes. The model demonstrated excellent predictive accuracy (concordance index of 0.752) and clinical utility, offering a practical tool to improve personalized risk assessment and management of patients with hepatocellular carcinoma in Southwest China.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the predominant form of primary liver cancer, representing 75%-85% of cases. Chronic hepatitis B or C infections, alcohol-related liver disease, metabolic fatty liver disease, and liver cirrhosis are major contributors to the development of HCC[1,2]. Early diagnosis and treatment of HCC are essential for improving patient outcomes. However, early-stage HCC often lacks distinct symptoms[3,4]. As the tumor progresses, patients may present with symptoms such as liver pain, hepatomegaly, jaundice, fatigue, and nausea[5]. Currently, HCC diagnosis relies primarily on imaging studies (e.g., ultrasound, computerized tomography, and magnetic resonance imaging) and evaluation of serum biomarkers, such as alpha-fetoprotein (AFP). However, these methods have limitations in the detection of early-stage HCC[6,7]. The five-year survival rate mainly reflects the poor prognosis of patients with this disease. The 5-year overall survival (OS) rate for HCC patients frequently falls below 20%[8]. China’s survival rate is 14.1%, approximately 10% lower than that of Japan and South Korea (over 27%) and Australia and the United States (over 17%). Accordingly, mortality rates in China and South Korea (> 10 per 100000) are higher than those in other countries (all < 6 per 100000). Consequently, enhancing the early detection of HCC is crucial for increasing the 5-year survival rate among Chinese patients[9-11].

HCC poses a significant health challenge in China because of its high incidence and mortality rates. Nomograms are commonly used as predictive models to evaluate prognosis in cancer patients[12-14]. Most current HCC prognostic nomograms are based on European, American, Japanese, and Korean populations, and few models are specifically designed for HCC patients in Southwest China. In addition, traditional models primarily focus on tumor stage and liver function and lack integration of inflammatory and immune biomarkers. By analyzing and integrating various clinical and demographic variables, such a model can provide insights into the underlying mechanisms of HCC and reveal novel biomarkers and therapeutic targets[15]. The objective of this research was to identify factors independently associated with the prognosis of HCC patients in Southwest China, develop a nomogram for prediction, and provide a reliable tool for clinical prognostic evaluation.

MATERIALS AND METHODS

Data source

This study analyzed data from 958 HCC patients collected between January 1, 2020, and December 31, 2022.

Inclusion criteria: (1) Age ≥ 18 years; (2) HCC confirmed by pathological examination; and (3) Receipt of at least one standardized antitumor treatment.

Exclusion criteria: (1) Absence of baseline clinical data and follow-up records; and (2) A history of other malignant tumors. Furthermore, patients with other malignant tumors, severe hepatic or renal failure, end-stage disease, who received only symptomatic supportive treatment, or any other conditions potentially affecting OS were also excluded. Follow-up was performed monthly for the first 2 years after the patient’s HCC diagnosis and every 6 months thereafter, with the endpoint on May 31, 2025. Figure 1 illustrates the data processing procedure and the overall research methodology.

Figure 1 Flow chart of the patients enrolled in the final study cohorts. AUC: Area under the receiver operating characteristic curve; DCA: Decision curve analysis; HCC: Hepatocellular carcinoma; LASSO: Least absolute shrinkage and selection operator; OS: Overall survival; ROC: Receiver operating characteristic; TNM: Tumor node metastasis.

Variables

In this study, we examined demographic data, including age, sex, type of medical insurance (%), prevalence of hypertension (%), diabetes (%), Karnofsky performance status (KPS), and tumor node metastasis (TNM) stage (%). We also extracted information on therapeutic methods, including radiotherapy, chemotherapy, and targeted therapy. We collected laboratory variables, including white blood cells, lymphocytes (LYMs), albumin (Alb), globulin, AFP, lactate dehydrogenase, fibrinogen (FIB), FIB degradation products (FDP), β2-microglobulin (β2-MG), the neutrophil-to-LYM ratio (NLR), and the platelet-to-LYM ratio. The time span from the initial diagnosis to death or the last follow-up was used to define the OS.

