Rethinking carcinoembryonic antigen and carbohydrate antigen 19-9 in gastric cancer surveillance: From statistical association to clinical utility

Abstract

Gastric cancer (GC) remains a leading cause of cancer-related mortality worldwide, and identifying reliable prognostic biomarkers for postoperative surveillance is a persistent clinical challenge. Serum carcinoembryonic antigen and carbohydrate antigen 19-9 have long been studied for their potential to predict recurrence and guide treatment in this context, yet their clinical utility remains incompletely validated. A recent study examined these biomarkers as predictors of early recurrence and gastrointestinal chemotherapy toxicity in patients receiving adjuvant S-1 plus oxaliplatin therapy after gastrectomy. While the study provides clinically interesting data, several methodological and statistical shortcomings limit the strength and generalizability of its conclusions. In this opinion review, we undertake a comprehensive critical reappraisal of the study’s design, analytical approach, and interpretive framework. We examine five key areas of concern: selection bias arising from exclusion of poorly compliant patients, temporal ambiguity in biomarker-outcome relationships, inadequate control of clinically relevant confounders, fundamental flaws in survival analysis methodology, and the absence of predictive performance evaluation. Beyond the specific critique, we place these issues within the broader landscape of biomarker research in GC, discussing the evolving role of molecular subtypes, liquid biopsy technologies such as circulating tumor DNA, and modern statistical frameworks, including decision curve analysis and the TRIPOD guidelines. We argue that the field must move beyond simple correlative biomarker studies toward methodologically rigorous, prospectively validated, and clinically actionable prognostic models. Finally, we outline a practical roadmap for future research, identifying the key steps required to translate promising biomarker signals into tools that can meaningfully improve postoperative decision-making in gastrointestinal oncology.

Key Words: Gastric cancer; Carcinoembryonic antigen; Carbohydrate antigen 19-9; Biomarkers; Recurrence; Chemotherapy toxicity; Liquid biopsy; Postoperative surveillance; Adjuvant therapy

Core Tip: This opinion review critically examines a recent study linking carcinoembryonic antigen and carbohydrate antigen 19-9 to gastric cancer recurrence and chemotherapy toxicity, identifying fundamental methodological gaps that limit clinical applicability. The concerns raised-selection bias, reverse causality, inadequate confounding control, mishandled survival analysis, and absent predictive performance assessment-are not unique to one study; they represent recurring weaknesses across the biomarker literature in gastrointestinal oncology. We argue that the field must embrace modern biostatistical frameworks, integrate molecular subtyping and liquid biopsy technologies, and adopt rigorous validation standards before any single biomarker can reliably guide postoperative management.

INTRODUCTION

Gastric cancer (GC) remains a formidable adversary in global oncology. Despite meaningful progress in surgical techniques, perioperative chemotherapy regimens, and immunotherapeutic strategies, it continues to rank as the fifth most commonly diagnosed malignancy and the fourth leading cause of cancer-related death worldwide[1]. For patients who undergo curative-intent gastrectomy, the postoperative period is a critical window: Recurrence rates remain stubbornly high, reaching approximately 30% to 60% depending on disease stage, even among those receiving adjuvant therapy[2,3]. The fundamental challenge, then, is not merely whether recurrence will occur, but whether we can detect it early enough-and with sufficient precision-to meaningfully alter the patient’s trajectory.

It is against this backdrop that serum biomarkers such as carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9) have attracted sustained clinical interest. Both markers have a long history in gastrointestinal oncology, having been studied for decades as potential tools for cancer detection, surveillance, and prognostication[4-7]. Their appeal is obvious: They are inexpensive, widely available, minimally invasive, and can be serially monitored. However, the clinical reality is considerably more complicated. Neither CEA nor CA19-9 has the sensitivity or specificity required for standalone diagnostic or prognostic use in GC[7]. A wide range of non-malignant conditions influences their levels, and their discriminatory power varies substantially across disease stages and molecular subtypes.

A recent study by Qi and Wu[8] sought to advance our understanding of these biomarkers by examining the correlation between postoperative serum CEA and CA19-9 levels and two clinically relevant outcomes: Recurrence within one year and gastrointestinal toxicity during adjuvant S-1 plus oxaliplatin (SOX) chemotherapy. The study measured biomarker levels at three postoperative time points, averaged the values, and examined their relationships with recurrence-free survival (RFS) and toxicity severity, graded according to CTCAE v5.0 standards. While the research addresses an important clinical question and provides hypothesis-generating data, several methodological and statistical limitations warrant careful consideration before these findings can be incorporated into clinical practice.

