A self-attention deep learning model for pre-treatment prediction of lymph node metastasis in FIGO IIB cervical cancer: Development, temporal validation and interpretability analysis

Objective To develop and validate an interpretable deep-learning model that predicts radiologically defined regional lymph node metastasis (LNM) from routinely available pre-treatment clinicopathological variables in FIGO 2018 stage IIB cervical cancer. Materials We screened and collected 616 cases with stage II cervical cancer patients in the Department of Gynecological Oncology of the Affiliated Hospital of Jinggangshan University from December 2006 to October 2020. The clinicopathological data were systematically collected. Methods In 616 patients treated with definitive chemoradiotherapy (2006–2020), LNM (regional node ≥10 mm short-axis on CT/MRI/PET-CT) occurred in 101, 16.4%. A feature-tokenising Transformer using nine pre-treatment variables was compared with seven baselines on identical folds with equal class weighting and hyperparameter budgets, using 10 × 5-fold nested cross-validation and a temporally independent cohort (2017–2020, n = 195). Results The model achieved AUROC 0.793 and Brier 0.102 in cross-validation and AUROC 0.771 (95% CI 0.688–0.851) on temporal validation, significantly exceeding logistic regression (ΔAUROC 0.062, p < 0.001) but not XGBoost (ΔAUROC 0.017, p = 0.11). Sensitivity was 0.76, specificity 0.71 and NPV 0.94; risk stratification separated LNM rates of 3.8%, 16.2% and 41.9%, with net benefit across thresholds of 0.08–0.35. Conclusions A self-attention model over routine pre-treatment variables provides modestly better-calibrated and better-discriminating estimates of nodal risk in FIGO IIB cervical cancer than logistic regression, and performs comparably to well-tuned gradient boosting. The model is suitable for hypothesis generation and for triage to further imaging, but requires prospective multicentre validation against a pathological reference standard before clinical use.