Nomogram model construction and evaluation

Patients were split at random, with 671 (70%) assigned to the training set and the remaining 287 (30%) to the validation set. A 10-fold cross-validation in least absolute shrinkage and selection operator (LASSO) regression identified the optimal lambda (λ) for variable selection, using the λ minimum criterion to select candidate variables for further analysis. The optimal tuning parameter, λ minimum, was selected to minimize the mean cross-validation error. Variables with nonzero coefficients at the minimum λ were identified as potential predictors. Variable selection in the training set was conducted using LASSO regression and stepwise multivariate Cox regression, with clinical relevance considered. The selected variables were subjected to multivariate Cox proportional hazards regression analysis to identify independent prognostic factors for HCC, which were subsequently used to develop a nomogram prediction model.

This study confirmed the performance of the nomogram model using a validation set. Prediction accuracy was evaluated using Harrell’s concordance index (C-index), and generalizability was assessed using the area under the receiver operating characteristic curve (AUC). Predictive accuracy was validated in both sets using a calibration curve generated through the bootstrap method with 1000 resamplings. Receiver operating characteristic curves and C-indices were used to evaluate the dynamic predictive performance of the model. The 95%CI was determined using either the influence function method or bootstrap resampling to ensure robust uncertainty estimates.

To more clearly assess the benefits of the nomogram developed in this study compared with the traditional TNM staging system, we systematically compared these models using AUC, integrated discrimination improvement (IDI), and net reclassification improvement (NRI). This comparison provides a basis for evaluating the additional clinical value of our nomogram compared with that of existing staging methods.

Statistical analyses

In this study, normally distributed data are reported as mean ± SD, and intergroup comparisons were conducted using the t-test. For non-normally distributed data, the median and interquartile range were used to describe the central tendency and variability, respectively, and nonparametric tests were employed for group comparisons. Frequencies and percentages were used to describe categorical variables, and group differences were evaluated with the χ² test. Furthermore, the proportional hazards assumption was examined in the Cox regression model. Multiple imputation by chained equations (MICE) was used to address missing data, employing the “MICE” package. Patients with more than 20% missing variables were excluded from the analysis, while variables with ≤ 20% missing values were imputed using the MICE algorithm. Multiple imputation is based on the missing-at-random assumption. To preserve the structure of correlations between variables as much as possible and to reduce potential bias, all candidate predictors, survival times, and outcome events were included in the imputation model. A total of 10 imputed datasets (n = 10) were generated with 50 iterations to ensure convergence of the chained equations algorithm. The imputation process was conducted prior to model development. R software (version 4.2.1) was used for statistical analysis, and a two-sided P value < 0.05 was considered to indicate statistical significance.

RESULTS

Baseline patient characteristics

Among the 958 patients, 563 had died. The patients’ average age was 57.9 years; 731 were male (76.30%). On the other hand, most patients did not have diabetes, and their body mass index were within the normal range. The number of patients with stage IV disease was significantly greater than that of patients with other stages. In terms of treatment, most patients chose surgery, and fewer than half chose other treatment options. Table 1 displays the descriptive data for our population, showing no significant differences between the datasets. Table 1 presents the distribution of sociodemographic and clinical characteristics among HCC patients.

Table 1 Patient demographics and clinical characteristics, n (%)/mean ± SD/median (interquartile range).

