A retrospective investigation evaluating routine preoperative blood tests for distal cholangiocarcinoma found that a five-variable laboratory model offered modest incremental predictive capability for detecting perineural invasion beyond baseline disease prevalence. Perineural invasion, characterized pathologically by cancer cells infiltrating the nerve sheath or surrounding neural structures, represents an aggressive biological feature frequently observed in distal bile duct malignancies. While surgical resection via pancreaticoduodenectomy remains the primary curative-intent intervention for resectable distal cholangiocarcinoma, postoperative recurrence remains common and long-term survival remains poor. Currently, perineural invasion is diagnosed almost exclusively through postoperative histopathological assessment of surgical specimens, which prevents clinicians from integrating this critical prognostic factor into preoperative risk stratification and surgical planning.
To investigate whether routinely measured circulating biomarkers could offer early insight into neural infiltration, researchers at Beijing Chao-Yang Hospital retrospectively evaluated 233 consecutive patients with histologically confirmed conventional distal cholangiocarcinoma who underwent curative-intent resection between January 2011 and June 2025. The operational definition of conventional distal cholangiocarcinoma specifically excluded adenosquamous carcinoma and other rare histological variants. Histological confirmation on surgical resection specimens served as the definitive reference standard, with perineural invasion independently diagnosed by two gastrointestinal pathologists blinded to the preoperative laboratory findings. Within the cohort, perineural invasion was documented in 206 patients, representing 88.4 percent of the sample, while only 27 patients, or 11.6 percent, were confirmed to be free of perineural invasion.
Candidate predictors were selected from an extensive panel of routinely measured preoperative indicators obtained primarily within one week prior to surgery. In individuals with multiple tests, the measurement closest to the operation date was analyzed. For patients who underwent preoperative biliary drainage, measurements of carbohydrate antigen 19-9, alkaline phosphatase, aspartate aminotransferase, alanine aminotransferase, and bilirubin fractions were standardized using values collected after drainage completion to minimize the confounding effects of obstructive jaundice. Candidate variables spanned serum tumor markers, including carbohydrate antigen 19-9, carcinoembryonic antigen, carbohydrate antigen 125, and alpha-fetoprotein, as well as nutritional, metabolic, coagulation, and hematological indices. Markedly right-skewed tumor markers were log-transformed prior to screening, while continuous predictors were Winsorized at the 1st and 99th percentiles to curb the influence of extreme values.
Predictor selection employed Elastic Net logistic regression with the mixing parameter set to 0.5 and the penalty parameter chosen via stratified cross-validation. To stabilize feature selection in a dataset characterized by high outcome imbalance, the authors implemented stability selection across 500 bootstrap resamples generated from the development cohort. Candidate markers retained in at least 70 percent of resamples were selected for inclusion in the final multivariable model. This stability threshold retained five specific laboratory variables: fibrinogen, mean platelet volume, mean corpuscular hemoglobin concentration, log-transformed carcinoembryonic antigen, and log-transformed carbohydrate antigen 19-9. Although alternative thresholds of 50 percent, 60 percent, and 80 percent were compared, retaining 13, 9, and 4 predictors respectively, the 70 percent cutoff was chosen after balancing predictive error and model complexity.
The multivariable prediction model was subsequently fitted using Firth penalized logistic regression to reduce small-sample estimation bias and mitigate numerical separation stemming from the small number of negative cases. Within the fitted multivariable framework, mean platelet volume demonstrated the sole statistically significant association with perineural invasion, exhibiting an odds ratio of 1.565 per one-unit increase (95 percent confidence interval, 1.072 to 2.382; P = 0.019). Fibrinogen demonstrated a positive but nonsignificant association, with an odds ratio of 1.381 per 100 mg/dL increase (95 percent confidence interval, 0.907 to 2.183; P = 0.135). Mean corpuscular hemoglobin concentration exhibited an inverse, borderline association, yielding an odds ratio of 0.680 per 10-unit increase (95 percent confidence interval, 0.453 to 1.006; P = 0.054). The log-transformed tumor markers carcinoembryonic antigen and carbohydrate antigen 19-9 yielded odds ratios of 1.868 (95 percent confidence interval, 0.788 to 4.676; P = 0.159) and 1.236 (95 percent confidence interval, 0.941 to 1.679; P = 0.132), respectively.
Model discrimination and accuracy metrics revealed that the apparent area under the receiver operating characteristic curve of 0.748 (95 percent confidence interval, 0.649 to 0.847) experienced modest attenuation during rigorous internal validation. In bootstrap validation evaluating the five fixed predictors across 1,000 resamples, the optimism-corrected area under the curve was 0.737 (95 percent confidence interval, 0.660 to 0.806), accompanied by a Brier score of 0.093, a calibration intercept of 0.208, and a calibration slope of 0.920. When internal validation accounted for data-dependent preprocessing and stability selection conditional on the 70 percent frequency threshold within outer resamples, the optimism-corrected area under the curve was 0.718 (95 percent confidence interval, 0.662 to 0.781), with a Brier score of 0.098 (95 percent confidence interval, 0.081 to 0.129), a calibration intercept of 0.185, and a calibration slope of 0.871.