In this opinion review, I undertake a structured, in-depth critical appraisal of Qi and Wu’s study[8], examining five major areas of concern. Beyond the specific critique, I situate these issues within the broader landscape of biomarker research in GC, discuss emerging alternatives, including liquid biopsy technologies and molecular subtyping, and propose a practical research roadmap to advance the field toward clinically actionable prognostic tools.

SELECTION BIAS

Perhaps the most consequential design decision in the study by Qi and Wu[8] was restricting enrollment to patients who completed all six cycles of SOX chemotherapy with “good compliance” (defined as achieving at least 80% of the planned dose intensity). On the surface, this seems like a reasonable approach to studying a homogeneous population. In practice, it introduces substantial selection bias that fundamentally compromises the study’s external validity[9].

Here is the core problem: Patients who fail to complete a full course of adjuvant chemotherapy are not a random subset of the study population. They are disproportionately the ones experiencing the most severe toxicity, the most aggressive disease biology, or both[10]. By systematically excluding them, the study effectively creates a “survivor cohort” a group that, by definition, was healthy enough and tolerant enough to endure the full treatment regimen. This has two direct consequences. First, it underestimates the true incidence of chemotherapy-related gastrointestinal toxicity, because the very patients most likely to experience severe adverse events are removed from the analysis. Second, it potentially distorts the observed associations between biomarkers and outcomes, since the patients most likely to have elevated biomarkers driven by aggressive tumor biology are also those most likely to have been excluded due to early treatment discontinuation.

This type of bias has been well described in the methodological literature. Hernán, Hernández-Díaz, and Robins provided a seminal framework for understanding how conditioning on a post-treatment variable can introduce structural bias that distorts causal inferences[9]. The practical implication is clear: Even if CEA or CA19-9 levels do correlate with outcomes in the selected population, we cannot confidently generalize those findings to the broader population of postoperative GC patients-the very population in which clinicians most urgently need prognostic guidance.

TEMPORAL AMBIGUITY

The study measured CEA and CA19-9 at three time points: The first postoperative day, one month, and three months (after two cycles of chemotherapy), then averaged these values for analysis. Recurrence was assessed at one year. This design introduces a significant temporal ambiguity that undermines causal interpretation.

The central question is straightforward but difficult to answer without an appropriate study design: Do elevated biomarker levels at one to three months after surgery predict future recurrence, or do they merely reflect recurrence that has already begun but remains clinically undetectable. If biomarkers capture early subclinical recurrence rather than predict future events, their utility shifts from prognostication to surveillance. The latter is far less clinically transformative, particularly given the availability of cross-sectional imaging. The same temporal concern applies to the toxicity analysis.

The same temporal concern applies to the toxicity analysis. Worsening gastrointestinal symptoms during the first three months post-surgery occur within a complex milieu of surgical recovery, chemotherapy-induced mucosal injury, nutritional depletion, systemic inflammation, and-in some cases-early peritoneal or local recurrence. Attributing these symptoms primarily to biomarker levels without accounting for competing etiologies is a significant interpretive leap. Future studies should incorporate landmark analysis designs, in which patients are classified based on biomarker status at a fixed time point and followed prospectively from that landmark, to establish clearer temporal directionality[11,12].

CONFOUNDING

The study reported no baseline differences between recurrence and non-recurrence groups for age, sex, differentiation grade, TNM stage, or metastasis status. However, the analysis relied entirely on unadjusted group comparisons, overlooking a substantial number of clinically relevant confounders that could independently influence both biomarker levels and clinical outcomes.

The list of unmeasured or uncontrolled variables is extensive and clinically significant. Comorbidity burden, baseline liver function, and biliary obstruction are particularly relevant for CA19-9, which is known to be elevated in a range of benign hepatobiliary conditions[4]. Performance status, nutritional status, and systemic inflammatory markers such as C-reactive protein and albumin (or composite scores such as the Glasgow Prognostic Score or neutrophil-to-lymphocyte ratio) have well-established prognostic significance in GC[13,14]. Detailed tumor characteristics-including T-stage depth, lymph node count and ratio, lymphovascular invasion, perineural invasion, and surgical margin details even within R0 resections-provide critical prognostic information that aggregate TNM staging does not capture[15].

Equally important are treatment-related variables that were not accounted for: Dose delays, cycle timing, antiemetic regimens, and supportive care interventions, all of which directly influence both toxicity outcomes and potentially biomarker kinetics. Molecular subtype information-human epidermal growth factor receptor 2 (HER2) status, microsatellite instability (MSI), Epstein-Barr virus (EBV) status-is increasingly recognized as a fundamental determinant of prognosis and treatment response in GC[16-18], yet it was entirely absent from the analysis.