Variables	Overall (n = 958)	Training set (n = 671)	Validation set (n = 287)	P value
Age	57.90 ± 11.45	58.14 ± 11.26	57.36 ± 11.87	0.333
Karnofsky performance status	82.87 ± 9.18	83.17 ± 8.66	82.16 ± 10.29	0.120
Sex				0.361
Male	731 (76.30)	506 (75.41)	225 (78.40)
Female	227 (23.70)	165 (24.59)	62 (21.60)
Hypertension				0.698
No	762 (79.54)	531 (79.14)	231 (80.49)
Yes	196 (20.46)	140 (20.86)	56 (19.51)
Diabetes				1.000
No	805 (84.03)	564 (84.05)	241 (83.97)
Yes	153 (15.97)	107 (15.95)	46 (16.03)
Body mass index				0.289
18.5-23.9	577 (60.23)	395 (58.87)	182 (63.41)
≥ 24	316 (32.99)	226 (33.68)	90 (31.36)
< 18.5	65 (6.78)	50 (7.45)	15 (5.23)
Tumor node metastasis				0.085
I-II	227 (23.70)	170 (25.34)	57 (19.86)
III	347 (36.22)	230 (34.28)	117 (40.77)
IV	384 (40.08)	271 (40.39)	113 (39.37)
Radiotherapy				0.340
No	877 (91.54)	610 (90.91)	267 (93.03)
Yes	81 (8.46)	61 (9.09)	20 (6.97)
Chemotherapy				0.375
No	609 (63.57)	420 (62.59)	189 (65.85)
Yes	349 (36.43)	251 (37.41)	98 (34.15)
Surgery				0.943
No	307 (32.05)	216 (32.19)	91 (31.71)
Yes	651 (67.95)	455 (67.81)	196 (68.29)
Immunotherapy				0.112
No	853 (89.04)	605 (90.16)	248 (86.41)
Yes	105 (10.96)	66 (9.84)	39 (13.59)
Targeted				0.347
No	650 (67.85)	462 (68.85)	188 (65.51)
Yes	308 (32.15)	209 (31.15)	99 (34.49)
White blood cell	6.61 ± 3.19	6.54 ± 3.14	6.77 ± 3.28	0.311
LYM	1.22 ± 0.50	1.20 ± 0.50	1.25 ± 0.52	0.142
Albumin	38.73 ± 6.13	38.96 ± 6.08	38.20 ± 6.22	0.078
Globulin	32.75 ± 6.75	32.66 ± 6.65	32.97 ± 6.99	0.519
Alpha-fetoprotein	11.05 (3.20, 334.60)	12.10 (3.24, 444.70)	10.40 (3.16, 168.86)	0.162
Lactic dehydrogenase	293.78 ± 315.56	294.11 ± 256.06	292.99 ± 423.80	0.960
FIB	3.33 ± 1.26	3.30 ± 1.23	3.41 ± 1.33	0.205
FIB degradation products	2.50 (2.50, 5.97)	2.50 (2.50, 5.80)	2.70 (2.50, 7.05)	0.441
β2-microglobulin	3.10 ± 1.42	3.13 ± 1.48	3.03 ± 1.27	0.338
Neutrophil-to-LYM ratio	3.50 (2.29, 5.30)	3.45 (2.30, 5.32)	3.58 (2.27, 5.20)	0.965
Platelet-to-LYM ratio	142.51 (97.36, 200.95)	143.75 (94.25, 198.86)	141.94 (102.54, 204.88)	0.664

FIB: Fibrinogen; LYM: Lymphocyte.

Independent prognostic factors identified in the training set

The analysis confirmed that all variables adhered to the proportional hazard assumption. The Cox proportional hazards model developed in this study met the essential assumptions and showed strong applicability and reliability (P > 0.05) (Figure 2). Independent prognostic factors were identified in the training set using Cox proportional hazards models (Figure 2C). Multivariate analysis revealed that KPS [hazard ratio (HR) = 0.987, 95%CI: 0.977-0.997)], sex (HR = 1.274, 95%CI: 1.014-1.602), diabetes (HR = 1.396, 95%CI: 1.050-1.855), and TNM stage (HR for stage III vs stage I-II = 2.146, 95%CI: 1.540-2.991) were independent predictors. The HR for stage IV disease compared with stage I-II disease was 3.016 (95%CI: 2.113-4.306). Both chemotherapy and surgery had an HR of 0.708 (95%CI: 0.551-0.909 for chemotherapy and 95%CI: 0.551-0.911 for surgery). FDP had an HR of 1.010 (95%CI: 1.000-1.019), β2-MG had an HR of 1.132 (95%CI: 1.050-1.221), Alb had an HR of 0.974 (95%CI: 0.956-0.993), and NLR had an HR of 1.042 (95%CI: 1.019-1.066).

