Age-related nomogram revealed optimal therapeutic option for older patients with primary liver cancer: less is more

Data source and patient selection

The data were obtained from National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) program between 2010 and 2016. SEER serves as an open-access resource providing demographic, clinicopathological, and some treatment information related to tumors. All the data used in this study were retrieved from the SEER*Stat Version 8.4.0 (http://www.seer.cancer.gov/seerstat). As the study utilized publicly available data, local ethical approval and declaration were not required.

The inclusion criteria were as follows: 1) the patients were diagnosed between 2010 and 2016; 2) primary tumor were located in the liver and intrahepatic bile duct; 3) there was only one primary tumor and the behavior was malignant. The exclusion criteria were: 1) incomplete patient demographics, including race and marital status; 2) incomplete clinicopathological characteristics, such as grade, AJCC TNM stage, tumor size, AFP level, metastasis status, survival information, and treatment details.

Clinical variables of patients

The information of demographics included age at diagnosis, sex, marital status (categorized as married and unmarried, including separated, divorced, widowed, single or domestic partner), race (white, black, and others, including American Indian/Alaska Native, Asian/Pacific Islander). The tumor-related factors included tumor size (<=5, 5-10, >10 cm), grade, AFP level, histology, AJCC TNM stage, and metastatic status involving bone, brain, and lung. The therapeutic options were categorized as follows based on the codes from the SEER program: no treatment (N), surgery alone (S), chemotherapy alone (C), radiation alone (R), surgery combined with chemotherapy (S+C), surgery combined with radiation (S+R), chemotherapy combined with radiation (C+R), and surgery combined with chemotherapy and radiation (S+C+R). The primary endpoint of interest was overall survival (OS), defined as the time from the date of the first-time diagnosis until the date of death caused by any cause or the most recent follow-up.

Construction of nomogram

We divided all patients into three groups according to age at diagnosis using the X-tile software. The X-tile software is a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization, obtained from Yale University School of Medicine, USA. The X-tile offers a single, global assessment of each possible way of dividing the population into low, middle and high level of marker expression [10]. For the construction of nomogram, the patients who were diagnosed between 2010 and 2014 from the SEER program were randomly divided into a training set and an internal validation set at a ratio of 7:3. The patients who were diagnosed between 2015 and 2016 from the SEER program were used as an external validation set. The training set was used to construct the nomogram, and its efficacy was assessed through internal and external validation sets.

In the training cohort, the independent risk factors were identified by univariate and multivariate Cox proportional-hazards regression analyses for OS [11]. A backward stepwise method, based on the smallest Akaike information criterion (AIC) value, was finally employed to incorporate the covariates into the multivariate Cox proportional hazards models. The AIC value suggests the minimal loss of prognostic information [12, 13]. Cox regression and the nomogram were developed by using the ‘rms’ package in R software. The predictive discrimination ability of the nomogram was assessed by using C-index and the area under curve (AUC) [14, 15]. Calibration curves-based boots method was utilized to examine the association between actual OS and predicted OS by the nomogram in the training and validation set [16]. The decision curve analysis (DCA) was also developed to estimate the clinical utility and benefits of the prediction model [17, 18]. We divided patients into three risk groups according to the total score of each patient and the Kaplan-Meier survival curve with a log-rank test was used to verify the value of the risk classification [19]. Additionally, an interactive dynamic nomogram web page was built to visualize our results using ‘DynNom’ package [20].

Statistical analysis

Statistical analysis was performed by using R software version 4.1.2. The difference among groups was compared using the Pearson’s chi-square test or Fisher’s exact test for categorical variables appropriately. Survival curve was compared using the log-rank test. P < 0.05 was considered statistically significant.

Data availability

The data of this study are available in the SEER database (https://seer.cancer.gov/).

Study cohort characteristics

We ultimately included a total of 7,244 eligible PLCs from the SEER database based on the above inclusion and exclusion criteria. The flowchart of the patient selection process is shown in Figure 1A. Generally, the majority of patients were male (73.9%), white (67.6%), married (57.8%), AFP positive (64%), T1 (42.7%), N0 (89.5%), M0 (86.5%) stage and the common histology is HCC (91%). 46% of tumors was moderately differentiated/Grade II. The lung represented the most common site of metastasis, accounting for 5% of cases (Table 1).

