Navigating the evidence for hepatocellular carcinoma treatment: Surgery vs radiofrequency ablation through sentiment and meta-analysis

Abstract

BACKGROUND

Hepatocellular carcinoma (HCC) is among the most aggressive primary liver cancers, leading to significant global mortality. While early diagnosis improves prognosis, treatment decisions, particularly between surgical resection and radiofrequency ablation (RFA), remain controversial.

AIM

To clarify this issue using sentiment analysis of medical literature alongside a meta-analysis of overall survival (OS).

METHODS

We included studies comparing liver resection and RFA, excluding case reports, editorials, and studies without relevant outcomes. A systematic search in PubMed and Web of Science identified 197 studies. Abstracts underwent sentiment analysis using Python’s Natural Language Toolkit library, categorizing them as favoring resection, ablation, or neutral. We also performed a meta-analysis using a random-effects model on 11 studies reporting hazard ratios (HRs) for OS.

RESULTS

Sentiment analysis revealed that 75.1% of abstracts were neutral, 14.2% favored resection, and 10.7% favored RFA. The meta-analysis showed a significant survival advantage for liver resection, with a pooled HR of 0.5924 (95% confidence interval: 0.540-0.649). Heterogeneity was moderate (I² = 39.98%). Despite the meta-analysis demonstrating clear survival benefits of liver resection, most abstracts maintained a neutral stance. This discrepancy highlights potential biases or hesitancy in drawing definitive conclusions.

CONCLUSION

The study emphasizes the need for clinicians to prioritize robust statistical evidence over narrative impressions. Liver resection remains the preferred treatment for HCC in eligible patients.

Key Words: Liver resection; Radiofrequency ablation; Hepatocellular carcinoma; Sentiment analysis; Meta-analysis; Treatment comparison; Surgical oncology

Core Tip: With the exponential growth of medical literature, critically evaluating content is becoming increasingly complex. Reading abstracts is often the first step in selecting articles, influencing the decision-making process of clinicians and researchers. However, this study demonstrates how such an approach can lead to misleading interpretations of actual clinical outcomes, highlighting the need for a more in-depth analysis to avoid erroneous or distorted conclusions.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the leading cause of death from primary liver neoplasms and is among the most common cancers worldwide, with a prevalence of approximately 4-10 cases per 100000 individuals, representing one of the most aggressive forms of hepatic neoplasia[1]. HCC is a complex disease impacting immune responses, coagulation, hepatocyte function, and cellular growth. Its early diagnosis is essential for achieving more favorable prognosis. Primary risk factors for HCC include alcohol abuse, chronic hepatitis B and C infections, obesity, exposure to toxic substances, and hepatic cirrhosis[2]. While most HCC cases occur in individuals with pre-existing liver disease, the condition can also develop in people with a healthy liver. The treatment of HCC is based on tumor staging. When detected in an early stage, the tumor can be treated with surgical therapy (transplantation or resection) or loco-regional ablative treatments such as radiofrequency ablation (RFA) or microwave ablation[3,4].

Numerous discoveries and progressive innovations in HCC therapy have led to a significant increase in studies published in medical journals and indexed in commonly used scientific databases including PubMed, MEDLINE, Scopus, and Web of Science. However, the choice between resection and local ablation remains debated and lacks definitive consensus, and decisions in individual cases often rely on the team’s experience, medical opinions, and assessments related to the patient’s specific conditions.

The modern methods of scientific archiving have distinguished precedents in medieval jurisprudential collections. During that time, the progressive recovery of the Roman legislative corpus was achieved through collections of opinions and commentaries. By selecting certain treatises whose precedents were considered authoritative[5,6], this approach attempted to anchor case evaluations to predetermined rules. Modern medicine offers selection tools linked to advanced information technology, but selecting and distilling opinions and viewpoints to determine a general truth is more complex than ever[7]. In the case of HCC, a bibliographic search over an extended time frame easily yields a collection of several thousand scientific titles, among which the most authoritative studies must be selected. This process initially involves reading abstracts, followed by examining the selected works in full, and evaluating synthesized sources such as guidelines and meta-analyses conducted through standardized evaluation processes.

Clinical practice is often guided by the overall perception reported in the available literature. Sentiment analysis provides an innovative method to assess prevailing viewpoints within the literature, helping to clarify patterns or biases in clinical recommendations[8,9].

Today, additional techniques exist to evaluate the methods and effects of scientific decisions. Text analysis methodologies have been enhanced by the availability of computational power and advanced statistical processes, enabling analyses and comparisons that were not feasible until recently.

This study aimed to identify the most effective treatment for early-stage HCC-surgical resection or loco-regional ablative therapy-by employing sentiment analysis techniques on a significant set of medical studies and comparing the results with a classical meta-analysis. Through this innovative approach, we sought to navigate the foggy background of conflicting evidence, distilling the vast and complex literature into actionable insights to guide clinical decision-making. By integrating advanced text analysis with traditional meta-analytical methods, this study endeavored to provide clearer guidance on the optimal therapeutic approach for early-stage HCC, highlighting the potential of novel methodologies in resolving longstanding clinical debates.