Figure 2 Results of the prognostic nomogram for overall survival. A: Construction of a predictive model of hepatocellular carcinoma in the dataset by least absolute shrinkage and selection operator regression; B: Each curve in the figure represents the changed trajectory of each independent variable coefficient; C: The hazard ratio plot of stepwise multivariate Cox regression; D: Nomogram for predicting the 1-year, 3-year, and 5-year overall survival of hepatocellular carcinoma patients. Alb: Albumin; FDP: Fibrinogen degradation products; KPS: Karnofsky performance status; NLR: Neutrophil-to-lymphocyte ratio; TNM: Tumor node metastasis.

Development of the prognostic nomogram model

Using LASSO regression and stepwise multivariate Cox regression, 10 independent prognostic factors for HCC were identified (Figure 2A and B). A predictive nomogram for OS rates in individuals with HCC was established and is shown in Figure 2D. The cumulative scores were then mapped onto a comprehensive point table to determine the OS probability.

We applied SHapley Additive exPlanations (SHAP) analysis to explore the importance of each predictor and its effect on the model, as shown in Supplementary Figure 1. Supplementary Figure 1A presents a global overview of the model’s key factors by ranking predictors based on their mean absolute SHAP values. TNM stage exhibited the highest mean SHAP value, indicating that it was the most influential feature, followed by chemotherapy and surgery. These variables played key roles in shaping the model predictions. Supplementary Figure 1B shows the SHAP summary plot, which illustrates the distribution of the SHAP values across all features. The TNM stage exhibited the broadest SHAP value distribution, highlighting its significant influence on the prediction, particularly in stage IV, where it contributed positively to the SHAP values. NLR and β2-MG showed significant distributions, with their feature values highly correlated with SHAP values. In contrast, the FDP values were clustered around zero, confirming their relatively minor contribution to the prediction.

Performance and validation of the nomogram model

The C-index for the training set was 0.752 (95%CI: 0.731-0.774), whereas that for the validation set was 0.716 (95%CI: 0.680-0.752). Time-dependent C-index and AUC analyses confirmed the superiority and consistency of the predictive accuracy and generalizability of the nomogram model over a 1-5-year period (Figure 3A and B). In the training set, the AUCs for 1-year, 3-year, and 5-year OS were 0.847, 0.828, and 0.846, respectively, whereas in the validation set, they were 0.799, 0.799, and 0.746, respectively (Figure 3C and D). The nomogram calibration curves demonstrated a strong alignment between the predicted and actual survival rates, reflecting high predictive accuracy (Figure 4A and B). The nomogram model provided significant net benefits, as illustrated in Figure 4C and D. Decision curve analysis curves indicated that the nomogram offered a greater net clinical benefit than the traditional TNM staging system at all risk thresholds.

Figure 3 The generalization ability of the nomogram model. A: The time-dependent concordance index of the nomogram model; B: The time-dependent area under the receiver operating characteristic curve of the nomogram model; C: The 1-year, 3-year and 5-year receiver operating characteristic curves for the nomogram model in the training set; D: The 1-year, 3-year and 5-year receiver operating characteristic curves for nomogram model in the validation set. AUC: Area under the receiver operating characteristic curve.