A critical benchmark in the study was the comparison of the multivariable model against a baseline prevalence-only model. The intercept-only prevalence model yielded a Brier score of 0.103, illustrating that the multivariable model, with an optimism-corrected Brier score of 0.098, achieved only a limited reduction in overall prediction error beyond assuming the baseline population risk. Furthermore, evaluating classification performance using the sample-derived Youden index cutoff of 0.908 yielded an apparent sensitivity of 53.4 percent and a specificity of 85.2 percent. Because of the overwhelming 88.4 percent underlying prevalence of perineural invasion in the cohort, the positive predictive value reached 96.5 percent while the negative predictive value was only 19.3 percent, demonstrating that threshold-based classification metrics were heavily driven by cohort disease prevalence rather than exceptional predictive discrimination.
Sensitivity analyses reinforced the consistency of the core laboratory signals while confirming the presence of methodological boundaries. In a calendar-period sensitivity analysis splitting the cohort at 2019 into an earlier cohort of 117 patients and a later cohort of 116 patients, the area under the curve was 0.759 in the earlier timeframe and 0.736 in the later period, with identical Brier scores of 0.090 and calibration slopes of 1.085 and 1.056. Expanding the initial candidate predictor pool to include hepatic dysfunction indicators, such as alkaline phosphatase, transaminases, direct bilirubin, total bilirubin, and gamma-glutamyl transferase, revealed that only alanine aminotransferase reached the 70 percent selection frequency. Adding alanine aminotransferase to the five primary variables produced an optimism-corrected area under the curve of 0.722, indicating no meaningful performance improvement over the original model.
Biological interpretations of the retained markers remain strictly correlative, as the investigators emphasized that observational statistical associations do not confirm physiological mechanisms. Mean platelet volume reflects average circulating platelet size and may correlate with platelet activation states, which experimental literature has connected to tumor microenvironment interactions, angiogenesis, and cell adhesion. However, no direct platelet activation assays or histological tissue signaling measurements were evaluated. Similarly, mean corpuscular hemoglobin concentration reflects erythrocyte hemoglobin content and can capture host systemic nutritional and inflammatory status or cancer-related anemia mechanisms rather than a biliary-specific pathway. While fibrinogen and circulating tumor markers are widely linked to systemic inflammation and tumor burden, their individual nonsignificant associations in multivariable modeling highlight the difficulty of isolating specific drivers of neural invasion.
The authors identified substantial structural constraints that prevent the nomogram from being used in direct patient management. A formal sample size adequacy assessment performed using the pmsampsize criteria revealed a major information shortfall: depending on anticipated C-statistics between 0.70 and 0.80, stable development of a prediction model with 33 candidate parameters required between 2,196 and 5,365 patients, whereas only 233 were available. The presence of only 27 perineural invasion-negative patients yielded an effective ratio of roughly 5.4 minority events per retained predictor, far below established thresholds for robust multivariable estimation. Furthermore, because comparing candidate stability thresholds was performed empirically during initial development and was not repeated within the bootstrap validation loops, residual selection optimism cannot be ruled out. Consequently, the authors concluded that the five-variable nomogram serves as an exploratory, hypothesis-generating tool that requires evaluation in larger, prospective, multicenter cohorts before clinical utility can be established.
Subject of Research: Preoperative laboratory-based prediction model for perineural invasion in distal cholangiocarcinoma
Article Title: Exploratory development and internal validation of a preoperative laboratory-based nomogram for perineural invasion in distal cholangiocarcinoma
Article References: Ye, X.-Y., Wang, K., He, Q., Ji, Z., & Wang, H. (2026). Exploratory development and internal validation of a preoperative laboratory-based nomogram for perineural invasion in distal cholangiocarcinoma. Frontiers in Medicine, 13, Article 1970156. https://doi.org/10.3389/fmed.2026.1970156
Image Credits: AI Generated
DOI: 10.3389/fmed.2026.1970156
Keywords: distal cholangiocarcinoma, perineural invasion, preoperative prediction, Firth penalized logistic regression, Elastic Net regression, mean platelet volume, fibrinogen, carcinoembryonic antigen, CA19-9, mean corpuscular hemoglobin concentration, bile duct cancer, internal validation, nomogram, biliary surgery, surgical oncology, bootstrap resampling, prediction model
News Source: Nathaniel Bowman. (November 13, 2026). Exploratory Laboratory Model Shows Modest Predictive Value for Bile Duct Cancer Invasion. Scienmag.