MATERIALS AND METHODS

Eligibility criteria

This analysis included studies comparing liver resection and RFA for the treatment of HCC. Eligible studies met the following criteria: (1) Studies comparing liver resection with RFA as primary treatment modalities; and (2) Studies reporting overall survival (OS) as an outcome of interest.

The exclusion criteria were as follows: (1) Studies not published in English; (2) Case reports or editorial comments; (3) Studies investigating combinations of treatments (e.g., transarterial chemoembolization with resection or ablation) not directly relevant to this analysis; and (4) Studies not reporting hazard ratios (HRs) or relevant comparative outcomes.

Outcomes

The primary outcome of this meta-analysis was OS, measured as the time from treatment to death from any cause. Additionally, recurrence was analyzed as a secondary outcome, using the same studies selected for the OS analysis. Recurrence was assessed considering both recurrence-free survival (RFS) and time to recurrence, as both measure the probability of remaining free from tumor recurrence over time.

Search strategy

A systematic literature search was conducted in PubMed and Web of Science using the following search terms: ‘Liver resection’, ‘radiofrequency ablation’ and ‘hepatocellular carcinoma’. The search was limited to publications from the past 10 years, and additional filters were applied to limit the results to human studies and articles published in peer-reviewed journals. The initial search retrieved 433 studies.

The articles were screened using Rayyan, an online platform for systematic reviews. Ten duplicates were removed, resulting in 423 unique studies. These were then assessed for eligibility based on title and abstract screening. A total of 226 studies were excluded due to the following: (1) They were published in a foreign language; (2) They were case reports or editorial comments; and (3) They investigated combinations of treatments or reported outcomes not relevant to the comparison of resection vs ablation.

This left 197 studies for full-text review and sentiment analysis.

Data source and cleaning

For the 197 selected studies, the following processes were performed separately for the sentiment analysis and the meta-analysis.

Data cleaning for sentiment analysis: The textual data from the abstracts of the selected studies were cleaned and pre-processed using Python. This involved the following steps: (1) Special character and punctuation removal: All non-alphabetical characters (e.g., punctuation marks) were removed using Python’s ‘re’ module; (2) Whitespace normalization: Redundant spaces were eliminated to ensure consistent tokenization; (3) Lowercasing: The text was converted to lowercase to avoid case sensitivity issues during keyword matching; and (4) Tokenization: Using the Natural Language Toolkit (NLTK) Python library, the cleaned text was tokenized into individual words for keyword-based sentiment identification.

In this stage, all studies were retained for keyword analysis, irrespective of their specific relevance to the meta-analysis.

Data extraction and preparation for meta-analysis: Separately, data from the 197 studies were also processed for the meta-analysis. This involved the following steps: (1) Full-text screening: Each study was manually reviewed to ensure that it explicitly compared liver resection and RFA; (2) Inclusion criterion refinement: The studies had to provide HRs for OS and report multivariate analysis results, including HRs with confidence intervals (CIs); and (3) Exclusion of irrelevant studies: Studies that did not directly compare resection with ablation or did not report the outcomes of interest were excluded.

This process left 23 studies that were evaluated for the final meta-analysis[10-32]. These studies were further examined in terms of the HRs they reported for OS, as well as their statistical significance in both univariate and multivariate analyses.

Sentiment analysis

Sentiment analysis[8,9] was conducted to evaluate the tone of the selected studies with respect to the two treatments: Liver resection and RFA. The analysis was performed using Python, with the NLTK and ‘re’ libraries employed to process and classify the sentiment of the study abstracts.

To ensure reproducibility, the sentiment classification followed a rule-based approach, reducing subjective interpretation and ensuring consistent results.

After pre-processing the text, the following steps were taken.

Keyword matching: Sentiment was classified based on predefined sets of keywords related to each treatment. (1) Resection keywords: ‘Liver resection’, ‘surgery’, and ‘surgical treatment’; and (2) Ablation keywords: ‘Radiofrequency ablation’, ‘local ablative therapies’, and ‘ablation’.

Positive and negative sentiment was determined based on the presence of specific sentiment-related keywords. (1) Positive-sentiment keywords: ‘Successful’, ‘effective’, ‘preferred’, ‘beneficial’, and ‘improved’; and (2) Negative-sentiment keywords: ‘Complications’, ‘ineffective’, ‘risks’, ‘limitations’, and ‘failed’.

Sentiment classification: The algorithm scanned each abstract for the defined sets of positive- and negative-sentiment keywords associated with the two treatments. A resection score and an ablation score were calculated by counting the occurrences of positive- and negative-sentiment words related to each treatment: (1) If the number of positive-sentiment words associated with liver resection exceeded those associated with ablation, the abstract was classified as favoring liver resection or surgery; (2) Conversely, if the number of positive-sentiment words associated with RFA exceeded those associated with resection, the abstract was classified as favoring RFA or local therapies; and (3) If both treatments were associated with a comparable number of positive- and negative-sentiment words, or if no clear sentiment was expressed, the abstract was classified as neutral.

To enhance transparency, we explicitly state that the classification process was fully automated, minimizing subjective interpretation. The use of predefined keywords ensured standardized criteria, while the computational approach implemented in Python using the NLTK and regular expression (‘re’) libraries provided a structured and replicable methodology for sentiment assessment.

