Tertiary Lymphoid Structures as Predictors of Recurrence in Colorectal Cancer: Development and Validation of a
Xian-Hua Lei1, Rong Li1, Dong-Mei Wang1
1Department of Pathology, Ganzhou Cancer Hospital, The Affiliated Cancer Hospital of Gannan Medical University, Ganzhou, Jiangxi, China.
Background:
Tertiary lymphoid structures (TLS) are prognostic immune aggregates in the tumor microenvironment, but the value of location-specific TLS features for predicting colorectal cancer (CRC) recurrence remains unclear. This study developed and internally validated a machine-learning (ML) model integrating TLS features for recurrence prediction in stage II-III CRC.
Methods:
We retrospectively included 224 patients with stage II-III CRC after curative resection and split them into training (n = 156) and held-out internal validation (n = 68) cohorts. TLS at intratumoral, invasive-front, and peritumoral sites were semi-quantitatively scored by two blinded pathologists. LASSO selected predictors from 39 candidate variables. Five ML algorithms were trained with stratified cross-validation, predefined tuning, and training-only preprocessing. Performance was evaluated using discrimination, calibration, decision-curve/clinical-impact analyses, reclassification indices, SHAP interpretation, and Kaplan-Meier risk stratification.
Results:
Recurrence occurred in 89 patients (39.7%). LASSO retained peritumoral TLS score, invasive-front TLS score, and preoperative CEA. TLS scoring showed high interobserver agreement (weighted kappa: 0.82-0.86). LightGBM provided the most balanced performance, with AUCs of 0.872 (95% CI: 0.814-0.923) in training and 0.718 (95% CI: 0.572-0.839) in internal validation. SHAP ranked invasive-front TLS score as the most influential predictor. The Youden cutoff (0.5017) separated high- and low-risk groups with significantly different disease-free survival (log-rank p < 0.0001; hazard ratio = 5.132, 95% CI: 3.263-8.073). In exploratory comparison, validation AUCs were 0.727 for TLS-only, 0.542 for clinical-only, and 0.737 for combined models.
Conclusions:
A TLS-based ML model showed moderate internal validation performance for CRC recurrence prediction. Location-specific TLS features may support postoperative risk stratification, but external multicenter and molecularly integrated validation is required before clinical use.