A sweeping review of prediction tools designed to flag which lung cancer patients will develop pneumonia after surgery has reached an uncomfortable conclusion: the models look impressively accurate on paper, yet virtually every one of them was built in a way that could inflate that accuracy. The systematic review and meta-analysis, published in Frontiers in Oncology, pulled together 20 studies describing 26 risk prediction models for postoperative pneumonia, a complication that can strike as many as one in four patients after lung cancer surgery and that lengthens hospital stays, drives up costs and raises perioperative mortality.
The review team searched nine databases, including CNKI, Wanfang, VIP, SinoMed, PubMed, Embase, the Cochrane Library, CINAHL and Web of Science, covering the literature from inception to July 20, 2026. After screening 3,044 records, they included 20 studies, all conducted in China and published between 2019 and 2025. Seventeen were retrospective, two were prospective cohorts, including one multicenter study, and the combined datasets spanned 18,558 lung cancer patients, with individual study sizes ranging from 72 to 5,269 participants. Reported rates of postoperative pneumonia ranged from 1.7 percent to 27.8 percent depending on the cohort.
Most of the models were built with logistic regression, the workhorse statistical technique for estimating how a set of patient characteristics translates into the probability of an outcome. Two studies instead turned to machine-learning approaches. Across the 20 studies, the researchers identified 40 final predictive factors. The most frequently used were age, appearing in twelve studies; operation time, in ten; smoking, in nine; diabetes mellitus, in seven; preoperative pulmonary function, in six; chronic obstructive pulmonary disease, in five; and TNM cancer stage, also in five. A long tail of other variables, including extent of excision, intraoperative blood loss, ASA grade, mechanical ventilation time, C-reactive protein, albumin, white blood cell count, body mass index, drainage retention time, the systemic immune-inflammation index, procalcitonin and HMGB1, each appeared in one to four studies.
On the surface, the performance numbers were striking. Reported area under the curve values, a standard measure of how well a model separates patients who develop the complication from those who do not, ranged from 0.670 to 0.963, spanning moderate to excellent discrimination. When the reviewers pooled the AUC values from the 21 validated models across 15 studies that met their meta-analysis criteria, using a random-effects model in STATA 17.0, the combined figure was 0.89, with a 95 percent confidence interval of 0.87 to 0.91.
But the pooled estimate came with heavy statistical caveats. Heterogeneity across studies was extreme, with an I-squared value of 92.8 percent and a p-value below 0.001, meaning the models differed so much in methods, predictors and populations that a single summary number conceals enormous variation. Egger’s test also flagged significant publication bias, with a statistic of 2.243 and p equal to 0.016, a sign that studies with flattering results may be overrepresented in the literature. A sensitivity analysis that dropped each model one at a time found the combined estimate essentially unchanged, which the authors read as evidence of robustness, yet subgroup analyses by validation type, internal versus internal plus external, still showed substantial heterogeneity of 92.9 percent and 85.0 percent respectively.
The deeper problem emerged when the reviewers applied PROBAST, the Prediction Model Risk of Bias Assessment Tool, which grades studies across participant, predictor, outcome and analysis domains. Every single one of the 20 studies was judged to carry a high risk of bias. In the analysis domain, 18 studies were high risk: eleven had too few outcome events to satisfy the conventional guideline of more than 20 events per variable, four converted continuous variables into categories, one excluded enrolled participants and handled missing data inadequately, eight selected variables based on univariate screening, eight did not thoroughly evaluate model performance, three failed to address overfitting or underfitting, and two did not disclose regression coefficients.
Bias risks also crept in through the predictors and the outcome definitions themselves. Four studies showed unclear risk in the predictor domain because the definitions of predictive factors overlapped with the definitions of the outcome indicators, and three more did not confirm that all predictors would be available at the time the model would actually be used. In the outcome domain, eight studies were high risk, four because outcome definitions did not exclude information on the predictors and four because the interval between predictor assessment and outcome determination was not appropriately ensured. These are not cosmetic flaws; when the thing being measured is entangled with the things doing the predicting, a model can appear far more accurate than it would be in genuine prospective use.
Applicability concerns compounded the picture. Sixteen of the 20 studies raised high applicability concerns, mainly because participants were drawn from narrow slices of the lung cancer population: six studies focused on particular histological types such as adenocarcinoma or squamous cell carcinoma, two targeted only elderly patients, and eleven restricted enrollment to patients undergoing thoracoscopic surgery. Four studies raised concerns about the timing of predictor measurements, and one study’s postoperative outcome evaluation window was judged too long.
The reviewers also noted structural limitations of their own synthesis. All included studies came from China, so the models may need adjustment before being applied in other countries and health systems. The meta-analysis merged only AUC values, not calibration, the measure of whether predicted probabilities match observed event rates, and the authors acknowledge this choice may itself inflate the pooled discrimination estimate.
The bottom line for clinicians and researchers is a call for restraint and rebuilding. The existing models show good discriminative performance in their own datasets, but that performance may be overestimated because of the pervasive high risk of bias, and their clinical applicability is limited by narrow participant selection and weak validation. The authors recommend that future work adhere to the TRIPOD reporting guidelines, use the PROBAST checklist during design, build models on larger samples with rigorous prospective designs, and subject them to multicenter external validation before any claim of clinical readiness. Until then, the apparent precision of these pneumonia risk scores should be read as a promise still awaiting proof rather than a tool to trust at the bedside.
Subject of Research: Systematic review and meta-analysis of risk prediction models for postoperative pneumonia after lung cancer surgery
Article Title: Risk prediction models for postoperative pneumonia in lung cancer patients: a systematic review and meta-analysis
Article References: Li, C., Jiang, Y., & Wang, Y. (2026). Risk prediction models for postoperative pneumonia in lung cancer patients: a systematic review and meta-analysis. Frontiers in Oncology, 16, 1967523. https://doi.org/10.3389/fonc.2026.1967523
Image Credits: AI Generated
DOI: 10.3389/fonc.2026.1967523
Keywords: lung cancer, postoperative pneumonia, risk prediction models, systematic review, meta-analysis, PROBAST, AUC, logistic regression, machine learning, risk of bias, external validation, thoracic surgery
News Source: Nathaniel Bowman. (October 30, 2026). Lung Cancer Pneumonia Risk Models Look Accurate but Almost All Are Flawed. Scienmag.