Data output and visual representation: The results of the sentiment analysis were summarized, and the distribution of abstracts classified as favoring resection, favoring ablation, or being neutral was visualized using Python’s ‘Matplotlib’ library: (1) A bar chart was created to show the frequency of each sentiment category across the dataset; and (2) A density plot was created to visualize the distribution of sentiment scores for the two treatments. This plot showed the overlap between abstracts favoring resection and ablation, with a neutral reference line (sentiment score of 0) indicating balanced sentiment.

Following this, studies containing comparative statements (e.g., the keyword ‘compared with’ in Rayyan) were selected for further review, narrowing down the total number of studies to 23 for the meta-analysis (Table 1).

Table 1 Overview of studies assessing outcomes of liver resection vs ablation in hepatocellular carcinoma.

Ref.	Study type	Pre-PSM sample size	Post-PSM sample size	Subgroups	Subgroup type
Bai et al[10]	Retrospective comparative study	2197	1346	Yes	Microvascular invasion risk
Chong et al[11]	Retrospective comparative study	214	118	No
Chua et al[12]	Retrospective comparative study	219	104	No
Gory et al[13]	Retrospective comparative study	146		Yes	Tumor size (< 3 cm; < 5 cm)
He et al[14]	Retrospective comparative study	435	259	No
Hsiao et al[15]	Retrospective comparative study	387		No
Kim et al[16]	Retrospective comparative study	365	122	Yes	AFP level > 100 ng/mL (yes; no)
Lai et al[17]	Retrospective comparative study	94		Yes	Age (< 60 years; > 60 years)
Liu et al[19]	Retrospective comparative study	836	600	Yes	Risk of recurrence (high; low)
Liu et al[18]	Retrospective comparative study	237	158	No
Lu et al[20]	Retrospective comparative study	332	240	Yes	Milan criteria (yes; no) and recurrence within 2 years (yes; no)
Ng et al[21]	Randomized clinical trial	218		No
Qiu et al[22]	Retrospective comparative study	259	154	No
Song et al[24]	Retrospective comparative study	156		Yes	Tumor size (< 2 cm; 2-4 cm)
Song et al[23]	Randomized clinical trial	150		No
Takayasu et al[25]	Retrospective comparative study	853	732	No
Vitale et al[26]	Retrospective comparative study	720	693	No
Wang et al[27]	Retrospective comparative study	672	420	Yes	Age (60-64 years; 65-72 years; > 73 years)
Wei et al[28]	Retrospective comparative study	382		Yes	Pre-operatorial imaging traits
Xia et al[29]	Randomized clinical trial	217		Yes	Tumor size (< 3 cm; > 3 cm) and AFP > 200 ng/mL (yes; no)
Ye et al[30]	Retrospective comparative study	388	308	Yes	Tumor size (3-4 cm; 4-5 cm)
Zeng et al[31]	Retrospective comparative study	1632	1142	Yes	Tumor size (0-2 cm; 2-5 cm; > 5 cm) and age (< 65 years; > 65 years)
Zhong et al[32]	Retrospective comparative study	847	454	No

AFP: Alpha-fetoprotein; PSM: Propensity score matching.

Meta-analysis

The meta-analysis[33] was performed using Python, specifically the ‘Pandas’, ‘NumPy’, ‘SciPy’, and ‘Matplotlib’ libraries. A total of 23 studies were reviewed, but only 11 studies met the criteria for inclusion in the final meta-analysis[10,13,14,18,20,22,25,26,29,31,32]. To ensure statistical robustness, only studies that demonstrated significant differences between the treatments in univariate analysis were considered for full inclusion. Furthermore, studies were excluded if they lacked multivariate analysis with HRs with CIs for OS[11,12,15-17,19,21,23,24,27,28,30] (Figure 1). This approach ensured that the meta-analysis was based on adequately adjusted, statistically significant and comparable findings.

Figure 1 Preferred reporting items for systematic reviews and meta-analyses flow diagram for study selection process. Flowchart depicting the selection process of studies included in the meta-analysis. The diagram outlines the number of records identified, screened, assessed for eligibility, and ultimately included in the final analysis, with reasons for exclusion specified. HR: Hazard ratio; CI: Confidence interval.

Data extraction and HR adjustment: The extracted data included HRs and 95%CIs for OS. For studies where ablation was used as the reference, the HRs were inverted (1/HR) to ensure consistency across the studies (i.e., HR < 1 favored liver resection). The CIs were similarly inverted to match the adjusted HRs. This ensured that the relative hazard was accurately reflected for all studies.

For studies that reported both overall cohort analyses and subgroup-specific results (e.g., those stratified by patient characteristics such as tumor size, age, or other clinical features), the data were treated as separate entries in the meta-analysis. Subgroup analyses were included only if they met the inclusion criteria, which required significant results in both univariate and multivariate analysis. This approach allowed the inclusion of subgroup analyses alongside the overall study results, ensuring that the distinct treatment effects observed in specific populations were not conflated with the broader study outcomes. Each of these analyses was treated independently, similarly to separate studies, to preserve the integrity of the subgroup findings and prevent the introduction of potential bias or heterogeneity. The studies requiring HR inversion included the following: (1) Vitale et al[26]; (2) Xia et al[29]; (3) Qiu et al[22]; (4) Liu et al[18]; (5) Gory et al[13]; and (6) He et al[14].