A particularly important but overlooked biological issue concerns the Lewis antigen phenotype. Approximately 5% to 10% of the population are Lewis antigen-negative and cannot produce CA19-9, leading to falsely low or undetectable levels even in the presence of aggressive disease[4]. Without accounting for this phenotype, the study's interpretation of CA19-9 levels is fundamentally compromised: Some patients classified as “low CA19-9” may in fact have advanced disease that simply cannot be detected by this marker.

Even with the study’s modest sample size, modern statistical approaches such as Cox proportional hazards models with a parsimonious set of covariates, penalized regression, or propensity score methods could have provided at least partial control for confounding[19]. Without such adjustments, the reported correlations between biomarkers and outcomes likely reflect underlying disease severity and host factors as much as-or more than-the independent prognostic value of CEA or CA19-9.

STATISTICAL METHODOLOGY: SURVIVAL ANALYSIS AND HYPOTHESIS TESTING

The study’s statistical shortcomings are not merely technical quibbles; they constitute fundamental violations of established analytical principles that directly affect the validity and interpretability of the findings (Table 1).

Table 1 Summary of key methodological concerns and recommended improvements.

Methodological issue	Concern in Qi and Wu’s study[8]	Recommended approach
Selection bias	Only patients completing 6 cycles with good compliance were included	Intention-to-treat analysis; sensitivity analyses including non-completers
Temporal ambiguity	Biomarkers measured 1-3 months postoperatively; recurrence at 1 year	Landmark analysis; trajectory modeling; joint longitudinal-survival models
Confounding	Unadjusted comparisons; key covariates not measured or controlled	Multivariable Cox models; penalized regression; propensity score adjustment
Survival analysis	RFS reported as mean without censoring; no KM curves or Cox models	KM curves, log-rank tests, Cox regression; hazard ratios with CIs
Multiple testing	Multiple comparisons without correction; no primary endpoint	Prespecified primary endpoint; Bonferroni/FDR correction
Predictive performance	No AUC, calibration, or incremental value assessment	ROC/AUC; calibration plots; decision curve analysis; TRIPOD adherence
Molecular context	No molecular subtyping; Lewis antigen status not assessed	Incorporate MSI, HER2, EBV status; assess Lewis antigen phenotype

RFS: Recurrence-free survival; KM: Kaplan-Meier; CI: Confidence interval; AUC: Area under the curve; ROC: Receiver operating characteristic; FDR: False discovery rate; MSI: Microsatellite instability; HER2: Human epidermal growth factor receptor 2; EBV: Epstein-Barr virus.

Mishandling of time-to-event data

RFS is the quintessential time-to-event endpoint, requiring methods that properly account for right censoring, variable follow-up durations, and the differential timing of events[11,14]. Presenting RFS as a simple mean and comparing groups without accounting for censored observations is a fundamental analytical error. This approach discards information from patients who had not yet experienced recurrence at the time of analysis, biases survival estimates in unpredictable directions, and precludes the calculation of essential clinical statistics such as median RFS, survival rates at specified time points, and hazard ratios. The omission of Kaplan-Meier curves, log-rank tests, and Cox proportional hazards models is particularly striking, as these are universally recognized as the standard analytical toolkit for survival data in oncology[11,12].

Inappropriate parametric assumptions

Both CEA and CA19-9 levels are characteristically right-skewed in oncology populations, with wide variance and frequent outliers, especially among patients with recurrence. The study reported means and standard deviations and used t-tests-all of which assume approximately normal distributions and equal variances across groups. In a small recurrence group (n = 18), these assumptions are almost certainly violated, increasing the risk of both type I and type II errors[20]. More appropriate alternatives include log-transforming biomarker data, nonparametric tests such as the Mann-Whitney U or Kruskal-Wallis test, and bootstrapped confidence intervals.

Multiple comparisons without correction

The study performed numerous statistical tests across multiple biomarker time points, recurrence strata, toxicity grades, and correlation analyses, without specifying a primary endpoint or applying multiplicity corrections. Each additional comparison inflates the cumulative probability of obtaining a false-positive result. In exploratory analyses involving small datasets, this risk is particularly acute. Established correction methods-Bonferroni, Holm-Sidak, or false discovery rate approaches-should have been employed, and the absence of a clearly defined primary endpoint is a significant design limitation.

Oversimplification of toxicity data

Recording only the highest toxicity grade across six treatment cycles collapses a complex, longitudinal phenomenon into a single ordinal value. This approach obscures the timing, duration, recurrence, and trajectory of adverse events. Ordinal statistical methods-such as ordinal logistic regression or nonparametric trend tests-would have better utilized the ordered nature of toxicity grades and provided more clinically meaningful insights into dose-toxicity relationships.

