Back to Evidence

Statistical Review of TACE Plus HAIC Evidence for Clinicians | QSevidence 2026

Evidence-Based Medicine18 min read

Evaluating combined interventional treatment for liver cancer requires a connected account of treatment choice, patient differences and statistical reporting. QSevidence can support clinical and research teams in organizing questions, tracing sources and comparing outcomes for professional review.

Statistical Review of TACE Plus HAIC Evidence for Clinicians | QSevidence 2026

Establish the comparison and verification level

The question here concerns transarterial chemoembolization (TACE) plus hepatic arterial infusion chemotherapy (HAIC) versus TACE alone. Combination treatment, HAIC alone and other interventional or systemic regimens represent different comparisons. Reviewers need the actual protocol, population and publication identity before interpreting a treatment label.

The supplied research material offers an audit plan but lacks independently confirmed full text, complete bibliographic details and patient-level data for its target paper. Some numerical results are explicitly simulated. This article therefore makes no finding of error against named authors and reports no simulated survival, hazard ratio or toxicity rate as clinical evidence. It is a review framework, not a completed reanalysis or systematic review.

Understand population and treatment selection

Intermediate or advanced liver cancer is an insufficient eligibility definition. Extract staging criteria, liver function, performance status, tumor burden, vascular invasion, extrahepatic spread, AFP and previous treatment. Child-Pugh classification and BCLC stage describe different dimensions and need separate fields. A shared disease label does not establish comparable populations.

When clinicians select treatment, patients able to tolerate combination therapy may differ prognostically. Baseline descriptions should retain clinical differences and missingness. Standardized differences can describe balance, but neither those measures nor nonsignificant baseline P values demonstrate absence of confounding.

Regression, matching and propensity score methods require a target population, justified pretreatment variables, overlap assessment and post-analysis balance review. Selecting confounders solely by univariable P values is inadequate. Post-treatment changes should not casually become baseline adjustment variables. Residual and unmeasured confounding remain possible after adjustment.

Align time zero with treatment assignment

Check eligibility, first TACE, HAIC initiation and follow-up origin. Classifying patients by later HAIC completion while measuring survival from earlier TACE can introduce a survival requirement for entering the combination group. Map the sequence before considering time-dependent analysis or another appropriate design; adding a routine covariate may not resolve the problem.

Subsequent targeted therapy, immunotherapy and further interventions also need timestamps. Their effects should not automatically be attributed to the starting combination. Specify whether the question concerns an initial treatment strategy or an association within a later treatment phase, and retain that interpretation throughout the report.

Keep survival, response and safety definitions distinct

Overall survival, progression-free survival and time to progression are not interchangeable. Preserve event definitions, assessment schedules, death handling, censoring rules and follow-up completeness. Numbers at risk help readers judge how much information supports a curve's tail.

A Cox hazard ratio is not a fixed-time mortality probability ratio and cannot be calculated by dividing median survival times. Assess proportional hazards. If inappropriate, consider a question-led alternative such as time-varying effects or restricted mean survival time with a specified horizon. Do not replace a primary analysis simply because another looks favorable.

Response rates need assessment criteria, observation windows, evaluable populations and raw counts. RECIST and mRECIST cannot be silently mixed. Odds ratios, risk ratios and risk differences require accurate labels. Excluding people without follow-up imaging changes the denominator and can alter interpretation.

Safety comparisons need grading rules, observation windows, assessed participants, liver deterioration and treatment discontinuation. Patients with at least one adverse event and total event counts are different units. Report efficacy alongside harm; one significant survival result does not establish a superior overall benefit–risk balance.

Review continuous and repeated outcomes

AFP, tumor diameter and liver enzymes may be skewed, contain outliers and change during treatment. Match the method to the estimand, distribution and sample. Failure to report a normality test does not automatically invalidate a t test; rank tests are not universally tests of medians. Robust methods or transformations may be appropriate with explicit interpretation.

Measurements at one, three and six months are proposed follow-up examples here, not verified visits in the target study. Repeated observations within a patient are correlated. Separate independent tests at every visit can neglect correlation and multiplicity. Longitudinal models need an explained time structure, treatment–time interaction and missing-data assumptions.

Values unavailable after progression or death should not be filled with healthy measurements or treated as ordinary omissions. Distinguish unperformed tests, loss to follow-up, withdrawal, death and assay failure. Compare reasons between groups and plan sensitivity analyses appropriate to the question. Imputation must not erase clinical events.

Interpret effects rather than significance alone

Consider effect size, confidence intervals, clinical context and whether analyses were prespecified. A confidence interval expresses estimation uncertainty; it is not a 95% probability statement about a fixed parameter lying in the calculated interval. Nonsignificance does not establish equivalence, and significance does not establish clinical importance. No universal survival-benefit threshold is proposed here.

Multiple endpoints, visits and subgroups require defined primary questions, comparison families and an explained multiplicity strategy. Significance in one subgroup and not another does not itself demonstrate interaction. Sparse categorical data need methods suited to their structure. A contingency-table analysis of ordered categories is not automatically erroneous, although it may discard ordering information.

A retrospective cohort can explain available sample size and resulting precision. Observed power does not repair its design. Future sample planning should use the primary endpoint, target effect, expected events and feasibility. Simulated examples must remain separate from clinical conclusions.

Build a source-linked review record

SAMPL supports basic statistical reporting, while STROBE supports transparent observational reporting. Neither certifies correct design or treatment effectiveness. A useful record separates the original passage, question, required evidence and expert assessment, distinguishing unreported information from unverifiable claims and demonstrated problems.

An Evidence-based medical AI workflow can support search questions, comparison matrices, endpoint dictionaries and source links. QSevidence outputs require clinical and statistical review against original sources. No product accuracy, time saving or patient benefit was measured here. Obtain the full publication and update verified facts before deciding whether data recalculation is possible.

Frequently asked questions

How can QSevidence support comparisons of TACE plus HAIC studies?

It can help organize eligibility, regimens, treatment timing and endpoint definitions with original-source tracing. Professionals then decide whether studies address the same question and how their findings can be compared.

How can QSevidence assist statistical review?

It can support checklists covering effect measures, denominators, censoring, repeated observations and missingness. Statistical reviewers use complete publications and available data to determine the nature of concerns and document traceable revisions.

How should teams evaluate AI-assisted review?

Start with source traceability, issue identification and expert corrections. Appropriately governed research can assess practical effects, reporting review-process measures separately from patient benefit or treatment value.

Sources and verification scope

  1. Original SAMPL reporting guidance: statistical methods, effect uncertainty and survival reporting sections reviewed.
  2. Official STROBE checklists: linked combined checklist PDF reviewed for observational design, bias, missingness, follow-up and reporting items.

Research notice: Sources were checked on October 29, 2026. This framework provides no individual treatment selection or interventional procedure instructions and claims no recalculation of the target study. Qualified teams interpret complete evidence alongside patient circumstances.