AI Model Predicts Colorectal Cancer Nodal Metastasis

Key Takeaways

• Among newly diagnosed colorectal cancer patients at two Chinese hospitals who underwent direct radical surgery without preoperative anticancer treatment, a six-variable model showed moderate external discrimination for pathological nodal disease.
• The model significantly outperformed clinical T stage alone in internal testing and independent external validation.
• Six selected predictors did not differ significantly in discrimination from all prespecified candidate predictors in either validation set.
• Decision-curve net benefit was exploratory; improved clinical decisions or patient outcomes were not demonstrated.

Regional nodal involvement can remain uncertain before surgery for newly diagnosed colorectal cancer (CRC), despite conventional imaging and clinical T staging. Clinical measurements, laboratory tumor markers, and endoscopic biopsy findings are available before resection pathology establishes lymph node metastasis (LNM) status. Patients proceeding directly to surgery provide postoperative findings against which preoperative estimates can be assessed.

In a JMIR Cancer study, investigators retrospectively evaluated 2,725 adults treated from July 2023 through December 2025 at two Chinese hospitals. All had newly diagnosed CRC and underwent direct radical resection with regional lymph node dissection, without preoperative anticancer therapy. Harbin contributed 1074 patients. Final postoperative pathology, rather than imaging or model output, defined the binary regional LNM outcome.

Of the Harbin patients, 753 formed the training set and 321 the internal test set. A separate Shanxi hospital supplied 1651 patients for independent external validation against the same pathological endpoint.

Training-set-only logistic-regression screening selected body mass index (BMI), preoperative carcinoembryonic antigen (CEA), primary tumor site, clinical T stage, biopsy histological type, and biopsy tumor differentiation. Clinical T stage came from imaging and endoscopic assessment; the histological inputs came from preoperative biopsy, not resection pathology. Among seven candidate algorithms, random forest (RF) was selected on combined internal-test performance, without demonstrated superiority over every alternative. Its output estimated the probability of pathological regional LNM. Shapley additive explanations (SHAP) ranked clinical T stage highest in contribution to predictions, not causal importance.

Discrimination, measured by area under the receiver operating characteristic curve (AUROC) with a 95% confidence interval (CI), was 0.803 (95% CI 0.756–0.849) in internal testing and 0.776 (95% CI 0.755–0.797) at the independent center. Both estimates concern patients on the direct-surgery pathway.

Clinical T stage alone reached an internal-test AUROC of 0.712, a difference of 0.091 from the six-predictor model (P<.001). This showed additional discrimination, not a measured improvement in care.

At the external hospital, clinical T stage alone reached an AUROC of 0.707, a difference of 0.069 (P<.001). This independent-center comparison also favored the multivariable model.

At the training-derived operating point, external sensitivity was 0.677 and specificity was 0.693. The investigators carried that operating point unchanged into both validation sets; it was an evaluation threshold, not a validated clinical cutoff.

External positive predictive value (PPV) was 0.569, while negative predictive value (NPV) was 0.781. These measures describe classification at the same operating point.

The external Brier score was 0.190, and the calibration slope was 1.061. Discrimination did not differ significantly between the six selected predictors and all prespecified candidate predictors in either validation set. Decision-curve analysis (DCA) explored net benefit but did not establish changes in treatment decisions or patient outcomes.

The retrospective evaluation was limited to two Chinese centers and patients receiving direct radical surgery. Performance was not validated for neoadjuvant-treatment, endoscopic-resection, local-excision, or nonoperative pathways. For U.S. clinicians, transportability to North American surgical populations remains untested. Biopsy histological type and differentiation agreed imperfectly with resection pathology, constraining two preoperative inputs.

The investigators characterized the model as a research-stage adjunct that supplied information beyond clinical T stage, not a replacement for imaging, pathology, or clinical judgment. They identified prospective validation in broader populations and clinical settings, including evaluation of clinical impact before implementation, as the next step.

Clinician Questions

How were missing preoperative predictors handled in the colorectal cancer nodal-risk model?

For Xu and colleagues’ colorectal cancer regional nodal-metastasis model, investigators estimated continuous-variable imputation and categorical-variable mode replacement using development training data. They applied those preprocessing rules to the internal and external validation sets without refitting; pathological nodal outcomes were not imputed.

Which colorectal cancer histological types counted as special types when biopsy findings were compared with resection pathology?

In the biopsy–resection agreement assessment, the special histological-type category included mucinous adenocarcinoma and signet-ring cell carcinoma; the alternative was adenocarcinoma. A special type found at resection could have been classified as adenocarcinoma on preoperative biopsy.

Recommended Reading

• For more on validating clinical AI models: The Science of Exceptions: Setting the Safety Bar for Clinical AI