PREDICTIVE PERFORMANCE AND CLINICAL UTILITY

Demonstrating a statistical association between a biomarker and an outcome is, at best, the first step in a long translational journey. Clinical utility-the ability of a biomarker to improve risk prediction, alter management decisions, and ultimately improve patient outcomes-is what matters at the bedside (Table 2)[21,22]. The study by Qi and Wu[8] did not attempt to evaluate predictive performance.

Table 2 Emerging biomarker technologies for postoperative gastric cancer surveillance.

Biomarker	Advantages	Limitations	Clinical readiness
CEA/CA19-9	Inexpensive, widely available, easily serialized	Low sensitivity/specificity; affected by non-malignant conditions	Established in guidelines but not validated as a standalone prognostic tool
Circulating tumor DNA (ctDNA)	Tumor-specific; detects MRD; identifies actionable mutations	Cost, assay standardization; CHIP interference; limited early-stage sensitivity	Investigational; multiple ongoing trials in GC
Circulating tumor cells	Captures metastatic phenotype; whole-cell analysis	Technical isolation challenges; low counts in early disease	Experimental; limited GC validation
Extracellular vesicles	Rich molecular cargo; immunomodulatory insights	Isolation variability; lack of standardized protocols	Pre-clinical/early translational
Methylated cfDNA	Epigenetic signal: High sensitivity in some assays	Assay complexity; limited prospective data	Investigational; promising early results
Multi-analyte panels (ctDNA + protein + fragmentomics)	Combines complementary signals; highest accuracy potential	Complexity; cost; validation needed	Active clinical trials; not yet guideline-endorsed

CEA: Carcinoembryonic antigen; CA19-9: Carbohydrate antigen 19-9; MRD: Minimal residual disease; CHIP: Clonal hematopoiesis of indeterminate potential; GC: Gastric cancer; cfDNA: Cell-free DNA.

Specifically, the study did not assess discrimination [the ability of the biomarkers to distinguish between patients who will and will not experience recurrence, typically quantified by the area under the receiver operating characteristic curve or time-dependent area under the curve (AUC) for survival endpoints], calibration (the agreement between predicted and observed risk), or incremental value (whether adding CEA or CA19-9 to standard clinicopathologic variables improves prediction beyond what those variables alone provide)[23,24]. Without these assessments, we have no way of knowing whether the observed statistical associations translate into clinically meaningful predictive improvement.

The TRIPOD guidelines provide a well-established framework for the development, reporting, and validation of prognostic models[24,25]. Adherence to these guidelines would have substantially strengthened the study’s translational relevance. Furthermore, decision curve analysis-a method specifically designed to evaluate the net clinical benefit of prediction models across a range of threshold probabilities-should be considered essential for any study claiming to identify clinically useful biomarkers[21].

An additional missed opportunity concerns the use of biomarker trajectories rather than simple averages. In clinical practice, the trajectory of CEA or CA19-9 over time-whether levels are rising, falling, or plateauing-often provides more actionable information than any single measurement or average. Persistent elevation, failure to normalize, or accelerating increases during the postoperative period may carry very different prognostic implications than a transiently elevated value that subsequently declines. By reducing complex temporal data to a single average, the study may have obscured precisely the patterns that could most usefully guide clinical decision-making.

MOLECULAR SUBTYPES, LIQUID BIOPSY, AND THE FUTURE OF PROGNOSTIC BIOMARKERS IN GC

It is worth stepping back to consider the study by Qi and Wu[8] within the broader context of where the field is heading. The reliance on traditional serum tumor markers like CEA and CA19-9, while understandable given their accessibility, is increasingly out of step with the molecular revolution transforming GC management.

Molecular subtyping

The Cancer Genome Atlas and the Asian Cancer Research Group classifications have identified distinct molecular subtypes of GC-EBV-positive, MSI-high, genomically stable, and chromosomally unstable-that carry fundamentally different prognostic and therapeutic implications[16,17]. MSI-high tumors, for example, have a favorable prognosis and respond dramatically to immune checkpoint inhibitors[18,26], while genomically stable tumors enriched for diffuse histology carry a worse prognosis and limited therapeutic options. Any prognostic biomarker study that ignores this molecular heterogeneity risks producing results that are uninterpretable at the patient level.

