This study retrospectively enrolled 728 HCC patients who received TACE combined ablation at Beijing Youan Hospital, Capital Medical University, between January 2018 and December 2023. HCC diagnosed by histological or radiological criteria as defined by the American Association for the Study of Liver Diseases (AASLD) guidelines (19). Patients were stratified into two groups based on the HALP score cutoff value (-56.8), as established in prior research (15): 422 patients with high HALP scores (> -56.8, H-HALP group) and 306 patients with low HALP scores (≤ -56.8, L-HALP group). The H-HALP cohort was further randomly divided into a training cohort (n=296) and an internal validation cohort (n=126) at a 7:3 ratio, with a randomization seed set at 400. Furthermore, an independent external validation cohort comprising 147 H-HALP patients treated at Beijing Ditan Hospital during the same period was included to test the model’s robustness. Inclusion criteria were as follows: (1) aged ≥ 18 years; (2) BCLC stage 0, A or B; (3) Child-Pugh class A or B liver function; (4) patients with tolerable general status: Eastern Cooperative Oncology Group (ECOG) performance status of 0–2 and stable organ function adequate for interventional therapy; (5) underwent TACE followed by ablation, with radiological confirmation of complete ablation, defined as the absence of contrast enhancement in the treated lesion on follow-up imaging one month after the procedure. Exclusion criteria included: (1) received systemic drugs before TACE combined with ablation, including sorafenib, lenvatinib, PD-1 inhibitors, etc. (2) presence of other primary malignant tumors. (3) lost to follow-up. (4) contraindication to TACE or ablation.

This study was conducted by the Declaration of Helsinki, and experienced clinicians determined patient eligibility for combined therapy based on guidelines. In addition, this study was approved by the Ethics Committee of Beijing Youan Hospital, Capital Medical University, and informed consent from the patients was waived due to its retrospective nature. All patient data were de-identified to protect privacy.

Demographic and clinicopathological data were retrospectively collected for analyzed. Baseline characteristics included age, gender, tumor size, tumor number, hypertension, diabetes, cirrhosis, antiviral treatment, BCLC stage and Child-Pugh classification. Laboratory values were obtained from the closest test performed within 7 days before treatment. If multiple measurements were available, the most recent value prior to treatment initiation was used for analysis. Laboratory parameters included Hb, red blood cell (RBC) count, white blood cell (WBC) count, ALB, LYM count, PLT count, alanine transaminase (ALT), aspartate aminotransferase (AST), total bilirubin (TBIL), direct bilirubin (DBIL), γ-glutamyl transpeptidase (GGT), globulin (Glob) and alpha-fetoprotein (AFP). The HALP score was calculated as Hb (g/L) × ALB (g/L) × LYM (10⁹/L)/PLT (10⁹/L).

All patients underwent conventional TACE (cTACE) performed by two interventional radiologists with over five years of experience. The procedure was carried out under local anesthesia using the Seldinger technique via femoral artery access. Tumor-feeding arteries were identified through digital subtraction angiography (DSA). A chemotherapeutic emulsion consisting of 20 mg of epirubicin mixed with 6–10 mL of lipiodol was selectively injected into the tumor-feeding vessels. This was followed by embolization using gelatin sponge particles until stasis of blood flow was achieved. If any adverse reactions occurred during the procedure, symptomatic treatment was administered accordingly.

