This single-center retrospective cohort study was conducted at the Department of Obstetrics and Gynecology, Shanxi Bethune Hospital, Shanxi Academy of Medical Sciences, Third Hospital of Shanxi Medical University, Tongji Shanxi Hospital, Taiyuan, China. We enrolled consecutive women who underwent loop electrosurgical excision procedure (LEEP) for histologically confirmed cervical intraepithelial neoplasia grade 2 or higher (CIN2+) between January 2018 and October 2024, with follow-up data extracted from electronic medical records through October 2025. This study was purely retrospective; all data were retrieved from existing hospital databases, and no participants were recruited or contacted for the purpose of this research. The study protocol adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (38). The development and validation of the prediction model were reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement (39).

The inclusion criteria were age ≥18 years at LEEP, histologically confirmed cervical intraepithelial neoplasia grade 2 or higher (CIN2+), including adenocarcinoma in situ, on index LEEP specimen reviewed by two independent pathologists, complete baseline clinical and pathological data, and minimum 12-month follow-up or until recurrence diagnosis. The exclusion criteria included previous hysterectomy, pregnancy at time of LEEP or within 6 months post-LEEP, invasive cervical cancer on index specimen (except microinvasive stage IA1 managed definitively with LEEP), HIV-positive serostatus, concurrent vaginal or vulvar high-grade neoplasia, documented loss to follow-up within 12 months, incomplete pathological specimens precluding margin assessment, or participation in investigational HPV therapeutic vaccine trials. The final cohort comprised 2,230 women with 334 recurrence events (15.0%) over median 31.8 months of follow-up (interquartile range 19.6–43.5 months).

All patients underwent standardized pre-LEEP evaluation, including structured clinical history and physical examination with height/weight measurement, liquid-based cytology using ThinPrep system (Hologic, USA) classified per Bethesda 2014 (40), high-risk HPV DNA testing via Hybrid Capture 2 (HC2, Qiagen) detecting 13 oncogenic types collectively (16, 18, 31, 33, 35, 39, 45, 51, 52, 56, 58, 59, and 68) with positivity threshold RLU/CO ≥1.0, HPV genotyping via Linear Array (Roche) for HC2-positive cases, colposcopic examination per IFCPC 2011 nomenclature documenting transformation zone type and lesion characteristics, and colposcopy-directed punch biopsy with endocervical curettage when indicated (41). Complete blood count with differential was obtained within 1 week pre-LEEP using Sysmex XN-series automated analyzer; neutrophil-to-lymphocyte ratio (NLR) was calculated as absolute neutrophil count divided by absolute lymphocyte count. All pre-LEEP histopathology was reviewed by two independent gynecologic pathologists using standardized criteria; discordant cases were adjudicated by a senior pathologist. CIN2 diagnoses underwent p16 immunohistochemistry (CINtec, Roche) when histologically equivocal.

LEEP procedures were performed in outpatient setting under local anesthesia (paracervical block with 1% lidocaine plus epinephrine 1:100,000) using standardized electrosurgical settings (blended current 35–45 watts for cutting, pure coagulation 40–50 watts for hemostasis). Single-pass excision technique was attempted for all lesions to produce intact specimens optimal for margin assessment; multiple-pass approach was employed only when single-pass was inadequate due to lesion size (>2.5 cm), extensive circumferential involvement (>75%), or deep endocervical extension. Specimens were oriented with suture at 12 o’clock position, measured in three dimensions using calibrated calipers (transverse diameter, anteroposterior diameter, and cone depth in millimeters), and submitted fresh to pathology.

All LEEP specimens underwent standardized processing in our CAP-accredited pathology laboratory. The oriented specimens were fixed in 10% neutral buffered formalin for 12–24 h, sectioned perpendicular to the endocervical axis at 2- to 3-mm intervals, entirely embedded, and stained with hematoxylin–eosin. The pathologists evaluated the specimens using the standardized synoptic reporting type of documenting: specimen dimensions, highest-grade CIN lesion (CIN2, CIN3, or adenocarcinoma in situ), extent of dysplasia, glandular involvement (binary: CIN extension into endocervical gland crypts present versus absent), and three-margin status. The margins were assessed as endocervical margin (superior edge), ectocervical margin (inferior edge), and deep/stromal margin (lateral circumferential and basal). Each margin was classified as positive (CIN2+ at margin or within 1 mm) or negative (>1 mm clearance) based on the minimum distance measured via calibrated ocular micrometer on multiple sections. Multifocal margin involvement was categorized as none (all margins negative), single (one margin positive), or multiple (≥2 margins positive). All specimens received mandatory dual-pathologist review; final diagnoses were recorded in an electronic pathology system.

Post-LEEP surveillance followed institutional protocol adapted from ASCCP and Chinese guidelines: co-testing (cytology plus HPV) at 6–9, 12–15, 18–21, and 24–27 months post-LEEP, then annually through year 5 (42, 43). Colposcopies with directed biopsy was triggered by any cytological abnormality (ASC-US or higher), positive high-risk HPV regardless of cytology, or clinical concern. When colposcopy revealed no visible lesion despite positive co-testing, endocervical curettage was performed. All surveillance procedures were documented in the electronic tracking system with automated reminders for missed appointments and telephone contact attempts before classifying patients as lost to follow-up.

