Inflammation Blood Test Joins Staging to Predict Lung Cancer Survival

Lung cancer remains one of the most lethal malignancies worldwide, and clinicians have long relied on a familiar toolkit to estimate how long a patient might live: age, smoking history, and the tumor’s stage under the American Joint Committee on Cancer (AJCC) system. These clinicopathological factors are indispensable, but they describe the tumor and the patient’s exposure history rather than the biology of the host. A new study published in BMC Cancer argues that a missing piece has been sitting in routine blood work all along: the state of systemic inflammation, measurable from standard complete blood count parameters. The research team, led by Lingge Yang, Wei Liu, and Quan Xie with corresponding authors Qingping Chen and Duanlin Du, built and stress-tested a prognostic model that fuses conventional clinical features with an aggregate index of systemic inflammation, known as the AISI, to predict overall survival in lung cancer patients.

The aggregate index of systemic inflammation is a composite derived from peripheral blood counts, combining markers such as neutrophils, lymphocytes, platelets, and monocytes into a single figure that captures the body’s inflammatory tone. Systemic inflammation has attracted growing attention in oncology because tumor-associated inflammation shapes immune surveillance, treatment tolerance, and disease progression. Yet most prognostic tools in lung cancer either ignore these host-related signals or treat them as afterthoughts. The study’s central question was straightforward: if you add this inflammatory signature to the variables clinicians already collect, do you get a model that stratifies patients reproducibly, or does it merely add statistical noise? Answering that question rigorously required an unusually disciplined validation design, which is where the paper distinguishes itself from many prognostic modeling exercises.

The development cohort consisted of 272 patients with pathologically confirmed lung cancer treated at a single center, among whom 60 deaths were recorded. The investigators first screened candidate variables through a multivariable Cox proportional hazards model, the workhorse of survival analysis, and retained four predictors in the final model: age, smoking history, AJCC stage, and the AISI. On top of this, they constructed a random survival forest, a machine-learning ensemble that grows many survival trees on bootstrap samples and aggregates their predictions. The head-to-head comparison between classical regression and the random forest is one of the study’s most instructive contributions, because machine-learning methods are frequently assumed to outperform traditional models whenever enough data exist.

The internal validation results deliver a sobering lesson in statistical honesty. The model’s apparent concordance index, or C-index, which measures how well the model ranks patients by risk, stood at an impressive 0.776 before correction. But when the authors applied full-pipeline bootstrap optimism correction, a procedure that re-runs the entire modeling process, including variable selection, inside each bootstrap sample, the C-index fell to a corrected 0.698. The five-year area under the receiver operating characteristic curve dropped in parallel from 0.871 to 0.816. This gap between apparent and corrected performance quantifies optimism bias, the tendency of any model to look better on the data that shaped it. Many published prognostic models report only the flattering apparent figures; this team reported both, and the corrected values still indicate useful, if modest, discrimination.

The machine-learning comparison produced an equally instructive null result. Under repeated cross-validation, the random survival forest achieved a C-index of 0.757 against 0.761 for Cox regression, and none of three additional machine-learning comparators exceeded the classical model on the same variables. In other words, with a few hundred patients and a handful of well-chosen predictors, the flexible forest conferred no advantage over a transparent regression equation. This finding echoes a broader theme in clinical prediction research: when sample sizes are modest and predictors are few, simpler models are often as accurate, far easier to implement at the bedside, and less prone to overfitting. The authors did not discard the random forest entirely, but they treated it as an exploratory comparator rather than the headline result.

The external evaluation is the study’s boldest element. Rather than refitting the model on data from a second institution, the team froze the Cox model exactly as developed, with its coefficients, its baseline survival function, and a risk cutoff derived from the development cohort, and applied it unchanged to 74 patients, with 28 deaths, at a second hospital. This frozen-model approach is a far harsher test than the common practice of recalibrating on the validation set, because it assumes the development estimates transport across centers without adjustment. The frozen model achieved an external C-index of 0.823, with a 95 percent confidence interval of 0.739 to 0.897, along with one-year and three-year AUCs of 0.798 and 0.868. A calibration slope of 1.173 was statistically compatible with the ideal value of 1, though the confidence interval, spanning 0.752 to 1.593, was wide.

Discrimination, however, tells only half the story. When the researchers examined absolute survival predictions, they found the model systematically overestimated survival, by 0.081 at one year and 0.173 at three years. In practical terms, a patient predicted to have a given probability of surviving three years would, on average, fare worse than promised. The authors also noted that the strikingly high external discrimination partly reflected case mix: after standardization, the difference between cohorts narrowed from a C-index gap of 0.773 to 0.801 in standardized terms. Meanwhile, the development-derived risk cutoff, applied without modification, separated the external cohort into groups with dramatically different outcomes, a hazard ratio of 42.48 with a confidence interval of 5.73 to 314.71. That enormous ratio reflects both genuine risk separation and the small number of events in the low-risk group, and it should be read as evidence of stratification rather than a precise effect estimate.

The study’s limitations are candid and consequential. Both cohorts were small and retrospective, drawn from single centers in China, and the external validation set contained only 74 patients and 28 deaths, which is why several confidence intervals stretch wide. The overestimation of absolute survival means the model, in its current form, cannot safely guide individual counseling or treatment intensity decisions. The authors are explicit that recalibration on larger datasets and multicenter prospective confirmation are prerequisites before any clinical deployment. They also caution that the reconstructed random survival forest, evaluated externally as a sensitivity analysis, produced concordant but exploratory findings rather than validated performance.

What survives all this scrutiny is a genuinely reproducible signal: integrating host inflammatory status, captured by a routine and inexpensive blood-derived index, with conventional staging yields risk stratification that held up under bootstrap correction, cross-validation, and a frozen external test. The paper’s methodological transparency, reporting apparent and corrected metrics side by side, freezing the model before external evaluation, and quantifying calibration errors, offers a template for how prognostic research should be conducted and reported. For a field crowded with overfit prediction models that evaporate outside their birth cohort, a modest but transportable C-index anchored in routine laboratory data is a meaningful step. The next test will be whether the inflammation-augmented model, once recalibrated, can earn a place in the clinic alongside the staging systems it aims to sharpen.

Subject of Research: A prognostic model combining systemic inflammatory markers with clinical features for overall survival in lung cancer

Article Title: A prognostic model integrating systemic inflammation and clinical features for overall survival in lung cancer: development, internal validation, and external evaluation

Article References: Yang, L., Liu, W., Xie, Q., Wang, X., Li, S., Yi, L., Xu, Q., Peng, J., Yang, Z., Peng, X., Fu, S., Xie, P., Dai, W., Peng, B., Chen, Q., & Du, D. (2026). A prognostic model integrating systemic inflammation and clinical features for overall survival in lung cancer: development, internal validation, and external evaluation. BMC Cancer. https://doi.org/10.1186/s12885-026-16809-2

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16809-2

Keywords: lung cancer, prognostic model, systemic inflammation, AISI, Cox regression, random survival forest, external validation, overall survival, risk stratification, machine learning, calibration, BMC Cancer