In this retrospective study, data from 632 patients diagnosed with HBV-related HCC and treated with a combination of transarterial chemoembolization (TACE) and ablation treatment at Beijing You’an Hospital, affiliated with Capital Medical University, were reviewed, from January 2014 to January 2020. All participants enrolled presented with a confirmed diagnosis of HCC according to the criteria stipulated by the American Association for the Study of Liver Disease (AASLD). The combination treatment was undertaken by physicians possessing a minimum of 5 years of expertise in the field. TACE was initially employed to target the cancerous lesions, delivering chemotherapy drugs and embolic agents directly to the tumor-feeding blood vessels. Following the TACE procedure, local ablation therapy was performed within a time frame of 1 to 2 weeks.

Here are the inclusion and exclusion criteria for the study. Inclusion criteria: (1) Patients diagnosed with hepatitis B virus (HBV)-related primary HCC; (2) Received the combination of transarterial chemoembolization (TACE) and ablation treatment; (3) Complete clinical and follow-up data. Exclusion criteria: (1) Non-primary HCC; (2) Not HBV-related HCC; (3) Received other anti-tumor therapies before the combination treatment; (4) Suffering from other malignancies or systemic diseases; (5) Incomplete clinical or follow-up data. Figure 1 illustrates the flowchart outlining the process of selecting qualified patients.

The study has obtained approval from the ethics committee of Beijing You’an Hospital, affiliated with Capital Medical University. The committee conducted a thorough review of the research protocol to ensure that the study adheres to ethical guidelines and regulatory requirements. Given the retrospective nature of the study, where data was collected from past medical records, and the stringent measures implemented to protect patient’s privacy, the need for obtaining informed consent from the subjects was deemed unnecessary.

The data collected in this study comprises general information, laboratory examination data, and pathological data. General information: age, gender, smoking history, drinking history, family history, antiviral history, hypertension, diabetes and so on. Laboratory examination data: white blood cell (WBC), red blood cell (RBC), lymphocyte (Lym), neutrophil (Neu), monocyte (Mon), eosinophils, basophils, platelet (PLT), hemoglobin (Hb), alanine aminotransferase (ALT), aspartate aminotransferase (AST), AST/ALT, gamma glutamyl transpeptidase (GGT), alkaline phosphatase (ALP), total protein (TP), albumin, globulin, prealbumin, total bilirubin (TBIL), direct bilirubin (DBIL), bile acid, uric acid, BUN, cholesterol, glucose, potassium, natrium, chlorine, creatinine, fibrinogen (Fib), thrombin time (TT), international normalized ratio (INR), activated partial thromboplastin time (APTT), activated partial thromboplastin time ratio (APTTR), prothrombin time (PT), prothrombin time activity (PTA), prothrombin time ratio (PTR) and so on. Pathological data: tumor size, tumor number, Barcelona Clinic Liver Cancer (BCLC) stage and Child-Pugh class.

Within one month after the combination treatment, all patients were required to undergo a follow-up check at the outpatient department of our hospital. They were followed every 3 months starting from the 2nd month to 1 year and every six months thereafter. Tumor recurrence refers to the discovery of new lesions in the liver through imaging examinations such as contrast-enhanced computed tomography (CT) and magnetic resonance imaging (MRI) or confirmed by pathological evidence based on tumor biopsy. Recurrence-free survival (RFS) was calculated as the time interval from the date of undergoing treatment to the date of tumor recurrence or the last follow-up date, and was presented in months. The last follow-up date for this study was December 1, 2023 and the median follow-up time was 23.2 months.