OS analysis: A random-effects model was employed to account for variations between studies in terms of sample size, study design, and patient populations. This model adjusted for heterogeneity and provided a more conservative estimate of the pooled HR. The model was selected due to anticipated differences across studies, reflecting the diversity in clinical settings and treatment protocols. Heterogeneity across the studies was assessed using Cochran’s Q test and the I² statistic. I² values of 25%, 50%, and 75% were interpreted as indicating low, moderate, and high heterogeneity, respectively. To pool the HRs, log-transformed HRs were used to standardize the effect sizes across studies. The pooled HR and its 95%CI were computed from these log HR values. The CI was derived using the formula (CI = log HR ± z × SE), where ‘z’ represents the value corresponding to the 95%CI (1.96), and SE is the standard error of the weighted mean of the log HRs. The pooled HR was then converted back to its original scale by exponentiating the log HR values, and a forest plot was generated to visualize individual study HRs and the overall pooled estimate. To assess the potential for publication bias, a funnel plot was constructed, displaying the relationship between the SE of each study’s estimate and the adjusted HR. Symmetry in the funnel plot indicates an absence of publication bias, while asymmetry suggests the presence of bias[34]. The pooled HR was marked on the funnel plot for comparison, and visual inspection was conducted to ensure that there was no significant skewness or missing studies affecting the overall analysis.

RFS analysis: Recurrence was analyzed as a secondary outcome, using the same studies selected for the OS analysis. A total of 8 studies provided HRs with 95%CIs for recurrence and were included in this analysis, while studies that did not report HRs with 95%CIs for recurrence were excluded. A random-effects model was applied to pool the HRs, maintaining methodological consistency with the OS analysis. The HRs were adjusted where necessary, inverting values for studies where ablation was the reference treatment.

RESULTS

Sentiment analysis of study abstracts

Sentiment analysis was performed on 197 study abstracts to assess their tone toward liver resection and RFA in the context of HCC treatment. The sentiment was categorized as either favoring resection, favoring ablation, or being neutral or unclear.

Of the 197 abstracts analyzed, 148 (75.1%) were classified as ‘neutral or unclear’, 28 (14.2%) favored liver resection, and 21 (10.7%) favored ablation or local therapies (Figure 2). The majority of studies presented a neutral or unclear stance, indicating a balanced or inconclusive tone in the literature regarding the superiority of one treatment over the other.

Figure 2 Sentiment distribution across study abstracts. Bar chart showing the frequency of each sentiment category in the dataset of 197 study abstracts comparing liver resection and ablation.

Sentiment distribution and overlap

The density plot in Figure 3 illustrates the distribution of sentiment scores for both liver resection and ablation. The distribution of abstracts favoring liver resection and ablation shows considerable overlap, particularly around the neutral-sentiment line. This suggests that while some studies favored one treatment modality, there was a degree of overlap in sentiment that reflected the nuanced nature of the debate between resection and ablation as treatment options for HCC.

Figure 3 Sentiment score overlap between resection and ablation. Density plot displaying the distribution of sentiment scores for the two treatments. The distribution of abstracts favoring liver resection and ablation shows considerable overlap, particularly around the neutral sentiment line (sentiment score of 0).

Meta-analysis of OS

A total of 11 studies comparing liver resection and RFA were included in the meta-analysis (Table 2 and Table 3). The baseline characteristics of patients in these studies are reported in Table 4. The pooled HR for OS across all included studies was 0.5924 (95%CI: 0.540-0.649), indicating a statistically significant survival benefit for patients undergoing liver resection compared to RFA (Figure 4A). The heterogeneity across the studies was moderate, with an I² statistic of 39.98%, suggesting that the variability in the treatments effects was not solely due to chance. Cochran’s Q test yielded a value of 21.66 (P = 0.061), indicating that the observed heterogeneity was moderate and not statistically significant (Table 5).

Figure 4 Forest plot. A: Adjusted hazard ratios for overall survival across 11 studies comparing liver resection and radiofrequency ablation. For studies that reported both overall cohort analyses and subgroup-specific results, the data were treated as separate entries in the meta-analysis. The pooled hazard ratio shows a significant survival benefit for liver resection; B: Summarizing the meta-analysis of recurrence. The plot presents the hazard ratios (HR) for each included study, comparing recurrence outcomes between liver resection and radiofrequency ablation (RFA). For studies that reported both overall cohort analyses and subgroup-specific results, the data were treated as separate entries in the meta-analysis. The pooled HR indicates a lower recurrence rate after liver resection compared to RFA. OS: Overall survival.

Table 2 Summary of meta-analysis of adjusted hazard ratios for overall survival comparing liver resection and radiofrequency ablation.

