Gastrointestinal cancers, including malignancies of the esophagus, stomach, intestines, liver, biliary tract, and pancreas, account for a substantial share of global cancer morbidity and mortality, and surgery remains central to curative-intent treatment for many patients. Yet postoperative outcomes differ widely among patients undergoing apparently similar operations. Beyond tumor site and technical complexity, nutritional reserve, systemic inflammation, anemia, hepatic function, renal function, and coagulation status all shape a patient’s ability to tolerate major gastrointestinal surgery. Clinical nutrition research has increasingly emphasized malnutrition as a potentially modifiable perioperative risk state. In an international cancer-surgery cohort, malnutrition was associated with adverse early surgical outcomes, and perioperative nutrition management and enhanced-recovery pathways are now considered part of gastrointestinal cancer surgery quality improvement. Recent gastrointestinal oncology studies have linked albumin, prognostic nutritional indices, frailty, C-reactive protein-albumin composites, and other nutritional markers to complications or survival, motivating the question of how much a combined laboratory panel adds to clinical variables alone.

The research team turned to two large perioperative electronic health record datasets to answer that question at scale. The primary cohort was derived from INSPIRE version 1.4.2, a publicly available perioperative dataset containing operation records, diagnosis codes, laboratory values, intraoperative vital signs, medication summaries, and outcome timestamps. External validation used MIMIC-IV version 3.1, an independent hospital electronic health record dataset. The original INSPIRE cohort comprised 11,471 operations among 10,728 patients with a recorded gastrointestinal malignancy diagnosis before or at operation start and a qualifying digestive procedure code. The largest cancer groups were colorectal or anal malignancies with 4,413 operations, gastric cancer with 3,906, liver cancer with 1,445, pancreatic cancer with 712, and gallbladder or biliary cancer with 694. Multiple operations were retained, with 658 patients contributing more than one operation, and patients rather than operations were allocated in a 75:25 split, yielding 8,587 development and 2,884 test operations with no patient overlap.

The laboratory panel itself reads like the standard preoperative workup that virtually every surgical patient already receives. The researchers included albumin, bilirubin, aspartate aminotransferase, alanine aminotransferase, alkaline phosphatase, prothrombin time international normalized ratio, platelet count, C-reactive protein, white blood cell count, hemoglobin, creatinine, lactate, glucose, HbA1c, blood urea nitrogen, sodium, and potassium. They also derived composite indices such as the albumin-bilirubin score, the AST-to-platelet ratio index, and the fibrosis-4 index. Clinical variables comprised age, sex, body mass index, American Society of Anesthesiologists class, emergency status, anesthesia type, cancer group, department, operation duration, and operating-room duration. The full model added intraoperative physiologic and treatment summaries. The primary outcome was all-cause death within 365 days after operation end, and 893 deaths were recorded in the full cohort, a mortality rate of 7.8 percent.

The modeling approach used nested L2-penalized logistic regression, a technique that applies a regularization penalty to keep coefficients stable when many correlated predictors are present. The primary full design contained 63 non-intercept columns and 660 development events, roughly 10.5 events per column, a ratio that respects the classical rule of thumb for avoiding overfitting in regression models. Training-set medians and standardization were applied to numeric predictors, with categorical missingness coded as a distinct category and all categories one-hot encoded. Confidence intervals were recomputed by 500 patient-level bootstrap samples, retaining all records of each sampled patient, and for internal optimism assessment the entire development pipeline, including preprocessing, was refitted in 500 patient-level bootstrap samples. This full-pipeline bootstrap analysis yielded an optimism-corrected AUROC of 0.837, a figure that accounts for the slight inflation that arises when a model is evaluated on the same data used to build it.

The results show a clear and graded improvement as laboratory data entered the picture. In the unchanged 2,884-operation test set with 233 deaths, the clinical model achieved an area under the receiver operating characteristic curve of 0.783, with a 95 percent confidence interval of 0.752 to 0.811. Adding the full laboratory panel raised the AUROC to 0.843, with a confidence interval of 0.820 to 0.868, and the full perioperative model reached 0.856, with a confidence interval of 0.834 to 0.879. The average precision rose from 0.289 to 0.379, and the Brier score fell from 0.188 to 0.156. Notably, adding albumin and hemoglobin alone yielded an AUROC of 0.828, an increment of 0.045 over the clinical model, while the remaining laboratory panel added a further 0.015 beyond those two measurements. Internal sensitivity analyses using narrower 7-day and 14-day laboratory windows, missingness indicators, and first operations only produced test AUROCs of 0.852, 0.853, 0.854, and 0.860 respectively, supporting the stability of the finding within the original cohort.