Figure 4 The prediction accuracy and clinical application value of the nomogram model. A: The 1-year, 3-year, and 5-year calibration curves of the nomogram model in the training set; B: The 1-year, 3-year, and 5-year calibration curves of the nomogram model in the validation set; C: The 1-year, 3-year, and 5-year decision curve analysis curves for nomogram models in the training set; D: The 1-year, 3-year, and 5-year decision curve analysis curves for nomogram models in the validation set. OS: Overall survival; TNM: Tumor node metastasis.

In the training set, the TNM staging model achieved AUC values of 0.730, 0.744, and 0.771 for predicting 1-year, 3-year, and 5-year survival, respectively; the corresponding values in the validation set were 0.746, 0.715, and 0.725, respectively. The AUC values of the nomogram model were significantly greater than those of the TNM staging model in both sets (all P < 0.001), indicating its superior generalizability. Furthermore, we compared the NRI and IDI between the nomogram model and the TNM staging model. The findings indicated an IDI of 0.064 (95%CI: 0.032-0.103, P < 0.001) and an NRI of 0.270 (95%CI: 0.103-0.414, P = 0.004). Compared with the TNM staging system, the nomogram model showed notable improvements in predictive accuracy, offering improved discrimination and risk stratification.

This study was refitted using complete datasets for sensitivity analysis. The results revealed no substantial change in the direction or extent of the association of key prognostic variables after multiple imputation was used for missing data, and the model results were robust and not materially affected by the methods used to handle missing data.

Nomogram model for risk stratification

On the basis of the Youden index for the total score of the nomogram, the optimal cutoff value was 171.51 points. Patients with scores ≥ 171.51 were categorized as high risk, while those with scores < 171.51 were categorized as low risk. The nomogram model evaluates each patient by assigning a risk score and categorizing them into low-risk or high-risk groups. As illustrated in Figure 5, Kaplan-Meier analysis confirmed that the model effectively identified high-risk patients in both sets.

Figure 5 Kaplan-Meier curves of different risk levels to estimate the overall survival of patients in two groups. A: Training set; B: Validation set.

Web-based calculator for the nomogram model

Our web-based nomogram calculator for HCC prognosis is available at https://cpcfcqch.shinyapps.io/HCC_OS/. Users can quickly calculate the patient’s survival probability by selecting the relevant parameters. For example, if the following patient/tumor characteristics are selected: (1) With diabetes; (2) KPS of 65; (3) Male; (4) TNM stage III; (5) Receiving chemotherapy and surgery; (6) FDP = 9; (7) β2-MG = 4; (8) Alb = 36; and (9) NLR = 8, then the 3-year survival probability for this patient is estimated to be 0.309 (95%CI: 0.200-0.480).

DISCUSSION

HCC ranks among the leading causes of cancer-related deaths worldwide, with China bearing a considerable share of this burden. Despite advancements in treatment modalities, prognostic uncertainty remains a challenge in the management of HCC patients. The development of a reliable risk prediction model for OS is crucial for facilitating the application of personalized treatment strategies and improving patient outcomes[16,17].

Clinical data, including demographic information, tumor characteristics, treatment history, and laboratory results, were systematically collected. The developed risk prediction model demonstrated high accuracy and discrimination ability in estimating OS probabilities for patients with HCC. The key prognostic factors identified in the model included tumor stage, liver function parameters, Alb level, and β2-MG level. The model’s performance was further validated internally, confirming its robustness and reliability across different patient populations. The calibration curve in our study showed a strong agreement between the predicted and actual survival probabilities, highlighting the model’s excellent repeatability and reliability. The performance of the nomogram was excellent, with C-index values of 0.752 in the training cohort and 0.716 in the validation cohort. It demonstrated an acceptable discriminative ability. Furthermore, the decision curve analysis curves demonstrate that the clinical value of the nomogram is superior to that of the TNM staging system. These results indicate that the nomogram model has good predictive accuracy and generalizability, supporting its potential clinical applicability. Our findings indicate that the nomogram is a dependable and precise prognostic tool for predicting OS in HCC patients.

