Objective:

To develop a predictive model utilizing a tabular foundation model for acute postoperative protein depletion in colon cancer patients and create an interpretable clinical web tool.

Approach:

• Study Design: Retrospective evaluation of perioperative data from 812 colon cancer patients treated between 2020 and 2025.
• Model Development: Utilized eight traditional machine learning algorithms and the TabICLv2 tabular foundation model, assessing discrimination, calibration, and clinical utility using various statistical metrics.
• Analysis Techniques: Employed AUC, paired DeLong tests, calibration curves, Brier score, NRI, IDI, and decision curve analysis for evaluation.
• Feature Contribution: Used SHAP to visualize feature contributions to model predictions.

Key Findings:

• TabICLv2 achieved the highest validation AUC of 0.766 (95% CI: 0.699–0.832) among evaluated algorithms.
• NRI/IDI analyses indicated significant total improvement over Random Forest.
• Age, prealbumin, and globulin were identified as leading contributors to model predictions.

Interpretation:

TabICLv2 integrates demographic, nutritional, and immunological profiles to predict acute postoperative protein depletion.

Limitations:

• The study is based on a single-centre cohort, limiting generalizability.
• Prospective multicentre validation is required before routine clinical implementation.

Conclusion:

The accompanying web-based calculator may facilitate individualised risk assessment, but further validation is necessary.