Ref.	Study type	Adjusted HR	95%CI lower	95%CI upper	Weight (%)	Treatment favored	Subgroup
Bai et al[10]	Retrospective comparative study	0.68	0.52	0.88	55.52	Resection	High risk MVI and Milan criteria
Bai et al[10]	Retrospective comparative study	0.47	0.26	0.84	11.17	Resection	High risk MVI size < 3 cm
Gory et al[13]	Retrospective comparative study	0.44	0.20	0.98	6.04	Resection	< 5 cm
He et al[14]	Retrospective comparative study	0.47	0.31	0.72	21.66	Resection	Total
Liu et al[18]	Retrospective comparative study	0.47	0.23	0.99	7.01	Resection	Total
Lu et al[20]	Retrospective comparative study	0.54	0.33	0.88	15.97	Resection	Total
Qiu et al[22]	Retrospective comparative study	0.57	0.41	0.83	30.88	Resection	Total
Takayasu et al[25]	Retrospective comparative study	0.82	0.46	1.45	11.66	No difference	Total
Vitale et al[26]	Retrospective comparative study	0.71	0.54	0.93	50.26	Resection	Total
Xia et al[29]	Randomized clinical trial	0.58	0.35	0.95	15.52	Resection	HCC > 3 cm
Xia et al[29]	Randomized clinical trial	0.54	0.34	0.87	17.19	Resection	AFP > 200 ng/mL
Zeng et al[31]	Retrospective comparative study	0.44	0.35	0.56	69.56	Resection	Age > 65 years
Zeng et al[31]	Retrospective comparative study	0.56	0.46	0.69	93.47	Resection	Age < 65 years
Zhong et al[32]	Retrospective comparative study	0.89	0.68	1.18	50.58	No difference	Total
Pooled estimate		0.59	0.54	0.65	100	Resection

HR: Hazard ratio; CI: Confidence interval; HCC: Hepatocellular carcinoma; MVI: Microvascular invasion; AFP: Alpha-fetoprotein.

Table 3 Excluded studies.

Ref.	Study type	Treatment favored (OS)
Chong et al[11]	Retrospective comparative study	Resection
Chua et al[12]	Retrospective comparative study	Resection
Hsiao et al[15]	Retrospective comparative study	Resection
Kim et al[16]	Retrospective comparative study	No difference
Lai et al[17]	Retrospective comparative study	Resection/no difference
Liu et al[19]	Retrospective comparative study	Resection
Ng et al[21]	Randomized clinical trial	No difference
Song et al[24]	Retrospective comparative study	No difference
Song et al[23]	Randomized clinical trial	No difference
Wang et al[27]	Retrospective comparative study	Ablation/no difference
Wei et al[28]	Retrospective comparative study	Resection
Ye et al[30]	Retrospective comparative study	No difference/resection

OS: Overall survival.

Table 4 Baseline characteristics of the original patient population in the included studies before propensity score matching and/or subgroup analyses, mean ± avg for numerical ones, n (%).

Ref.	Age (years)	Sex	Child-Pugh/BCLC	Tumor size (cm)	MELD
Bai et al[10]	LR: 277 (20.6) > 60 years; RFA: 213 (24.9) > 60 years	LR: M = 1138 (84.6); RFA: M = 694 (81.3)	LR: Child-Pugh A = 1243 (92.4); RFA: Child-Pugh A = 694 (81.3)	LR: 690 (51.3) < 3 cm; RFA: 710 (83.2) < 3 cm
Gory et al[13]	LR: 59.3 ± 10.7; RFA: 65.1 ± 10.0	LR: M = 42 (81); RFA: M = 71 (74)	LR: Child-Pugh A = 46 (88.4); RFA: Child-Pugh A = 75 (78.1)	LR: 3.03 ± 1.08; RFA: 2.3 ± 0.96	LR: 8.4 ± 2.1; RFA: 10.1 ± 3.2
He et al[14]	LR: 47.4 ± 12.4; RFA: 53.9 ± 12.8	LR: M = 257 (83); RFA: M = 110 (88)	LR: Child-Pugh A = 305 (98); RFA: Child-Pugh A = 118 (94)	LR: 3.3 ± 1.0; RFA: 2.6 ± 1.0
Liu et al[18]	LR: 60 ± 13; RFA: 64 ± 12	LR: M = 78 (72); RFA: M = 84 (66)	LR: BCLC 0 = 109 (100); RFA: BCLC 0 = 128 (100)	LR: 109 (100) ≤ 2 cm; RFA: 128 (100) ≤ 2 cm	LR: 7.8 ± 1.3; RFA: 8.4 ± 2.5
Lu et al[20]	LR: 50.1 ± 10.9; RFA: 52.9 ± 11.8	LR: M = 124 (89.9); RFA: M = 172 (88.7)	LR: Child-Pugh A = 138 (100); RFA: Child-Pugh A = 194 (100)	LR: 2.8 ± 1.9; RFA: 1.9 ± 0.9
Qiu et al[22]	LR: 55.5 ± 12.9; RFA: 56.6 ± 11.2	LR: M = 98 (79.6); RFA: M = 118 (86.7)	LR: Child-Pugh = 5.8 ± 1.0; RFA: Child-Pugh = 6.1 ± 1.1	LR: 3.1 ± 1.3; RFA: 2.9 ± 1.1	LR: 9.2 ± 2.7; RFA: 10.7 ± 4.8
Takayasu et al[25]	LR: 139 (79) > 60 years; RFA: 397 (80.9) > 60 years	LR: M = 104 (59.1); RFA: M = 297 (60.5)	LR: Child-Pugh A = 151 (85.8); RFA: Child-Pugh A = 394 (80.2)	LR: 176 (100) ≤ 2 cm; RFA: 491 (100) ≤ 2 cm
Vitale et al[26]	LR: 132 (44.6) > 70 years; RFA: 129 (53.8) > 70 years	LR: M = 227 (76.7); RFA: M = 181 (75.4)	LR: Child-Pugh B = 28 (9.5); RFA: Child-Pugh B = 62 (25.8)	LR: 296 (100) ≤ 3 cm; RFA: 240 (100) ≤ 3 cm	LR: ≥ 8 = 199 (67.2); RFA: ≥ 8 = 170 (70.8)
Xia et al[29]	LR: 50.0 (24.0-58.0); RFA: 52.0 (25.0-59.0)	LR: M = 107 (89.2); RFA: M = 109 (90.8)	LR: Child-Pugh A = 120 (100); RFA: Child-Pugh A = 120 (100)	LR: 81 (67.5) < 3 cm; RFA: 78 (65.0) < 3 cm
Zeng et al[31]	LR: 66-75 years = 262 (65.0); RFA: 66-75 years = 152 (60.8)	LR: M = 281 (69.7); RFA: M = 172 (68.8)		LR: 180 (44.6) = 2-5 cm; RFA: 184 (73.6) = 2-5 cm
Zhong et al[32]	LR: 90 (29.3) ≥ 60 years; RFA: 184 (34.1) ≥ 60 years	LR: M = 245 (79.8); RFA: M = 482 (89.3)	LR: Child-Pugh A = 300 (97.7); RFA: Child-Pugh A = 523 (96.9)	LR: 135 (44.0) < 3 cm; RFA: 425 (78.7) < 3 cm