Several studies have reported nomograms for predicting OS in HCC patients. According to Lin et al[18] clinical and demographic factors play a crucial role in predicting OS in patients with HCC. The predictive model incorporated factors such as tumor stage, AFP levels, liver function tests, and patient demographics to stratify patients into risk categories. However, compared with previous research, our study has several strengths. LASSO regression, along with univariate and multivariate Cox regression analyses, revealed TNM stage, chemotherapy status, surgical status, β2-MG level, Alb level, and NLR as independent prognostic parameters. These conclusions are generally consistent with those of previous reports. Second, in addition to the recognized high-risk factors for HCC, such as TNM stage and diabetes, we identified β2-MG levels and NLR as additional high-risk factors.

β2-MG, an integral component of major histocompatibility complex I, plays a crucial role in antigen presentation and immune surveillance. Abnormal β2-MG expression may indicate dysregulation of immune recognition and is associated with tumor progression and poor prognosis[19]. Research has indicated that tumor cells may downregulate or abnormally secrete β2-MG, impairing antigen presentation and decreasing the ability of cytotoxic T LYMs to recognize tumor cells, thus facilitating immune evasion and tumor progression. Moreover, elevated circulating β2-MG levels may reflect increased tumor burden, accelerated cellular turnover, and activation of the inflammatory microenvironment, all of which are closely associated with enhanced invasiveness and poor survival in patients with HCC[20]. Dysregulation of β2-MG expression is associated with immune evasion mechanisms and tumor escape from immune surveillance in patients with HCC[21]. The NLR is a biomarker that reflects the balance between neutrophil-mediated and LYM-mediated inflammatory responses[22]. Neutrophils contribute to tumor angiogenesis, proliferation, and metastasis through the secretion of vascular endothelial growth factor, matrix metalloproteinases, and inflammatory cytokines, whereas LYMs, especially cytotoxic T cells, are pivotal in antitumor immunity[23,24]. Our study revealed a correlation between a low NLR and improved outcomes, including prolonged OS and lower recurrence rates after treatment. Patients with a high NLR may benefit from more aggressive interventions or combination therapies to improve their outcomes[25].

Diabetes plays a significant role in influencing the prognosis of patients with HCC[26], and its effective management is crucial for improving long-term outcomes. Research has shown that the risk of HCC in diabetes patients has increased by 2-3 times. Diabetes is associated with reduced OS and progression-free survival in patients with HCC[27,28]. Moreover, Shen et al’s studies[29] have shown that, compared with pure HCC, intrahepatic cholangiocarcinoma and combined hepatocellular-cholangiocarcinoma are often associated with poorer clinical prognosis, higher recurrence rates, and shorter OS[30,31]. Encouraging patients to actively receive treatment, including chemotherapy and surgery, can significantly increase their survival rate[32-34]. This was also reflected in our study, in which patients who received chemotherapy and surgery had significantly improved survival rates of 29% and 31%, respectively.

This study has the following limitations. As a single-center study, it lacks external data to sufficiently validate its generalizability. In the future, we plan to conduct a multicenter study based on the results of this research to further validate the model’s generalizability and clinical applicability. Secondly, owing to data availability, imaging and genetic data were not included as predictive factors. We aim to incorporate these factors in future research to enhance the predictive accuracy of the model. These findings underscore the need for further research to improve and validate nomogram models for predicting OS in HCC patients.

CONCLUSION

The nomogram developed in this study suggests a potential method for improving clinical outcomes and patient care. Future research may involve further refinement of the model, prospective validation in larger cohorts, and exploration of its utility in guiding therapeutic interventions. Ultimately, the implementation of such predictive tools could revolutionize the management of HCC and increase patient survival rates.

ACKNOWLEDGEMENTS

We thank all participants, as no meaningful research could have been conducted without them.