We further stratified the patients who diagnosed between 2010 and 2014 (n=5081) into three groups namely low-age group (n=2761): <64 years old, middle-age group (n=1475): 64-74 years old, old-age group (n=845): >74 years old, according to the best cut-off value based on the X-tile (Figure 1B). The Kaplan-Meier survival curve for OS showed significant difference in different age groups (P < 0.001) (Figure 1C). There was no doubt that the younger patients have better OS and the median OS of three groups was 20,16 and 9 months, respectively (Figure 1C and Table 1). Also, there was significant difference among three groups in terms of sex, marital status, race, tumor size, T stage, N stage, AFP level and therapeutic options in Table 1. In the low-age group, the most common tumor size was less than 5 cm, while it was 5-10 cm in old-age group. In the low and middle-age group, surgery alone (S) was the most treatment option, followed by chemotherapy alone (C), but in the old-age group, 38% of patients did not receive any treatment (N). Another cohort comprising 2,163 patients diagnosed between 2015 and 2016 was selected to as an external validation set. The demographics and clinicopathological characteristics were summarized in Supplementary Table 1.

Low-age group

The total 2761 patients in this group were divided randomly into the training set (n=1920) and the internal validation set (n=841). The baseline characteristics of the group is shown in Supplementary Table 2. The 1-, 3-, and 5-year OS rates were 60.9%, 39.3%, and 31.0% in the initial set. There was no significant difference between distribution of these variables in the training and internal validation sets. The 1,013 patients who diagnosed between 2015 and 2016 were selected as an external validation set.

As illustrated in Table 2, the univariate Cox analysis conducted in the training set showed that the significant indicators were sex, race, marital status, histology, grade, tumor size, T stage, N stage, M stage, AFP level, treatment and bone, brain, lung metastasis. The model yielded the smallest AIC value (AIC=17734.15) when including the aforementioned 13 independent indicators into the multivariate Cox analysis, except brain metastasis. Based on the model, we constructed a satisfactory nomogram for the prediction of 1-, 3- and 5-year OS probability of PLCs in low-age group (Figure 2A). Each variable has a specific value on the points scale, and the total Nomo-score was calculated by summing these scores to predict the 1-, 3- and 5-year survival probability for individual patient.

The AUC values of the nomogram to predict 1-, 3- and 5-year OS were 0.876 (95% CI, 0.860-0.891), 0.886 (95% CI, 0.871-0.901) and 0.883 (95% CI, 0.868–0.899) in the training set, respectively (Figure 3A). In the internal validation set, the AUC values to predict 1-, 3- and 5-year OS were 0.853 (95% CI, 0.826-0.879), 0.876 (95% CI, 0853-0.899) and 0.872 (95% CI, 0847-0.896), respectively (Figure 3B). In the external validation set, the AUC values to predict 1-, 3- and 5-year OS were 0.862 (95% CI, 0.839-0.885), 0.872 (95% CI, 0.849-0.895) and 0.850 (95% CI, 0.760-0.939), respectively (Figure 3C). The C-index of nomogram were 0.802, 0.786 and 0.797 in the training, internal validation and external validation sets.

Middle-age group

In this group, 1475 patients were randomly divided into the training set (n=1028) and internal validation set (n=447). The information was summarized in Supplementary Table 3. Similarly, there was no significant difference between distribution of these variables in the training and internal validation sets. The 800 patients who diagnosed between 2015 and 2016 were selected as an external validation set. In Table 3, the univariate Cox analysis in the training set showed the significant indicators. The model yielded the smallest AIC value (AIC=9526.1) in the multivariate Cox analysis with 10 variables: race, primary site, grade, tumor size, T stage, N stage, M stage, AFP level, treatment and lung metastasis. We constructed a nomogram to predict the 1-, 3- and 5-year OS probability of PLCs in middle-age group (Figure 2B).

The C-index were 0.766, 0.765 and 0.801 in the training, internal validation and external validation set. The AUC values to predict 1-, 3- and 5-year OS was 0.843 (95% CI, 0.818-0.867), 0.849 (95% CI, 0.825-0.873) and 0.864 (95% CI, 0.840–0.888) in the training set, respectively (Figure 3D). In the internal validation set, the AUC values were 0.855 (95% CI, 0.820-0.891), 0.857 (95% CI, 0823-0.892) and 0.866 (95% CI, 0.831-0.901), respectively (Figure 3E). In the external validation set, the AUC values of the nomogram to predict 1-, 3- and 5-year OS were 0.869 (95% CI, 0.843-0.894), 0.864 (95% CI, 0.838-0.890) and 0.814 (95% CI, 0.678-0.950), respectively (Figure 3F).