Circulating tumor DNA and liquid biopsy

The most transformative development in postoperative surveillance may be the emergence of circulating tumor DNA (ctDNA) as a biomarker for minimal residual disease (MRD). Several studies have now demonstrated that detectable ctDNA after curative-intent gastrectomy is a potent and specific predictor of subsequent recurrence, often detecting relapse months before conventional imaging[27-31]. The CRITICS trial showed that ctDNA positivity during perioperative treatment provided independent prognostic information beyond traditional tumor markers[31]. Unlike CEA and CA19-9, ctDNA is tumor-specific, can identify actionable mutations, and can provide real-time information on clonal evolution and treatment resistance[28,32].

This is not to suggest that ctDNA-based approaches are without limitations. Assay standardization, cost, turnaround time, sensitivity in early-stage disease, and the challenge of distinguishing tumor-derived mutations from clonal hematopoiesis of indeterminate potential all remain active areas of investigation[27]. However, the trajectory is clear: The field is moving toward molecular, tumor-specific biomarkers that can be integrated with clinical and pathologic data to create individualized risk profiles.

Immunotherapy and predictive biomarkers

The incorporation of immune checkpoint inhibitors into the treatment of advanced and perioperative GC[26] has created an additional need for biomarkers that can predict immunotherapy benefit. PD-L1 expression, MSI status, tumor mutational burden, and EBV positivity have all shown predictive value in this context. Future biomarker studies must account for the rapidly evolving treatment landscape and the interactions between biomarker profiles and therapeutic modalities.

RECOMMENDATIONS FOR FUTURE RESEARCH

Based on the methodological concerns outlined above and the evolving state of the field, I propose the following recommendations for investigators seeking to advance biomarker-based prognostication in postoperative GC.

First, study designs must minimize selection bias. Intention-to-treat analyses that include all patients who begin adjuvant therapy-regardless of compliance-should be the default. Where compliance-based subgroup analyses are performed, they should be clearly identified as such and interpreted with appropriate caveats.

Second, temporal relationships must be carefully established. Landmark analysis designs, in which biomarker status is assessed at a prespecified time point, and patients are followed prospectively from that landmark, are essential for separating prediction from detection. Serial biomarker measurements should be analyzed as trajectories rather than averages, using methods such as joint longitudinal-survival models.

Third, multivariable adjustment is non-negotiable, even in modest-sized studies. Cox proportional hazards models with a limited set of well-chosen covariates, or penalized regression methods designed for small samples, should replace unadjusted group comparisons. Confounders should be selected a priori based on established clinical and biological knowledge.

Fourth, survival analysis must be performed correctly. Kaplan-Meier methods, log-rank tests, and Cox models are the absolute minimum for time-to-event data. Results should be reported as median survival with confidence intervals, hazard ratios, and survival curves-not as simple means.

Fifth, predictive performance must be formally evaluated. Discrimination (AUC, time-dependent AUC), calibration, and-critically-incremental value over standard clinicopathologic variables should be reported. Decision curve analysis should be performed to establish clinical utility at actionable risk thresholds.

Sixth, the molecular context must be incorporated. Future studies should, at a minimum, assess and adjust for MSI status, HER2 expression, and EBV status. Where feasible, integration of ctDNA-based MRD detection with traditional serum markers may offer the most comprehensive and actionable prognostic framework.

Seventh, external validation is essential. No biomarker model should be considered clinically ready without validation in an independent cohort, ideally from a different institution or geographic population. The TRIPOD guidelines provide a clear roadmap for this process.

CONCLUSION

The study by Qi and Wu[8] addresses a clinically important question and provides hypothesis-generating data suggesting that postoperative CEA and CA19-9 levels may correlate with early recurrence and gastrointestinal toxicity in patients receiving adjuvant SOX chemotherapy. These findings underscore the potential of serum biomarkers for risk stratification and surveillance and warrant further investigation.

However, the study’s interpretability and clinical applicability are substantially limited by the methodological and statistical issues detailed in this review. The concerns I have raised-selection bias, temporal ambiguity, inadequate confounding control, fundamentally flawed survival analysis, and absent predictive performance evaluation-are not unique to this one study. They represent recurring weaknesses across the biomarker literature in gastrointestinal oncology, and they explain why, despite decades of research, no single serum biomarker has achieved reliable, validated clinical utility as a standalone prognostic tool in postoperative GC.

The path forward requires a fundamental shift in how we design, conduct, and evaluate biomarker studies. We must move beyond simple correlative analyses toward rigorously designed, appropriately powered, prospectively validated studies that integrate modern molecular insights, employ state-of-the-art biostatistical methods, and demonstrate not just statistical significance but genuine clinical utility. Only then will we be able to deliver on the promise of personalized, biomarker-guided postoperative management for patients with GC.