A decision tree model to predict liver cirrhosis in hepatocellular carcinoma patients: a retrospective study

Introduction

Primary liver cancer is one of the most prevalent forms of cancer in the world (Qiu et al., 2022; Sauzeau et al., 2022; Xu et al., 2022). It is the fourth most common malignant tumor and the second leading cause of cancer death in China. Hepatocellular carcinoma (HCC) accounts for 75–85% of primary liver cancer (Craig et al., 2020; Sung et al., 2021). Liver cirrhosis caused by excessive drinking, obesity, hepatitis B virus (HBV) infection, or hepatitis C virus (HCV) infection is an essential contributing factor towards the development of HCC (Ginès et al., 2021).

Studies have found that about 1–8% of cirrhotic patients develop HCC yearly (Ioannou et al., 2007). HCC incidence also correlates with increased portal pressure or liver stiffness values, which are measured by transient elastography (TE) (Masuzaki et al., 2009; Ripoll et al., 2009). Radical resection is a vital method for the long-term survival of HCC patients, and the liver reserve function should be carefully evaluated before surgery. Future liver remnant (FLR) of more than 40% is necessary for surgery in cirrhotic patients. Due to the stealthy onset of HCC, less than 30% of patients are suitable for surgical treatment at initial diagnosis (Zhou et al., 2020). Local therapy, immunotherapy, and targeted therapy are the main treatments for patients with advanced HCC (Wang et al., 2022b). Studies have shown that the degree of liver cirrhosis and liver function status affect the long-term efficacy of systemic anti-tumor therapy and transcatheter arterial chemoembolization (TACE) (Lencioni et al., 2016; Roth et al., 2023). Therefore, evaluating the level of liver cirrhosis in HCC patients is of great significance.

Liver biopsy (LB) remains the gold standard for cirrhosis assessment. However, the invasive method is expensive and risks serious complications such as bleeding and infection. Sampling error in LB can also lead to an unreliable diagnosis (Okada et al., 2007; Patel & Sebastiani, 2020). TE is a non-invasive technique used to evaluate liver cirrhosis, but its results are associated with many factors, such as elevated aminotransferase or bilirubin levels, ascites, and extrahepatic cholestasis (Millonig et al., 2008; Yan et al., 2013). This limits the clinical application of TE in HCC patients. Furthermore, non-invasive tools, including the Child-Pugh (C-P) scoring system, aspartate aminotransferase to platelet ratio index (APRI), and fibrosis-4 (FIB-4), are less accurate in predicting cirrhosis in HCC patients (Roth et al., 2023; Xiao et al., 2016).

Therefore, we conducted this study intending to establish a new non-invasive model to improve the diagnostic accuracy of cirrhosis in HCC patients.

Materials and Methods

Results

Development and validation of the decision tree model

Since PLT is included in the formulas of APRI, FIB-4, and GPR, we investigated the effect of sex on the diagnostic accuracy of existing models for cirrhosis. The results showed that sex had little effect on diagnostic accuracy (Table 4). In addition, neither imaging nor existing models had an area under the receiver operating characteristic (AUROC) greater than 0.67 in the entire cohort. That is why we intended to establish a novel model to improve the diagnostic accuracy of liver cirrhosis in HCC patients. According to the algorithm, sex was not included in the decision tree. For instance, patients with CT/MRI-reported cirrhosis and FIB-4 ≥ 6.0 are predicted to have cirrhosis, while patients without CT/MRI-reported cirrhosis and APRI < 0.38 are expected not to have cirrhosis (Fig. 1). The decision tree model had excellent diagnostic performance in the training and validation cohorts, with AUROC values of 0.853 (95% CI [0.783–0.923]) and 0.817 (95% CI [0.733–0.902]), respectively (Figs. 2A and 2B). Both calibration curves showed good calibration of the decision tree, indicating that the cirrhosis predicted by the model was consistent with the actual cirrhosis occurrence. The p-values of the Hosmer-Lemeshow test were 0.678 and 0.462, respectively (Figs. 2C and 2D). DCA for the decision tree model, imaging-reported cirrhosis, APRI, FIB-4, and GPR were presented in Fig. 3, which indicated that the decision tree model provided a larger net benefit than the other four models.