Categorical variables were expressed as numbers and proportions, and their comparison was executed through the Chi-square test. Continuous variables were represented as mean values accompanied by standard deviations, and their comparison was carried out using Student’s t-test or Mann-Whitney U test. The significance level was P=0.05, where P<0.05 indicated statistical difference. Platelet-albumin-bilirubin (PALBI) score was calculated as: 2.02 * log₁₀ bilirubin - 0.37 * (log₁₀ bilirubin)² - 0.04 * albumin - 3.48 * log₁₀ platelets + 1.01 * (log₁₀ platelets)², bilirubin was expressed in μmol/L, albumin in g/L and blood platelet count in 1000/μL. PALBI was divided into three grades, with higher grades indicating poorer liver function. PALBI grade 1: PALBI score≤ -2.53; PALBI grade 2: -2.53 < PALBI score ≤ -2.09; PALBI grade 3: PALBI score > -2.09. In this study, PALBI grade 1 was classified as low PALBI, while grade 2 and 3 were included as high PALBI. Survival analysis was conducted using Kaplan-Meier curves, and the log-rank test was utilized for data comparison. Subsequently, patients categorized under the low PALBI group were randomly split into training and validation cohorts in a 7:3 ratio. Then, we utilized a combination of advanced analytical techniques, including eXtreme Gradient Boosting (XGBoost), random survival forest (RSF), and multivariate Cox analysis, to delve into the factors influencing RFS following treatment for HBV-related HCC with low PALBI score. Based on the screened risk factors, we developed a nomogram to predict the 1-year, 3-year and 5-year RFS of low PALBI HBV-related HCC patients. Furthermore, using the scores generated by the nomogram, patients were effectively categorized into two distinct risk groups: low-risk and high-risk groups. The survival curves of the two groups were also compared by the log-rank test. Moreover, the discrimination ability, calibration, and clinical utility of the nomogram were then comprehensively evaluated to ensure its robustness and reliability.

A total of 632 patients diagnosed with HBV-related HCC who underwent TACE combined with ablation therapy were included in this retrospective study. The patients were stratified into two groups according to their preoperative PALBI scores: the low PALBI group comprised 247 individuals, while the high PALBI group consisted of 385 patients. It was confirmed that the high PALBI group exhibited higher recurrence rates compared to low PALBI group. However, delving deeper into the data, we unearthed a rather intriguing revelation: while there existed a disparity between these two groups, the RFS of patients with low PALBI scores did not demonstrate a substantial advantage over that of their high PALBI counterparts, as shown in Figure 2 . It implies that despite falling into the low PALBI category, patients were not necessarily immune to recurrence risks and warrant meticulous attention in managing their prognosis. This revelation underscores the imperative of developing specialized prognostic prediction models tailored to this subgroup, enabling clinicians to accurately assess recurrence risks and implement targeted interventions that optimize patient outcomes.

The group of 247 patients with low PALBI scores was divided into two cohorts, the training cohort (n=172) and the validation cohort (n=75), through random assignment, maintaining a ratio of 7 to 3. To ensure the validity of our analysis, a meticulous comparison of baseline characteristics between the training and validation cohorts was conducted. This comparative analysis aimed to ascertain whether there were any differences in the distribution of variables between the two cohorts. The results, as summarized in Table 1 , unequivocally demonstrated that there existed no statistically significant difference (P > 0.05) in the baseline characteristics between the training and validation cohorts, indicating their suitability for further studies.

In this study aimed at screening the factors influencing RFS, a multifaceted approach combining machine learning techniques and traditional statistical analysis was employed. At first, we utilized XGBoost, a powerful machine learning algorithm known for its effectiveness in handling complex datasets, to screen for relevant variables affecting RFS (16, 17). XGBoost is particularly adept at capturing nonlinear relationships and interactions within the data, making it suitable for identifying intricate patterns that might influence RFS outcomes. Figure 3A shows the results of XGBoost and presents the top 15 important variables, including age, GGT, cholesterol, PTA, bile acid, prealbumin, DBIL, albumin, AST, PLT, Fib, globulin, creatinine, Lym and potassium.

We then employed another machine learning technique - random survival forest (RSF), to scrutinize variables affecting RFS. Figure 3B depicts the outcomes of the RSF analysis, showcasing the top 15 significant variables, which encompass tumor number, age, GGT, PTR, INR, tumor size, PT, PTA, gender, uric acid, albumin, AST, creatinine, Child-Pugh class and glucose.