BCLC: Barcelona clinic liver cancer; LR: Liver resection group; RFA: Radiofrequency ablation group; MELD: Model of end stage liver disease; M: Male.

Table 5 Meta-analysis summary (overall survival).

Statistic	Value
Number of studies reviewed	23
Number of studies included in meta-analysis	11
Pooled hazard ratio	0.5924
95% confidence interval for hazard ratio	0.5404-0.6493
Heterogeneity (I²)	39.98%
Cochran’s Q statistic	21.66
P value for heterogeneity	0.0608
Model used	Random effects
Significance threshold	HR < 1 favors liver resection

HR: Hazard ratio.

Given the moderate level of heterogeneity (I² = 39.98%), a random-effects model was employed. This model accounted for variability across studies and provided a more conservative and robust estimate of the overall treatment effect. As such, the pooled HR reflected the variations in the study designs, populations, and methodologies, allowing for a more accurate interpretation of OS outcomes.

To assess the potential for publication bias, a funnel plot was generated (Figure 5). The plot displays symmetry, suggesting no significant evidence of publication bias. The SEs of the study estimates are plotted against the adjusted HRs, with the pooled HR (0.5924) marked on the graph. The lack of substantial asymmetry indicates that the findings from the included studies were unlikely to have been influenced by publication bias.

Figure 5 Funnel plot for publication bias. The plot evaluates the presence of publication bias in the meta-analysis. The lack of substantial asymmetry suggests that the findings from the included studies were unlikely to be influenced by publication bias. HR: Hazard ratio.

Meta-analysis of recurrence

A total of 8 studies were included in the meta-analysis. The pooled HR for recurrence was 0.6134 (95%CI: 0.554-0.679), indicating a significantly lower recurrence rate after liver resection compared to RFA (Figure 4B) (Table 6).

Table 6 Summary of meta-analysis of adjusted hazard ratios for recurrence comparing liver resection and radiofrequency ablation.

Ref.	Study type	Adjusted HR	95%CI lower	95%CI upper	Weight (%)	Treatment favored	Subgroup
Bai et al[10]	Retrospective comparative study	0.78	0.63	0.97	82.50	Resection	High risk MVI and Milan criteria
Bai et al[10]	Retrospective comparative study	0.51	0.32	0.81	17.81	Resection	High risk MVI e size < 3 cm
Gory et al[13]	Retrospective comparative study	0.48	0.28	3.52	2.42	No difference	< 5 cm
He et al[14]	Retrospective comparative study	0.43	0.31	3.2	2.75	No difference	Total
Liu et al[18]	Retrospective comparative study	0.41	0.26	3.8	2.12	No difference	Total
Qiu et al[22]	Retrospective comparative study	0.63	0.46	2.1	6.56	No difference	Total
Takayasu et al[25]	Retrospective comparative study	0.52	0.35	0.7	24.72	Resection	Total
Zeng et al[31]	Retrospective comparative study	0.43	0.32	0.56	49.07	Resection	Age > 65 years
Zeng et al[31]	Retrospective comparative study	0.57	0.45	0.73	65.65	Resection	Age < 65 years
Zhong et al[32]	Retrospective comparative study	0.68	0.57	0.82	116.19	Resection	Total
Pooled estimate		0.61	0.55	0.68	100	Resection

HR: Hazard ratio; CI: Confidence interval; MVI: Microvascular invasion.