Table 4:

The AUROC and cutoff values of diagnostic tools for liver cirrhosis.

		Entire cohort
Sex		Male	Female	All
Imaging-reported cirrhosis		0.649 (0.564, 0.734)	0.665 (0.507, 0.823)	0.658 (0.584, 0.731)
APRI	AUROC	0.650 (0.569, 0.731)	0.645 (0.484, 0.805)	0.653 (0.582, 0.724)
	Cutoff	0.47	0.51	0.47
FIB-4	AUROC	0.573 (0.485, 0.661)	0.669 (0.513, 0.826)	0.603 (0.527, 0.678)
	Cutoff	2.39	2.55	2.39
GPR	AUROC	0.585 (0.501, 0.670)	0.577 (0.409, 0.745)	0.577 (0.502, 0.652)
	Cutoff	0.47	0.23	0.43

DOI: 10.7717/peerj.15950/table-4

Note:

AUROC, area under the receiver operating characteristic; APRI, aspartate transaminase to platelet ratio index; FIB-4, fibrosis-4; GPR, γ-glutamyl transferase to platelet ratio.

Discussion

Cirrhosis, a late stage of chronic liver inflammation where regenerative cirrhotic nodules replace normal liver structure, can eventually lead to liver failure. About one million people die of liver cirrhosis worldwide annually, and one-third of cirrhotic patients will develop HCC (Ginès et al., 2021). Assessment of cirrhosis in patients with chronic liver disease (CLD) or HCC should be the cornerstone of treatment decisions and prognosis (Patel & Sebastiani, 2020; Wang et al., 2022a).

Radical resection is the best treatment option for HCC patients with Barcelona Clinical Liver Cancer (BCLC) stage 0 or A (Orimo et al., 2022). The criteria for assessing the safety of resection are as follows: FLR > 20–30% and indocyanine green (ICG)-R15 < 10% (patients with C-P class A and without liver cirrhosis); FLR > 40% (patients with liver cirrhosis); FLR > 50% and IGG-R15 of 10–20% (patients with liver cirrhosis) (Zhao & Cai, 2021). Preoperative evaluation of the liver cirrhosis status of patients is an important basis for determining the scope of surgical resection. For patients within the Milan criteria, liver transplantation (LT) is recommended as the first-line option for those unsuitable for liver resection (Marrero et al., 2018). Liver cirrhosis is a risk factor for postoperative recurrence of HCC (Marasco et al., 2019), and cirrhotic patients have a higher risk of developing HCC. Therefore, HCC patients with severe cirrhosis are ideal candidates for LT, accounting for 30–35% of European waiting list patients. Local and systemic treatments are options for HCC patients with BCLC stage B or C. The GIDEON study demonstrated that median overall survival (OS) was longer in C-P class A patients than in C-P class B patients (13.6 months vs 5.2 months) (Marrero et al., 2016). In addition, the prospective study by Pressiani et al. (2013) also showed that the OS and time to progression of C-P class A patients treated with sorafenib were longer than that of C-P class B patients. The C-P scoring system is the most widely used tool for assessing cirrhosis (Roth et al., 2023). Hence, assessing liver cirrhosis status in advanced HCC patients can predict treatment efficacy and prognosis. Compensated cirrhosis is asymptomatic, but it can progress to decompensated cirrhosis. Decompensated cirrhosis is associated with various complications (ascites, hepatic encephalopathy, variceal bleeding), leading to poor quality of life and frequent hospitalizations (Ginès et al., 2021). Thus, early identification of cirrhotic HCC patients for symptomatic and supportive treatment is imperative for mitigating the causative factors and preventing the progression of cirrhosis.