Ablation was conducted within two weeks post-TACE, utilizing radiofrequency ablation (RFA; Cool-tip RF Ablation System, Covidien, USA) and microwave ablation (MWA; ECO Microwave Ablation System, China). Among all patients, 312 (42.9%) received RFA and 416 (57.1%) received MWA. The ablation range completely covered the tumor to the edge of 0.5-1.0 cm to prevent marginal residue and recurrence. For MWA, a typical power setting of 40 to 60 watts was used, with an average ablation time of approximately 5 minutes per site. For more aggressive or multi-focal lesions requiring repeated ablation, total energy application ranged accordingly. In RFA procedures using emission-frequency modes, higher power outputs (120–160 watts) were employed, with each ablation site treated for 6–8 minutes. The selection of power and duration was individualized based on tumor location, size, and surrounding structures. The ablation protocol comprised the following steps: (1) determining the appropriate ablation position using contrast-enhanced computed tomography (CT) or magnetic resonance imaging (MRI); for MRI, a 1.5T system with T1-weighted dynamic contrast-enhanced sequences and T2-weighted imaging was used to assess tumor extent and vascular involvement. (2) inserting the ablation needle at the marked site and monitoring the procedure via imaging; (3) expanding the ablation area as necessary, considering multiple sites and potential overlapping or repeated ablation; (4) heating the needle track in the final phase to prevent tumor implantation and postoperative bleeding; and (5) conducting post-ablation imaging to evaluate treatment efficacy and complications, with follow-up contrast-enhanced CT or MRI performed one-month post-procedure.

All patients were followed up every three months during the first year and every six months thereafter in the outpatient clinic, in accordance with the Chinese Guidelines for Primary Liver Cancer (2024 edition) (20) and the AASLD guidelines (19). Follow-up assessments included liver function tests, routine blood tests, serum alpha-fetoprotein (AFP) levels, and imaging examinations (contrast-enhanced CT or MRI).

The final follow-up date was December 31, 2023. The median follow-up time was 25.1 months. Radiographic recurrence was defined as the appearance of new intrahepatic or extrahepatic enhancing lesions consistent with viable tumor, based on the modified Response Evaluation Criteria in Solid Tumors (mRECIST) (21). Recurrence-free survival (RFS) was calculated from the date of initial treatment to the date of radiologic diagnosis of recurrence.

All statistical analysis were performed using R software (version 4.1.3, http://www.rproject.org), with a two-sided P < 0.05 considered statistically significant. Demographic and clinical characteristics were compared among the training, internal validation and external validation cohorts. Continuous variables were presented as mean ± standard deviation and compared using the Student’s t-test or Mann-Whitney U test, depending on data distribution. Categorical variables were described as frequencies (percentages) and compared by Chi-square test. RFS was estimated using the Kaplan-Meier method, and differences between groups were assessed by the log-rank test. Independent predictive factors of RFS were evaluated using univariate and multivariate Cox proportional hazards models, incorporating variables with P < 0.05 from the univariate analysis into the multivariate analysis. Furthermore, LASSO-Cox regression was performed in the training cohort to identify independent prognostic factors for RFS, which were used to construct a nomogram. The predictive performance of the nomogram was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC). Additionally, the nomogram’s predictive accuracy was also compared with the latest version of the American Joint Committee on Cancer (AJCC) TNM staging system and the factors (cirrhosis, tumor number and GGT) used to construct the nomogram. Calibration curves and decision curve analysis (DCA) were employed to assess the model’s calibration accuracy and clinical utility, respectively. External validation was conducted using an independent cohort from Beijing Ditan Hospital. The established prediction model was applied to the external data to calculate individual risk scores, and model performance was evaluated as described above. Patients in each cohort (training, internal validation, and external validation) were independently stratified into high- and low-risk groups based on the median risk score within the respective cohort. Kaplan-Meier analysis and the log-rank test were conducted separately in each cohort to compare RFS between risk groups.

This study enrolled a total of 728 patients with HCC who received TACE combined with ablation at Beijing Youan Hospital. Among these, 422 patients stratified into the H-HALP group and 306 patients into the L-HALP group. Although a statistically significant difference was observed (P = 0.024), the median recurrence-free survival (mRFS) was similar for both groups: 1.84 years (95% CI: 1.44-2.04) for the H-HALP group and 1.60 years (95% CI: 1.39-1.89) for the L-HALP group ( Figure 1 ). The 1-, 3-, and 5-year RFS rates for the H-HALP group were 70.8%, 36.2%, and 21.5%, respectively, which was only marginally higher than those of the L-HALP group (1-year: 66.3%; 3-year: 28.2%; 5-year: 16.7%). These findings suggest that despite the slightly improved outcomes in the H-HALP group, the prognosis and recurrence patterns still require further exploration. Furthermore, as shown in Table 1 , both univariable and multivariable Cox regression analyses identified HALP as an independent predictor of RFS (HR: 1.30; 95% CI: 1.09-1.55; P = 0.04). Based on these findings, subsequent analyses focused on the subset of patients within the H-HALP group.