The heterogeneity was moderate (I² = 38.48%), with a Cochran’s Q test P value of 0.101, suggesting that variability among studies was present but not statistically significant.

Three studies did not provide HRs with 95%CIs for recurrence and were therefore excluded: (1) Lu et al[20]; (2) Vitale et al[26]; and (3) Xia et al[29].

DISCUSSION

The present study aimed to assess the relative effectiveness of liver resection vs RFA in patients with HCC by combining quantitative meta-analysis with a sentiment analysis of the published literature. The results of the primary meta-analysis indicated a clear survival benefit favoring liver resection, with a pooled HR of 0.5924 (95%CI: 0.540-0.649). Additionally, the secondary meta-analysis assessing recurrence as an endpoint showed a pooled HR of 0.6134 (95%CI: 0.554-0.679), indicating a lower recurrence rate after liver resection compared to RFA.

Despite this quantitative evidence, the sentiment analysis revealed that a substantial proportion of the included studies did not take a strong stance in favor of either treatment. Indeed, most abstracts were classified as ‘neutral or unclear’, suggesting a hesitancy among authors to express definitive conclusions about the superiority of one treatment over the other.

This discrepancy between the narrative tone of the studies and the quantitative findings is noteworthy. While the meta-analysis clearly supported liver resection as the superior treatment in terms of OS, many studies failed to reflect this in their qualitative conclusions. Specifically, only 28 abstracts favored liver resection, while 148 were classified as neutral or unclear. This highlights a potential issue in the way findings are communicated within the scientific literature. Several factors may contribute to the predominance of neutral abstracts, including publication bias, editorial policies, and institutional or funding-related influences. Authors may be reluctant to express strong preferences due to concerns about study limitations, potential biases, or the complexity of individual patient cases, particularly in observational studies. Additionally, the pressure to maintain neutrality may stem from journal editorial policies, which often favor cautious wording to prevent the overgeneralization of findings, especially when studies are not randomized or have methodological constraints.

Beyond editorial guidelines, publication bias may also play a role, as authors might avoid making definitive claims that could reduce the likelihood of their manuscripts’ acceptance or provoke controversy within the scientific community. Institutional affiliations and funding sources could further shape the phrasing of abstracts, as researchers may hesitate to present conclusions that contradict prevailing clinical guidelines or challenge well-established treatment paradigms. These factors may have collectively contributed to the observed disconnect between the quantitative evidence from the meta-analysis and the narrative framing of the literature. Addressing these underlying factors is essential to improving the clarity of scientific communication. While maintaining scientific rigor is crucial, ensuring that abstracts accurately reflect study conclusions can help prevent misinterpretations that may affect clinical decision-making. Using precise and unambiguous language in abstracts, rather than overly cautious or equivocal phrasing, may reduce the risk of misinterpretation and provide clinicians with clearer insights into the comparative effectiveness of treatment options.

Our findings also underscore the importance of a rigorous quantitative approach when evaluating treatment efficacy. If the conclusions of clinicians and decision-makers were solely based on the abstracts’ narratives, the evidence supporting liver resection might have been overlooked.

Heterogeneity among the included studies was moderate, as indicated by an I² of 39.98%, suggesting that some variability in the reported outcomes was likely due to differences in the study designs, patient populations, and methodologies. Regional differences, variations in surgical techniques, and center-specific approaches to RFA may have contributed to inconsistencies in the treatment outcomes. Differences in operator expertise, equipment availability, and treatment protocols could have further accounted for variability in the results across studies. These factors highlight the complexity of comparing outcomes between liver resection and RFA, as procedural nuances may influence patient prognosis beyond the primary treatment modality.

Additionally, patient-related factors such as comorbidities, particularly diabetes, may have influenced treatment outcomes. Diabetes is a well-recognized risk factor for HCC development[35], and its impact on post-treatment prognosis remains debated. Some studies suggest that diabetes is associated with worse perioperative outcomes, whereas others indicate that with appropriate perioperative management, its influence may be mitigated. The careful selection of candidates, the adoption of minimally invasive techniques, and optimized perioperative care are essential to minimizing complications in diabetic patients. Both surgical and ablative treatments appear to carry an increased risk of adverse events in diabetic individuals, particularly if metabolic and cardiovascular risk factors are not well controlled[36-39]. Although the random-effects model was employed to account for inter-study variability, these sources of heterogeneity should be considered when interpreting the findings. Notably, while the results consistently favored liver resection, the magnitude of this effect varied across studies, particularly those requiring HR inversion due to the use of ablation as the reference treatment.

The sentiment analysis of the 197 abstracts further revealed a largely neutral tone regarding the comparison between liver resection and RFA. This finding highlights a potential disconnect between narrative-driven perceptions and the hard data presented in the meta-analysis.

The decision to analyze abstracts rather than full texts was made to simulate real-world research practices, where clinicians and researchers primarily rely on abstracts as an initial screening tool to assess the findings of scientific studies. Given the exponential growth of medical literature, researchers and clinicians often rely on the initial reading of abstracts to select relevant studies, making them a key component in shaping early impressions of treatment efficacy. However, as demonstrated in this study, reliance on abstracts alone can sometimes lead to misleading interpretations, as they may not fully reflect the actual clinical outcomes.