APRI, FIB-4, and GPR include five variables: age, PTL, ATL, AST, and GGT. Age is associated with cirrhosis (Sterling et al., 2006), and the incidence of HCC increases with age in all populations. Thrombocytopenia occurs in about 78% of patients with liver cirrhosis, and its degree of severity is an early prognostic marker. Decreased thrombopoietin (TPO) levels, hypersplenism secondary to portal hypertension, and myelosuppression induced by antiviral therapy may all lead to decreased PLT (Peck-Radosavljevic, 2017). Hepatocellular injury caused by chronic hepatitis can lead to elevated transaminases (especially AST) and GGT (Xiao et al., 2016). To summarize, the variables included in serum diagnostic models correlate with liver fibrosis. The male-to-female ratio of HCC patients is about 2–2.5:1. However, this study found that being female is an independent risk factor for cirrhosis. This may be related to the sample size of the training cohort, which needs to be confirmed by further research.

A previous systematic review and meta-analysis showed that APRI and FIB-4 could identify HBV-related cirrhosis with moderate accuracy and sensitivity (Xiao, Yang & Yan, 2015). Meanwhile, Xiao et al. (2016) found that the AUC values for APRI and FIB-4 were 0.676 and 0.652 for diagnosing cirrhosis in HBV-related HCC patients. However, Kim et al. (2016) proposed that APRI and FIB-4 are unsuitable for evaluating the liver fibrosis of chronic hepatitis B (CHB) patients. This may be because the APRI and FIB-4 models were established in patients with HCV or HCV/HIV co-infection. In summary, the cause of liver cirrhosis and HCC is complicated, and the APRI and FIB-4 models lack high diagnostic accuracy in HCC patients. GPR was established by HBV patients in sub-Saharan African countries, where it was more accurate than APRI or FIB-4 (Lemoine et al., 2016). The study by Lee et al. (2018) reported that GPR was a valuable marker for assessing the degree of liver fibrosis in CHB patients (AUROC: 0.84; cut-off value: 0.299).

We focused on a noninvasive diagnostic model of cirrhosis status in HCC patients, and resected specimens from all patients were analyzed for liver fibrosis. APRI, FIB-4, and GPR are models established in CLD, and LB is the reference standard. LB specimens account for only 1/50,000 of the liver volume and may not represent the entire liver parenchyma (Bravo, Sheth & Chopra, 2001). In addition, LB specimen length (25 mm is considered optimal) and diameter (should be greater than 1 mm) also affect diagnostic accuracy (Patel & Sebastiani, 2020). Therefore, using surgical tissues for pathological examination is the advantage of our research. Like existing models, our decision tree model was also developed based on the binary classification of pathological examinations. This may not reflect the dynamic changes of liver fibrosis (Patel & Sebastiani, 2020). Although TE can identify liver stiffness in real-time, the Baveno IV consensus suggested that two tests on different days are required to confirm the validity (de Franchis, 2015), which will significantly increase the economic burden of HCC patients. Besides, the operator’s inexperience is the main factor in the failure of TE detection (Castéra et al., 2010). Neither TE nor LB can be used as a routine detection method in areas with limited resources. Abnormal liver function, ascites, narrow intercostal space, obesity, and age over 52 years all affect TE detection (Castéra et al., 2010; Millonig et al., 2008; Yan et al., 2013).

Our research also has some shortcomings. First, selection bias brought about by retrospective research cannot be avoided. Second, there is a lack of external validation cohorts and prospective studies to further assess the model’s accuracy. Third, the prevalence situation of HCC in China led us to include mainly HBV-associated patients. Finally, we did not build predictive models for significant and advanced liver fibrosis. This is because radiologists can only evaluate liver cirrhosis, and the diagnosis of cirrhosis is more clinically significant for HCC patients (Wang et al., 2020).

Conclusions

In conclusion, the novel user-friendly decision tree model demonstrated excellent discrimination and calibration in predicting cirrhosis in HCC patients (Fig. 4). Since it is vital to know the cirrhosis status of HCC patients, the model can help clinicians in decision-making.

Supplemental Information