Patients in the H-HALP were randomly divided into the training cohort (n=296) and the internal validation cohort (n=126) in a 7:3 ratio. The external validation cohort consisted of 147 individuals. Baseline characteristics are summarized in Table 2 . In both cohorts, the majority of patients were male (73.6% vs. 74.6% vs. 82.3%) and aged over 55 years old (58.1 ± 8.68 vs. 57.9 ± 8.79 vs. 57.16 ± 7.97). In the training cohort, 259 (87.5%) patients were diagnosed with cirrhosis, 86 (29.1%) with hypertension, and 75 (25.3%) with diabetes mellitus, respectively. In the internal validation cohort, 110 (87.3%) had cirrhosis, 39 (31.0%) had hypertension, 30 (23.8%) had diabetes mellitus. In the external validation cohort, cirrhosis was diagnosed in 135 patients (91.8%), with hypertension and diabetes affecting 36 (24.5%) and 35 (23.8%), respectively. Furthermore, more than half of the patients had received antiviral treatment (59.5% vs. 58.7% vs. 55.8%). In terms of tumor characteristics, most patients had solitary tumor (72.0% vs. 69.8% vs. 68.0%) and tumor size were less than 3cm (63.2% vs. 65.0% vs. 74.1%). Statistical analysis indicated no significant differences between the training and validation cohorts (P > 0.05).

The LASSO regression analysis was used to screen parameters, and the variation characteristics of the coefficient of these variables were shown in Figure 2A . The 10-fold cross-validation method was applied to the iterative analysis, and a model with excellent performance but minimum number of variables was obtained when λ was 0.065 (Log λ = -1.19) ( Figure 2B ). Eight candidate predictors were identified, including gender, antiviral, cirrhosis, BCLC stage, tumor number, TBIL, GGT and Glob ( Figure 3 ). These predictors were then assessed through multivariate Cox regression, revealing cirrhosis (P = 0.027), multiple tumor (P = 0.021) and GGT (P < 0.001) were independent predictors of RFS.

The constructed nomogram incorporated above three screened features identified through multivariate Cox regression ( Figure 4 ). To facilitate clinical translation, we developed an interactive web-based version of the nomogram, which is publicly accessible at https://joenomogogogo.shinyapps.io/DynNomapp/. To comprehensively evaluate its discriminatory performance, we compared the nomogram with the AJCC staging system and the individual predictors (cirrhosis, tumor number, and GGT) using time-dependent ROC analysis at 1-, 3-, and 5-year time points. Notably, five-year RFS data were unavailable in the external validation cohort due to insufficient follow-up time in 18.4% of patients. Therefore, 4-year AUC was reported as an alternative endpoint. The nomogram consistently demonstrated superior or comparable AUC values across all cohorts.

In the training set, the nomogram achieved AUCs of 0.665 (95% CI: 0.599-0.731), 0.694 (95% CI: 0.619-0.768), and 0.671 (95% CI: 0.564-0.779) at 1-, 3-, and 5-year time points, respectively, outperforming the AJCC staging system (AUCs: 0.587 [95% CI: 0.527-0.647], 0.580 [95% CI: 0.520-0.640], and 0.578 [95% CI: 0.484-0.673]; P = 0.00361, 2.4e-05, and 0.0244) as well as individual predictors ( Figures 5A-C ). In the internal validation cohort, the nomogram yielded AUCs of 0.622 (95% CI: 0.506-0.738), 0.606 (95% CI: 0.485-0.727), and 0.561 (95% CI: 0.372-0.751) at the same time points, comparable to those of the AJCC system (0.612 [95% CI: 0.515-0.709], 0.556 [95% CI: 0.462-0.650], and 0.577 [95% CI: 0.424-0.730]; P = 0.789, 0.179, and 0.826) ( Figures 5D-F ). Similar trends were observed in the external validation cohort, where the nomogram achieved AUCs of 0.569 (95% CI: 0.465-0.673), 0.615 (95% CI: 0.499-0.730), and 0.662 (95% CI: 0.505-0.819) (at 1-, 3-, and 4-year time points), compared with 0.554 (95% CI: 0.470-0.637; P = 0.656), 0.611 (95% CI: 0.528-0.693; P = 0.931), and 0.636 (95% CI: 0.538-0.734; P = 0.667) for the AJCC system ( Figures 5G-I ).

