External validation of predictive models for high-grade cervical lesions in Thai women undergoing colposcopy: a
Atikom Eidbur1, Thiti Atjimakul2
1Department of Obstetrics and Gynecology, Faculty of Medicine, Prince of Songkla University, Hat Yai, Songkhla, 90110, Thailand.
Background:
Predictive models for high-grade cervical lesions (CIN2+/HSIL) have been developed in various populations; however, their generalizability across different healthcare settings is uncertain. External validation in specific populations is required before clinical implementation.
Methods:
This retrospective external validation study included Thai women who underwent colposcopy between 2022 and 2024 at a tertiary referral center. The primary outcome was histologically confirmed CIN2+/HSIL. Two previously published predictive models were applied without any modifications. Model performance was evaluated in terms of discrimination using the area under the receiver operating characteristic curve (AUROC), calibration using calibration plots, calibration metrics, Brier scores, and clinical utility using decision curve analysis.
Results:
A total of 536 women were included, of whom 28.7% were diagnosed with CIN2+/HSIL. Despite differences in demographic characteristics, HPV genotype distribution, cytological findings, and colposcopic features compared to the original development cohorts, both models demonstrated good performance in the Thai population. The Xue model demonstrated good discrimination, with an AUROC of 0.82 (95% CI: 0.77-0.86), whereas the Sheng model demonstrated good discrimination, with an AUROC of 0.80 (95% CI: 0.77-0.85). Both models showed acceptable calibration. The Xue model had a calibration-in-the-large of 0.03 (95% CI: -0.22 to 0.28) and a calibration slope of 0.77 (95% CI: 0.64-0.91), indicating mild overfitting that may benefit from recalibration. The Sheng model showed good calibration, with a slope of 0.88 (95% CI: 0.68-1.08) and a calibration-in-the-large of 0.17 (95% CI: -0.05 to 0.38). Brier scores were 0.148 and 0.163, respectively. Decision curve analysis demonstrated a positive net clinical benefit across clinically relevant threshold probabilities. At the original model-specific predefined probability thresholds, the Xue model achieved a sensitivity of 56.5% and specificity of 93.7%, whereas the Sheng model achieved a sensitivity of 93.5% and specificity of 34.3%, demonstrating a marked trade-off that highlights the importance of threshold selection.
Conclusions:
Both predictive models showed good external validity and clinical utility in Thai patients undergoing colposcopy. These findings support individualized risk stratification in women with abnormal screening results to guide clinical decisions.
