Participants were consecutive adults who underwent curative-intent resection of primary colorectal adenocarcinoma between September 2024 and April 2025. Patients undergoing emergency surgery for obstruction, perforation or other acute conditions were excluded, as were those with another malignancy, severe hepatic, renal or cardiac insufficiency, or radiologically confirmed liver metastasis, including oligometastatic disease. The final cohort comprised 205 patients. Body weight was recorded before surgery and at four postoperative points: postoperative day 7 or discharge, one month, three months and six months. Percentage weight loss was calculated relative to each patient’s preoperative baseline, producing a longitudinal record intended to capture differences hidden by assessments made at a single time point.

Across the cohort, mean weight loss declined from 7.8% at postoperative day 7 or discharge to 5.6% at one month, 3.5% at three months and 2.5% at six months. A linear mixed-effects model found a significant effect of time on weight loss, with every pairwise comparison between the four time points significant after Bonferroni correction. The authors described the overall pattern as partial recovery followed by stabilization, although the group mean remained below baseline at six months. Individual trajectories were more varied. Latent class growth modeling identified a three-class solution with an entropy of 0.993, indicating highly distinct classification. A four-class model was considered but rejected in favor of the more parsimonious and clinically interpretable three-class solution.

The 58 patients in the persistent weight-loss group, representing 28.3% of the cohort, maintained an approximately 8% deficit throughout follow-up: 8.0%, 7.9%, 8.1% and 7.8% at the four postoperative measurements. The 79 patients in the gradual-recovery group, or 38.5%, began with similar early loss but progressively approached baseline, recording 7.8%, 6.0%, 2.9% and 0.6%. The remaining 68 patients, 33.2%, formed the rapid-recovery group and were near baseline by three months; their corresponding values were 7.6%, 3.0%, 0.4% and 0.1%. These groups describe measured differences in recovery patterns rather than causal effects. Multinomial logistic regression identified several factors associated with class membership, including age, American Society of Anesthesiologists physical-status grade, tumor location, sex and adjuvant chemotherapy.

Relative to persistent weight loss, patients older than 65 had lower odds of gradual recovery, with an odds ratio of 0.403 and 95% confidence interval of 0.179–0.910, and lower odds of rapid recovery, with an odds ratio of 0.302 and 95% confidence interval of 0.112–0.812. ASA grade III–IV was also associated with lower odds of both recovery patterns. By contrast, colon rather than rectal tumor location was associated with higher odds of gradual recovery: the odds ratio was 3.432 for left-colon cancer and 6.368 for right-colon cancer. For rapid recovery, corresponding odds ratios were 2.966 and 9.443. Female sex and adjuvant chemotherapy were additionally associated with lower odds of rapid recovery. These observational associations cannot establish why patients followed particular trajectories or whether chemotherapy caused nutritional deterioration. The analysis also showed that a six-month window captured a more stable phase than a three-month assessment, when many patients in the gradual-recovery group were still regaining weight.

At six months, investigators diagnosed malnutrition using the two-stage Global Leadership Initiative on Malnutrition framework. The Nutritional Risk Screening 2002 score was used first; because the framework treats cancer-related disease burden and inflammation as an etiologic condition, that criterion was considered satisfied for all participants. Malnutrition was then assigned when nutritional risk was present together with at least one phenotypic feature: weight loss greater than 5% within six months, low body mass index, or reduced skeletal muscle measured by CT at the third lumbar vertebra. The BMI thresholds were 18.5 kg/m² for people younger than 70 and 20 kg/m² for those aged 70 or older, while sex-specific CT skeletal muscle index cutoffs were 40.8 cm²/m² for men and 34.9 cm²/m² for women. Sixty-eight of the 205 patients met the resulting definition, producing a six-month prevalence of 33.2%.

To build the prediction models, investigators used Lasso regression and the Boruta algorithm in sequence. Feature selection was repeated inside each training fold of five-fold cross-validation, limiting information leakage from the held-out validation sets. The intersection of the two approaches consistently selected four variables: age, preoperative BMI, tumor location and planned adjuvant chemotherapy. All four were available by the time postoperative pathology was reported, generally five to seven days after surgery and no later than hospital discharge. Collinearity diagnostics indicated that the predictors were not strongly correlated, with variance inflation factors from 1.003 to 1.015. The model therefore relied on four clinical variables that could be assembled near the postoperative index date, although the study did not test the model prospectively in routine care.

Ten algorithms were then compared using cross-validation: support vector machine, gradient boosting machine, neural network, k-nearest neighbors, naive Bayes, multilayer perceptron, XGBoost, random forest, decision tree and logistic regression. The assessment covered discrimination, calibration and clinical utility rather than a single accuracy measure. XGBoost produced the highest AUC, 0.836, alongside 76.6% sensitivity, 77.9% specificity, 77.5% accuracy and an F1 score of 0.693. Its Brier score of 0.098 was the lowest among the ten models, and calibration curves showed favorable agreement between estimated and observed probabilities. Decision-curve analysis indicated net clinical benefit across threshold probabilities of 9–91%. The other leading models included naive Bayes with an AUC of 0.822, a multilayer perceptron at 0.820 and random forest at 0.812. These estimates came from internal cross-validation in one center, rather than independent external testing.

SHapley Additive exPlanations, a method for decomposing individual model predictions, ranked preoperative BMI as the strongest contributor, followed by adjuvant chemotherapy, age and tumor location. The directional analysis associated lower preoperative BMI, receipt of adjuvant chemotherapy, older age and rectal tumor location with higher predicted risk of malnutrition. The findings distinguish predictive importance from causal responsibility: SHAP values show how the fitted model uses a variable, not whether that variable caused malnutrition. The apparent importance of BMI also differed from its role in the trajectory analysis. BMI did not significantly distinguish the three recovery groups, but it could still be predictive of six-month malnutrition because patients starting with a lower BMI might remain below nutritional thresholds even while regaining weight.

Several limitations constrain the conclusions. This was a single-center study with a modest sample, and the cohort required at least three postoperative weight measurements and a six-month abdominal CT scan, potentially excluding patients with poorer follow-up or more difficult trajectories. The model received internal cross-validation but no external multicenter validation. The multivariable trajectory analysis selected covariates partly on the basis of univariate significance, potentially omitting relevant factors. Nutritional status was assessed only at six months, so outcomes beyond that point remain unknown. In addition, automatically treating the GLIM etiologic criterion as present in every patient may have affected the estimated prevalence, and the authors called for future studies using objective inflammatory measures. The results therefore support further investigation of a compact risk-screening approach, but the study did not establish whether model-guided monitoring or treatment improves patient outcomes.

Subject of Research: Postoperative weight trajectories and machine-learning prediction of six-month malnutrition in colorectal cancer patients

Article Title: Dynamic weight loss trajectories and machine learning-based prediction of malnutrition in colorectal cancer patients after curative surgery

Article References: Chen, X., Jian, J., Lin, Z., Xie, M., & Li, Y. (2026). Dynamic weight loss trajectories and machine learning-based prediction of malnutrition in colorectal cancer patients after curative surgery. Frontiers in Nutrition, 13, 1982995. https://doi.org/10.3389/fnut.2026.1982995

Image Credits: AI Generated

DOI: 10.3389/fnut.2026.1982995

Keywords: colorectal cancer, postoperative malnutrition, weight-loss trajectories, machine learning, XGBoost, risk prediction, GLIM criteria, nutrition oncology, clinical validation, Dynamic, weight, loss