Although the differences in the validation cohorts did not reach statistical significance, likely due to relatively small sample sizes and limited follow-up time, the nomogram demonstrated a consistent trend of improved discrimination across all cohorts.

To further evaluate predictive accuracy, we calculated the concordance index (C-index) for each model. In the training set, the nomogram achieved a C-index of 0.646 (95% CI: 0.608-0.684), comparable to AJCC (0.678; 95% CI: 0.615-0.741), while cirrhosis, tumor number, and GGT yielded C-indices of 0.650 (95% CI: 0.542-0.759), 0.692 (95% CI: 0.624-0.759), and 0.579 (95% CI: 0.537-0.621), respectively. In the internal validation cohort, the nomogram’s C-index was 0.614 (95% CI: 0.546-0.683), close to that of AJCC (0.662; 95% CI: 0.565-0.759), and tumor number (0.661; 95% CI: 0.551-0.771), and clearly higher than cirrhosis (0.402; 95% CI: 0.243-0.561) and GGT (0.560; 95% CI: 0.49-0.631). In the external validation cohort, the nomogram attained a C-index of 0.577 (95% CI: 0.519-0.636), similar to AJCC (0.612; 95% CI: 0.517-0.706) and tumor number (0.622; 95% CI: 0.523-0.720), and superior to cirrhosis (0.524; 95% CI: 0.359-0.689) and GGT (0.546; 95% CI: 0.485-0.607). These results suggest that the nomogram offers stable and balanced discriminatory performance across datasets, supporting its potential clinical utility in predicting RFS in patients with HCC undergoing TACE combined with ablation.

Furthermore, calibration curves illustrated well agreement between the predicted outcomes and the actual observations ( Figures 6A-C ). And the DCA confirmed the clinical utility of the nomogram, with net benefits consistently exceeding default strategies across risk thresholds ( Figures 6D-L ).

The above analyses demonstrated the good predictive effect of the nomogram. We calculated the prediction score based on the three variables in the nomogram. A median cutoff value was used to separate the patients in the training cohort into a low-risk group (n = 149) and a high-risk group (n = 147). In the training cohort, a significantly prolonged RFS have been observed in the low-risk group (3.76 years, 95% CI: 2.29-4.44) compared with the high-risk group (1.33 years, 95% CI: 1.03-1.50, P < 0.0001) ( Figure 7A ). Additionally, the cumulative RFS rates for low-risk patients at 1-, 3-, and 5-year were 0.81, 0.52 and 0.29, respectively, while for high-risk patients, the rates were 0.59, 0.19 and 0.11. In the internal validation cohort, there were 63 patients in each of the low-risk and high-risk groups. The low-risk group achieved a median RFS of 4.48 years (95% CI: 2.48-not reached), with 1-, 3-, and 5-year RFS rates of 0.81, 0.53 and 0.40, respectively. In contrast, the high-risk cohort displayed markedly reduced survival, registering a median RFS of 1.53 years (95% CI: 1.11-2.18, P = 5e-04). Corresponding 1-, 3-, and 5-year RFS rates were 0.63, 0.25 and 0.14, respectively ( Figure 7B ). External validation further corroborated this risk stratification pattern. Patients in the low-risk group had significantly longer RFS compared to patients in the high-risk group (1.84 years [95% CI: 1.51-2.79] vs. 1.38 years [95% CI: 1.13-1.97], P = 0.025) ( Figure 7C ). The 1-, 3-, and 5-year RFS rates were 0.74, 0.34 and 0.22 in the low-risk patients, while 0.63, 0.20 and 0.08 in the high-risk patients, respectively. All these results proved that our model can effectively distinguish the recurrence risk.