In clinical practice, sentiment-laden abstracts, which are often the first point of reference for clinicians and researchers, may influence treatment decisions or guide biases, especially when abstracts do not take a definitive stance. Given that our meta-analysis showed a significant survival benefit for liver resection over RFA, reliance on neutral or equivocal sentiments in abstracts could obscure the effectiveness of resection as the preferred treatment. The evident advantage of liver resection over RFA should be clearly articulated in scientific abstracts and publications. Avoiding vague or ambiguous expressions and instead taking a definitive stance on the superiority of one treatment can help prevent errors in clinical decision-making. Ensuring that findings are communicated in a straightforward and decisive manner is crucial to effectively guide treatment choices. Therefore, clinicians must base decisions on comprehensive, quantitative evidence rather than rely on potentially misleading narrative impressions.

Limitations

This study has several limitations that warrant consideration. First, the meta-analysis included 11 studies, with some variation in the study designs, patient populations, and treatment protocols. Although the random-effects model was employed to account for inter-study heterogeneity, as indicated by an I² statistic of 39.98%, residual heterogeneity may have still been present, potentially due to unmeasured factors such as differences in follow-up periods or baseline characteristics.

Second, inverting HRs for studies that used RFA as the reference group ensured consistency in interpreting the results (i.e., HR < 1 favors liver resection). While this step was essential for maintaining methodological rigor, it may have introduced some variability in the interpretation, as the original studies may have employed different approaches to report outcomes.

Third, the sentiment analysis was based on abstracts. While representative of the overall study findings, they may not have fully captured the depth of discussions and detailed results presented in the full-text articles. The abstracts may have also been subjected to bias or oversimplification, particularly in complex comparative studies like those evaluating resection vs ablation. However, using abstracts is a common approach in literature reviews and still provides a meaningful snapshot of the overall positioning of each study. While a full-text analysis would have provided a more comprehensive perspective, it would have presented significant computational challenges, including the need for extensive text cleaning to remove figures, tables, and other non-textual elements that frequently contain critical results. Additionally, many conclusions in full-text articles are conveyed through visual data rather than explicitly stated in the text, further complicating automated extraction and analysis. Given these limitations and considering that this study aimed to evaluate sentiment trends specifically in abstracts, a full-text analysis was beyond the scope of this work. However, we acknowledge that such an analysis would be a valuable extension to confirm and expand upon our findings in future research.

Additionally, the focus of this meta-analysis was primarily on OS as the primary endpoint, which is a critical and relevant outcome in the context of HCC treatment. Since recurrence was a secondary outcome, its analysis was conducted using the same studies selected for the primary outcome. However, other important endpoints such as complication rates and quality of life were not included in this analysis. Furthermore, the number of studies included in the meta-analysis was limited due to the rigorous selection criteria applied to ensure methodological robustness and data reliability. While these criteria were essential for maintaining this study’s quality, they resulted in a smaller sample size, which may impact the generalizability of the findings. Finally, while methodologically appropriate, the keyword-based sentiment analysis had intrinsic limitations, such as an inability to capture contextual subtleties or the potential for misclassification. Although we carefully selected and applied sentiment-related keywords, this approach may not have fully reflected the complexity of discussions in the literature. Nevertheless, it offered the advantage of reproducibility, ensuring consistent results across analyses. More advanced methods, such as transformer-based models, including Bidirectional Encoder Representations from Transformers or Biomedical Bidirectional Encoder Representations from Transformers, provide deeper contextual understanding. However, they introduce variability, as their outcomes are influenced by probabilistic components and can differ across repeated runs, making standardization more challenging[40].

Despite these limitations, the sentiment analysis effectively complemented the quantitative findings, shedding light on the tone and tendencies in the literature regarding these treatment modalities.

Key points

(1) There was a notable discrepancy between the predominantly neutral sentiment of the abstracts and the clear survival benefit of liver resection demonstrated by the meta-analysis; (2) Clinicians should prioritize quantitative data and rigorous analysis over narrative impressions when making treatment decisions; (3) Liver resection should be considered the preferred treatment for HCC in eligible patients, as it offers a significant survival benefit; and (4) RFA remains a valuable treatment option for patients who are not candidates for surgery due to significant comorbidities or intrinsic contraindications.

CONCLUSION

The observed discrepancy between the neutral sentiment of the abstracts and the clear survival benefit of liver resection in the meta-analysis emphasizes the importance of relying on robust statistical data when making clinical decisions. Abstracts that do not clearly state the superior outcomes of resection may lead clinicians to undervalue its benefits, further reinforcing the need for the rigorous interpretation of both qualitative and quantitative evidence. This study underscores the need for clinicians and decision-makers to rely on detailed meta-analytical data rather than abstract summaries when evaluating treatment options. Future research should aim to reconcile these differences and encourage more definitive conclusions in the reporting of treatment efficacy. These findings also point to the importance of ongoing efforts to standardize reporting practices in clinical research. By ensuring that both narrative and quantitative conclusions are aligned, the medical community can better support evidence-based decision-making, ultimately improving patient outcomes in HCC.