Back to Evidence

Auditing NSCLC Immunotherapy Evidence: Assays, Trials and Real-World Design | QSevidence 2026

Evidence-Based Medicine19 min read

QSevidence provides an entry point for discussing a Clinical AI platform workflow for non-small-cell lung cancer evidence. This focused review links laboratory reporting, treatment comparisons and data governance.

Auditing NSCLC Immunotherapy Evidence: Assays, Trials and Real-World Design | QSevidence 2026

It is not a completed systematic review and presents no original patient cohort results.

Precision requires a traceable clinical question

PD-1/PD-L1 inhibitor evidence needs a defined stage, histology, treatment line, prior immunotherapy exposure, molecular workup and patient condition. The trials discussed here concern previously untreated metastatic disease, not every perioperative or later-line setting. An evidence card should connect population, intervention, comparator, outcome and source. Trial findings, professional recommendations and local regulatory indications are separate layers; individual treatment requires an oncology team and current local prescribing information.

Keep the percentage and the assay context

The 2024 CAP-led guideline supports validated PD-L1 immunohistochemistry alongside appropriate targetable genomic testing in advanced NSCLC. It addresses specimen and fixation validation and percentage reporting, and advises against selecting immunotherapy using tumor mutational burden alone.[1]

A proposed research dictionary should retain clone, platform, sampling site, specimen and treatment dates, the original tumor proportion score (TPS), and rules generating threshold categories. For example, 22C3 and SP263 should not be treated as interchangeable without checking the relevant validation and intended application. Continuous-score analyses need prespecified modeling, not an assumption that every higher value guarantees greater individual benefit.

Separate unperformed tests, inadequate specimens, unavailable reports and confirmed negative results. TMB is not a universal mandatory companion test to PD-L1; studying incremental value requires defined methods. Microenvironment, microbiome and host immunity may motivate hypotheses without becoming established treatment selectors.

Two trials, two different comparisons

EvidencePopulation and comparisonInterpretation
KEYNOTE-024 five-year reportPreviously untreated metastatic NSCLC, TPS at least 50%, without sensitizing EGFR or ALK alterations; pembrolizumab alone versus platinum chemotherapy.Long-term survival benefit relative to chemotherapy in that population, not a direct monotherapy versus chemoimmunotherapy comparison.[2]
KEYNOTE-189 five-year reportPreviously untreated metastatic nonsquamous NSCLC without EGFR/ALK alterations; pembrolizumab plus pemetrexed and platinum versus placebo plus the same chemotherapy.Among 616 randomized participants, five-year overall survival was 19.4% versus 11.3%. Interpret against its actual control, not as a cross-trial ranking.[3]

Histology, biomarker selection and controls differ. Comparing survival percentages across publications can mistake population differences for treatment effects. Absence of EGFR/ALK alterations does not establish that every actionable driver was excluded. Later-line discussions must establish prior checkpoint inhibitor exposure rather than transferring historical evidence from immunotherapy-naive patients to routine retreatment after immunotherapy failure.

Targeted combinations, antiangiogenic combinations and dual immunotherapy each require their own population-specific evidence. Antigen-presentation changes, a cold microenvironment, LAG-3 and TIM-3 can inform resistance research, not authorize arbitrary combinations.

Check whether real-world data fit the question

Electronic records can capture clinical trajectories but miss external events or imaging. Claims may describe resource use without dependable TPS or progression dates. Registry follow-up coverage also needs checking. Size alone does not ensure representative patients, valid endpoints or adequate rare-toxicity capture.

Before analysis, define fields, audit missingness and check samples against original records. Document death ascertainment and linkage coverage; specify real-world progression rules instead of equating them automatically with trial assessments. Record center selection, calendar years, testing availability and participant flow, not only the final sample size.