Old-age group

The 845 patients were randomly divided into the training set (n=587) and internal validation set (n=258). The demographic and clinicopathological characteristics were summarized in Supplementary Table 4. The 350 patients who diagnosed between 2015 and 2016 were selected as an external validation set. Based on the OS, the univariate Cox analysis in the training set showed that the significant indicators were sex, race, grade, T stage, N stage, M stage, AFP level, treatment and bone, brain, lung metastasis in Table 4. The model yielded the smallest AIC value (AIC=5612.22) when 9 variables, sex, race, grade, tumor size, T stage, N stage, M stage, AFP level, and treatment, were included into the multivariate Cox analysis. Then we constructed a reliable nomogram for the prediction of 1-, 3- and 5-year OS of PLCs in old-age group (Figure 2C).

The C-index were 0.781, 0.762 and 0.748 in the training, internal and external validation set. Respectively, the AUC values of the nomogram to predict 1-, 3- and 5-year OS were 0.856 (95% CI, 0.826-0.886), 0.879 (95% CI, 0.849-0.909) and 0.868 (95% CI, 0.830–0.905) in the training set (Figure 3G); 0.871 (95% CI, 0.829-0.914), 0.864 (95% CI, 0.817-0.910) and 0.878 (95% CI, 0.829-0.926) in the internal validation set (Figure 3H); and 0.800 (95% CI, 0.754-0.846), 0.817 (95% CI, 0.765-0.870) and 0.868 (95% CI, 0.776-0.960) in the external validation set (Figure 3I).

Furthermore, the calibration curves indicated an optimal agreement between the actual and predicted probability of OS in the training set (Figure 4A, 4D, 4G), internal set (Figure 4B, 4E, 4H) and external validation set (Figure 4C, 4F, 4I) in three age groups. This suggested our model was reliable and valid. We also observed that the AUC values of our nomogram model were superior than the independent factors such as TNM stage and grade, regardless of the age group (Figure 5). Additionally, the DCA curves demonstrated that the nomogram exhibited better clinical applicability and net benefits at different threshold probabilities compared to both the treat-all and treat-none schemes for guiding clinical intervention, irrespective of the age group. It also outperformed TNM stage and other independent factors. Meanwhile, patients in each age group can be categorized into three risk subgroups based on the total score (Figure 6). The Kaplan-Meier survival curve for OS showed significant difference in three risk subgroups in each age group (Figure 2D–2F). We also developed an accessible online dynamic page to simplify and visualize our model (http://124.222.247.135/).

The effect of treatment in three age groups

Overall, in the low-age group, the patients who received surgery alone (S) or surgery combined with chemotherapy (S+C) had the best prognosis, followed by surgery combined with radiation (S+R) and surgery combined with chemotherapy and radiation (S+C+R) (Figure 7A). Further subgroup analysis revealed that in the female (Figure 7B), ICC (Figure 7C), AFP negative subgroup (Figure 7E), the prognosis of patients who treated with surgery alone (S) was better than that of patients treated with surgery combined with chemotherapy (S+C). While, in the male, HCC, AFP positive subgroup, the prognosis of patients treated with surgery combined with chemotherapy (S+C) was better.

In the middle-age group, the prognosis of patients who received S was the best, followed by S+C and S+R (Figure 7A). Further subgroup analysis suggested in the female, ICC, poorly differentiated (Grade III/IV), T3, tumor size 5-10cm subgroup, the prognosis of patients who treated with S was better. Meanwhile, in the male, HCC, well differentiated (Grade I/II), T1/2, N0, M0 subgroup, the prognosis who treated with surgery alone and surgery combined with chemotherapy or radiation were similarly good (Figure 7).

The patients who treated with S had a better prognosis than other treatments in high-age group. After subgroup analysis, in the male, well differentiated (Grade I/II), tumor size <5cm subgroup, the patients who treated with S+C also had good prognosis. Meanwhile, the patients in ICC, N1, M1 subgroup had the poor prognosis regardless of the type of treatment received (Figure 7).