When the researchers transported a portable model, built from age, sex, cancer group, the 17 laboratory measurements, and four derived indices, to the external MIMIC-IV dataset, performance dropped substantially. That dataset included 8,456 admissions among 5,546 patients, with 3,800 recorded deaths within 365 days. After harmonizing C-reactive protein units from the milligram-per-liter to the milligram-per-deciliter scale, the portable-model AUROC was 0.676 for one-year mortality and 0.724 for recorded hospital death. Calibration drift was substantial: for one-year mortality the observed-to-expected ratio was 1.792, meaning far more deaths occurred than the model predicted. Cross-fitted recalibration reduced the one-year Brier score from 0.277 to 0.227 without materially improving discrimination, indicating that adjusting absolute risks alone did not resolve limited risk separation. Hospital-death discrimination deteriorated when procedure-day laboratory measurements were excluded, with the AUROC falling from 0.724 to 0.611, showing that the result was sensitive to laboratory timing and limiting interpretation of the model as a strictly preoperative tool.

Descriptive risk profiles across predicted-risk strata revealed a striking graded pattern. After correcting an admission-level many-to-many merge, INSPIRE risk tertiles totaled exactly 2,884 operations and 233 deaths: low-risk operations saw 15 of 962 deaths, a mortality rate of 1.56 percent; intermediate-risk operations saw 36 of 961 deaths, or 3.75 percent; and high-risk operations saw 182 of 961 deaths, or 18.94 percent. The corrected high-to-low mortality ratio was 12.15. In MIMIC-IV, mortality ranged from 29.02 to 63.39 percent across tertiles. Higher-risk groups showed lower albumin and hemoglobin levels, consistent with the idea that nutritional and hematologic vulnerability tracks with predicted risk. However, the researchers emphasized that these profiles are descriptive because the same laboratory variables contribute to the risk score, and they do not independently establish biological phenotypes or treatment thresholds. Cancer-specific discrimination also varied, with AUROCs of 0.867 in colorectal cancer, 0.903 in gastric cancer, 0.779 in liver cancer, and 0.698 in pancreatic cancer, showing that performance is not uniform across gastrointestinal malignancies.

The contribution of albumin and hemoglobin fits a growing body of research on nutrition-related vulnerability in gastrointestinal cancer. Studies of nutritional indices, oral frailty, and albumin combined with functional measures provide clinical context for interpreting these laboratory associations. The additional information from the wider panel suggests that risk prediction may benefit from considering nutrition-related measurements alongside inflammatory and organ-function indicators, though the researchers caution that this interpretation concerns prognostic information rather than a specific nutritional mechanism. The datasets lacked the direct nutritional and functional assessments needed to distinguish nutritional depletion from other sources of laboratory abnormalities, and the 2025 ESPEN surgical nutrition guideline considers frailty, sarcopenia, and prehabilitation within perioperative care, highlighting dimensions that were not captured here. Nutritional intervention and prehabilitation studies address whether treatment improves outcomes; the present prediction analysis does not answer that question.

Several limitations affect interpretation and potential use. Broad procedure codes do not establish curative intent or tumor resection, and the colorectal grouping includes anal malignancies. Tumor stage, histology, neoadjuvant treatment, frailty, sarcopenia, dietary intake, formal nutritional assessment, nutritional intervention, and postoperative complications were not consistently available, so laboratory signals may reflect disease burden or organ dysfunction as well as nutritional vulnerability. Median imputation and missingness indicators cannot eliminate bias from informative missingness, and source outcome timestamps and non-equivalent intensive care unit definitions limit interpretation of secondary endpoints. No prospective intervention, actionable threshold, reclassification rule, or impact evaluation was defined, so clinical benefit from score-guided care remains untested. In conclusion, nutrition-related laboratory measurements improved discrimination for recorded one-year mortality beyond clinical variables in the original cohort, and their incremental prognostic information supports further study of perioperative risk stratification in patients with gastrointestinal cancer. Limited external discrimination, calibration drift, and timing-sensitive secondary outcomes warrant further validation before any clinical use.

Subject of Research: Perioperative nutrition-related laboratory profiles predict outcomes after digestive procedures in patients with gastrointestinal cancer

Article Title: Perioperative nutrition-related laboratory profiles predict outcomes after digestive procedures in patients with gastrointestinal cancer

Article References: Perioperative nutrition-related laboratory profiles predict outcomes after digestive procedures in patients with gastrointestinal cancer. (n.d.). https://doi.org/10.3389/fnut.2026.1927720

Image Credits: AI Generated

DOI: 10.3389/fnut.2026.1927720

Keywords: Perioperative, nutrition-related, laboratory, profiles, predict, outcomes, digestive, procedures, patients, gastrointestinal, cancer, scientific research