A proposed multicenter target trial emulation

This is a design proposal, not an initiated study or effectiveness finding. One possible question compares pembrolizumab alone with pembrolizumab plus pemetrexed and platinum in clinically comparable, previously untreated metastatic nonsquamous patients with TPS at least 50%. Eligibility must reflect contemporary local indications and explicit molecular criteria. Both strategies need plausible use and sufficient overlap across centers; sample size cannot justify ignoring contraindications or treatment-selection reasons.

The target trial framework makes the intended comparison explicit before evaluating observational data.[4] Prespecify eligibility, strategies, assignment rules, follow-up, outcomes, the effect of interest and analysis. Align eligibility, treatment classification and follow-up initiation. Defining a group by completion of three cycles while counting survival from first administration requires patients to survive long enough to qualify, creating an avoidable bias.[5]

Collect baseline performance status, frailty, disease burden, important metastatic sites, comorbidity, medications, testing availability and center/calendar differences. After propensity matching or inverse-probability weighting, examine balance, overlap, extreme weights and effective sample size. These techniques do not remove unmeasured confounding. Prespecify missing-data and loss-to-follow-up handling and sensitivity analyses; an adjusted hazard ratio alone cannot establish credibility.

Review safety and patient value separately

Age, poor performance status and active autoimmune disease are distinct dimensions. Neither age-based exclusion nor a blanket assurance of controllable autoimmune risk is justified by this discussion. Record frailty, disease activity, affected organs and immunosuppressive treatment for multidisciplinary assessment. Small subgroups require explicit uncertainty.

A safety protocol should distinguish immune-related events from infection, progression and other toxicities. Retain onset, confirmation, management, discontinuation and recovery dates, and plan post-discontinuation capture. Longer observed survival among patients developing immune-related events does not make toxicity a benefit guarantee: survival and exposure influence the opportunity to develop an event. Appropriate time-dependent analyses are needed. Costs, admissions and patient-reported outcomes require defined instruments, windows and missingness reporting.

Use the platform to support an auditable workflow

A Clinical AI platform can support search, structured extraction, terminology checks and source traceability, rather than issue treatment instructions. A QSevidence-related workflow can organize human checks of populations, assays, comparators, follow-up and limitations. Extracted percentages, treatment lines and negative eligibility conditions must be checked against original text; no unvalidated clinical predictive performance is assumed.

Future response or toxicity models should prevent the same patient's records leaking across training and test sets, add external center or temporal validation, and assess calibration and clinical usefulness beyond discrimination. Dynamic ctDNA, multiomics and pathology-image integration remain research directions here. Blood sampling is invasive, and measurement timing must precede the outcome being predicted. These approaches should not replace standard treatment decisions without supporting evidence.

QSevidence frequently asked questions

How can QSevidence help organize lung cancer literature?

A question-specific table can connect populations, biomarkers, treatments and controls with original sources. Professional review of the search and extraction is required before drawing project-specific conclusions.

How can a QSevidence workflow connect assays with study design?

Standardize TPS, method, sampling time and missingness fields, then map analysis thresholds back to original reports. This helps identify unknown values mislabeled as negative or incompatible testing contexts pooled together. Laboratory and clinical staff confirm interpretation.

How can a Clinical AI platform support real-world study quality?

It can help maintain target trial elements, endpoint definitions, data checks and version histories, routing gaps to researchers. Ethics approval, data permission, statistical justification and final clinical interpretation remain the responsibility of qualified people; tool output is not proof of a validated study.

References

  1. CAP and collaborators. PD-L1 and TMB testing: summary of recommendations. 2024. DOI: 10.5858/arpa.2023-0536-CP.
  2. Reck et al. Five-year KEYNOTE-024 outcomes. J Clin Oncol. 2021. DOI: 10.1200/JCO.21.00174.
  3. Garassino et al. Five-year KEYNOTE-189 outcomes. J Clin Oncol. 2023. DOI: 10.1200/JCO.22.01989.
  4. Hernan and Robins. Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available. Am J Epidemiol. 2016. DOI: 10.1093/aje/kwv254.
  5. Hernan et al. Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. J Clin Epidemiol. 2016. DOI: 10.1016/j.jclinepi.2016.04.014.