Subsequently, we took the intersection of the top 15 important variables for each of these two methods, as shown in Figure 3C . It can be seen that age, AST, albumin, GGT, creatinine and PTA are their shared variables. This strategic approach enabled us to pinpoint variables that demonstrated consistent importance across different analytical frameworks. By focusing on the shared variables identified through this process, we ensured a more robust and reliable selection of predictors for further analysis.

Following that, we engaged in a rigorous screening process using multivariate Cox regression analysis. This statistical technique allowed us to assess the significance of each variable while controlling for potential confounding factors. Variables with a significance level below 0.05 were deemed statistically significant and identified as independent risk factors influencing RFS outcomes. Table 2 shows the results of multivariate Cox analysis. We finally identified age, AST, and PTA as independent risk factors. This meticulous approach ensured that only the most relevant and influential variables were considered in our analysis.

Through the above rigorous statistical analysis, factors (age, AST and PTA) significantly associated with RFS were identified and incorporated into the nomogram ( Figure 4 ). This nomogram serves as a practical aid in clinical settings, providing a visual representation of the identified factors’ contributions to the overall risk assessment for RFS. The basic principle of the nomogram, simply put, is to assign scores to each value of each influencing factor based on the contribution (i.e., the values of regression coefficients) of these factors to the RFS in the multivariate Cox regression model. These scores are then summed up to obtain a total score, which is used to predict the 1-year, 3-year and 5-year RFS probabilities of patients.

In both the training and validation cohorts, patients were categorized into two risk groups—low-risk and high-risk—based on the scores generated by the nomogram. Subsequently, Kaplan-Meier curves were plotted for both the low and high-risk groups to assess their RFS outcomes. In the training cohort, the Kaplan-Meier curves revealed a clear demarcation between the low-risk and high-risk groups ( Figure 5 ). Patients classified as low-risk exhibited a more favorable prognosis, with a higher likelihood of survival over the follow-up period. Conversely, those categorized as high-risk faced a greater risk of recurrence, as evidenced by a steeper decline in survival probability. Similarly, the validation cohort exhibited comparable trends, further validating the prognostic utility of the nomogram-based risk stratification ( Supplementary Figure S1 ).

The evaluation of the nomogram’s performance is crucial for assessing its accuracy and reliability in predicting outcomes. Several metrics can be employed to gauge its effectiveness. Harrell’s concordance index (C-index) and Area Under the Curve (AUC) were used to evaluate its discrimination ability. Calibration curve was used to assess the agreement between predicted probabilities generated by the nomogram and observed outcomes. Decision curve analysis (DCA) was employed to evaluate its clinical utility. In the training cohort, we found that the nomogram exhibited strong discrimination ability. Its C-index was 0.689 (95% CI: 0.64–0.738) and its AUC values for 1-year, 3-year and 5-year RFS were 0.684, 0.747 and 0.813, respectively ( Figure 6 ). The 1-year, 3-year and 5-year calibration curves closely approximated the ideal diagonal line, indicating consistent predictions with actual observations ( Figure 7 ). The results of 1-year, 3-year and 5-year DCA curves demonstrated significant utility and clinical relevance ( Figure 8 ). The analysis indicated that the nomogram’s predictions offer substantial net benefit across a range of thresholds, highlighting its potential for guiding clinical decision-making effectively. In the validation cohort, the nomogram also demonstrated excellent predictive efficacy, receiving favorable evaluations across these metrics. The C-index of the nomogram in the validation cohort was 0.642 (95% CI: 0.569–0.714). Its 1-year, 3-year and 5-year AUC values were 0.689, 0.716 and 0.747, respectively ( Supplementary Figure S2 ). The calibration ( Supplementary Figure S3 ) and DCA ( Supplementary Figure S4 ) curves both showcased outstanding performance, affirming the model’s accuracy and significant clinical utility.