Impact of liver fibrosis score on prognosis after common therapies for intrahepatic cholangiocarcinoma: a propensity score matching analysis

The database used in our study was the incidence–SEER 18 Regs Custom data (with additional treatment fields), Nov2018 Sub (1975–2016 varying). A total of 729 ICC patients were identified at clinical stages I–IV from a population of 141,625 candidates diagnosed with primary liver cancer between 2004 and 2015. Of the identified patients, 63.8% of them (n = 465) were assigned to the low fibrosis score (low-Fb score) group (Fb score of 0–4), and the remaining ones (n = 264, 36.2%) were assigned to the high fibrosis score (high-Fb score) group (Fb score of 5–6). The flowchart of this study is presented in Fig. 1, and details are provided in Supplementary file 1.

Variables from the patient, tumor, and treatment levels were adopted in the statistical analysis. The patient-level parameters included age at diagnosis, sex, race, and marital status. The tumor-level variables included tumor number, tumor size, year of diagnosis (2004–2009 vs. 2010–2015), alpha-fetoprotein (AFP, cutoff value of 15 ng/mL), lymph node status, metastasis of diagnosis, tumor pathological grade, tumor node metastasis (TNM) stage, and Ishak Fb score. The treatment-level variables mainly included the surgery, radiation, and chemotherapy records of ICC patients. A full description of the variables is available in the SEER Data Management System User Manual (https://​seer.​cancer.​gov/​tools/​codingmanuals/​index.​html) and CS Coding Instructions v0205.

Continuous variables were demonstrated as means with standard deviations (SD) and were compared using the student’s t-test or Mann–Whitney U test. Categorical variables, presented as frequencies and percentages (%), were compared using the chi-square test or Fisher’s exact test.

To balance the differences between the two groups, we constructed a propensity score model by using the variables of the entire logistic regression model for patients with high-Fb scores. The candidate variables included all variables significantly associated with high-Fb scores, as determined via a univariate analysis with a threshold of P < 0.2. Then, propensity score matching (PSM) and inverse probability of weighting (IPW) methods were adopted to reduce the standard mean differences (SMDs) of the covariates. In PSM, individuals with low-Fb scores were matched to those with high-Fb scores by using a matching ratio of approximately 1:1, with the closest estimated propensity score values limited within 0.1 of the SD in the study population [12]. All patients with high- and low-Fb scores were weighted by the propensity score model via the IPW method [13, 14]. Two matched cohorts (PSM cohort and IPW cohort) were generated to validate the associations of the Fb scores with the clinical outcomes of ICC patients. The Kaplan–Meier (KM) curves for the OS and compete risk survival analysis were derived to visualize the comparison between the high- and low-Fb scores. Cox regression analysis was performed to explore the risk factors influencing patient survival. Additionally, the treatment benefits for the ICC patients with high- and low-Fb scores were further explored and confirmed in the matched cohorts via stratified analyses based on treatment modalities. All the statistical analyses were performed in R (R Foundation for Statistical Computing, Vienna, Austria, Version 3.6.0), and the R packages of “forestplot,” “glm,” “ggolot2,” “matching,” “survey,” and “cmprsk” were used. Two-tailed P < 0.05 was considered to indicate statistical significance.

According to the inclusion and exclusion criteria, 729 eligible ICC patients from the SEER database (2004–2015) were analyzed (Fig. 1). The baseline clinicopathological characteristics of the ICC patients in the high- and low-Fb score groups are listed in Table 1. As shown in Fig. 2, despite the rising incidence, the proportion of patients with high-Fb scores did not significantly change over time.

The median survival times in the low- and high-Fb score groups were 22.43 and 16.67 months, respectively. The OS analysis (Fig. 3a) showed that a high-Fb score was a significant risk factor of OS in ICC patients (HR 95%CI = 1.371 [1.154–1.628], P < 0.001). Before matching, the one-, three-, and five-year OS rates were 50.9, 28.0, and 16.1% in the low-Fb score group and 39.3, 20.1, and 8.0% in the high-Fb score group (P < 0.001), respectively. Table 2 shows the results of the univariate and multivariate Cox regression analyses in the unmatched cohort. We found that a high-Fb score was an independent risk factor of OS in ICC patients (HR 95%CI = 1.282 [1.067–1.541], P = 0.008). In addition to the Fb score, we identified being male, presence of multiple tumors, tumor size > 3 cm, presence of distance metastasis, and advanced stage as risk factors of OS. Meanwhile, surgery treatment and chemotherapy were identified as protective factors of OS in ICC patients. In the competing risk survival analysis, the unadjusted P values of the Fb score for the death circumstance of ICC and other reasons were 0.005 and 0.369 in the primary cohort (Fig. 4a).

A propensity score model was constructed on the basis of the multivariate logistic regression analysis (Table 3). Compared with the patients in the low-Fb score group, those in the high-Fb score group were mostly single and male and were more likely to have higher AFP levels, small tumor size (≤3 cm), and multiple tumors. The 6th AJCC stage III was more inclined to occur in patients with high-Fb scores than in their counterparts.

According to the propensity score model (Table 3), PSM, excluding the prognosis indexes (vital status and survival time), achieved an adequate balance between the high- and low-Fb score groups. Notably, we found a decrease in SMDs between the two groups in the IPW analysis but not in PSM (Supplemental Table 1 in Supplemental file 1). After PSM and IPW analyses, the variables between the two groups were significantly balanced because all the available P values were > 0.05.

After matching, the KM curves (Fig. 3b and c) showed that a high-Fb score was a significant risk factor of OS in the matched cohort (HR 95%CI = 1.323 [1.075–1.628], P = 0.008). IPW analysis revealed similar results (HR 95%CI = 1.256 [1.031–1.531], P = 0.024). After PSM, the one-, three-, and five-year OS rates were 45.0, 26.0, and 10.2% in the low-Fb score group and 36.0, 8.1, and 2.3% in the high-Fb score group (P = 0.008), respectively. In the competing risk survival analysis, the propensity score-adjusted P values were 0.032 and 0.516 in the matching cohort after PS matching, (Fig. 4b).

Figure 5 shows the comparison results of the OS rates of patients who received surgery with different Fb scores after PSM. Compared with the patients in the low-Fb score group, those in the high-Fb score group had poorer OS rates, with the P values of the log-rank test between the two groups being 0.015 (Fig. 5a). In the comparison of the OS rates of patients without surgery records, different results were observed between the two groups (P = 0.466) (Fig. 5b). The subgroup survival analyses outcomes between the two groups of patients with and without the other two therapies are illustrated in Supplemental Figure 1. A subgroup analysis of the prognosis of ICC patients with different therapies was conducted. Supplemental Figure 2 shows that in the surgery group, the patients’ survival outcomes were not influenced by the treatment they received (i.e., chemotherapy and radiotherapy treatments) (P = 0.96 and 0.79). By contrast, chemotherapy and radiotherapy were significantly related to the prognosis of ICC patients in the none-surgery group (P < 0.001).