The primary outcome was histologically confirmed CIN2+ recurrence, defined as detection of CIN2, CIN3, adenocarcinoma in situ, or invasive cervical cancer on biopsy or excisional specimen obtained ≥6 months following index LEEP. This 6-month threshold distinguished true recurrence from residual/persistent disease. Recurrence required histological confirmation by two pathologists; cytological abnormalities alone were not considered recurrence. Follow-up duration was calculated from index LEEP to recurrence diagnosis, death from any cause, documented loss to follow-up, or study end (October 31, 2025). Electronic medical records were cross-referenced with regional cancer registry, pathology database, and civil death registry for outcome ascertainment.

Demographic variables: Age at LEEP was calculated as completed years from birth to procedure date, analyzed continuously per 10-year increment and categorically (<40 versus ≥40 years). Body mass index (BMI) was calculated as weight (kg) ÷ height² (m²) from measurements obtained at pre-LEEP visit, analyzed continuously per five-unit increment and categorized per WHO Asian-Pacific criteria (underweight <18.5, normal 18.5–22.9, overweight 23.0–24.9, obese ≥25.0 kg/m²) (44). Menopausal status was classified as premenopausal (regular cycles within 12 months or irregular cycles with last period within 12 months and age <45 years) versus postmenopausal (≥12 months amenorrhea without alternative cause in women ≥45 years or bilateral oophorectomy regardless of age).

Reproductive history: Parity was defined as the number of pregnancies reaching ≥24 weeks of gestation resulting in live birth or stillbirth, categorized as nulliparous (0), one to two children, or ≥3 children. New sexual partner during follow-up was defined as self-reported acquisition of new partner between LEEP and recurrence/last visit. This information was retrieved from standard nursing assessment forms recorded in the electronic medical record, which are administered as part of routine clinical care at each surveillance visit (binary: yes/no).

Lifestyle factors: Smoking status was classified as never smoker (lifetime <100 cigarettes), former smoker (≥100 cigarettes but cessation >6 months pre-LEEP), or current smoker (active use within 6 months of LEEP or continued during follow-up), with former and current combined as “ever smoker” for analysis. HPV vaccination status was documented via immunization registry, vaccination card, or medical records, classified as unvaccinated (no doses), partially vaccinated (incomplete schedule), or fully vaccinated (completed age-appropriate regimen: three doses for bivalent/quadrivalent or two doses for nonavalent if age <15 years at initiation).

Pre-LEEP disease characteristics: Previous cervical treatment was defined as any documented prior therapeutic intervention (cryotherapy, laser, previous LEEP, and conization) identified through medical records (binary). Referral cytology was recorded as the most severe result within 6 months pre-LEEP (ASC-US, LSIL, HSIL, ASC-H, and AGC per Bethesda 2014). Pre-LEEP biopsy histology was documented as highest-grade diagnosis on punch biopsy/ECC within 3 months pre-LEEP (CIN1, CIN2, and CIN3). Transformation zone type was assessed colposcopically per IFCPC 2011: type 1 (squamocolumnar junction fully visible), type 2 (junction visible with endocervical component), or type 3 (junction not fully visible).

LEEP procedure variables: Indication was classified as diagnostic (biopsy-proven CIN2+ requiring LEEP for definitive assessment) versus therapeutic (see-and-treat or treatment following biopsy). Excision type was recorded as single-pass (entire lesion in one piece) versus multiple-pass (≥2 pieces due to size/extent). Cone depth was measured in millimeters by a pathologist on fresh specimen as the vertical dimension from the ectocervical surface to the deepest endocervical extent, analyzed continuously per 5-mm increment.

Pathological variables: Highest CIN grade in LEEP specimen was classified as CIN2 (moderate dysplasia), CIN3 (severe dysplasia/carcinoma in situ), or adenocarcinoma in situ, determined by two pathologists. Glandular involvement was defined as CIN extension into endocervical gland crypts on H&E sections (binary: present/absent). Individual margin status (endocervical, ectocervical, deep) was classified as positive (<1-mm clearance) or negative (>1-mm clearance). Multifocal margin involvement was categorized as none, single margin positive, or multiple margins positive (≥2 margins).

HPV molecular variables: Pre-LEEP high-risk HPV status was determined via HC2 within 3 months pre-LEEP (positive: RLU/CO ≥1.0, negative: <1.0). HPV genotype was determined via Linear Array for positive cases; women with multiple genotypes were assigned to most oncogenic type per hierarchy (HPV 16 > HPV 18 > other high-risk types). The primary genotype categories analyzed were HPV 16, HPV 18, and other high-risk types (31, 33, 35, 39, 45, 51, 52, 56, 58, 59, 68). Post-LEEP high-risk HPV status was tested at 6–9 months using identical HC2 methodology (positive/negative). Persistent HPV infection was defined as same high-risk genotype detected both pre-LEEP (within 3 months before) and post-LEEP (6–9 months after) via genotype-specific testing; coded as persistent versus non-persistent (different genotypes, cleared, or new infection).

Inflammatory biomarker: Neutrophil-to-lymphocyte ratio (NLR) was calculated from CBC obtained within 1 week pre-LEEP as absolute neutrophil count ÷ absolute lymphocyte count, analyzed continuously per one-unit increment. Samples with concurrent acute infection (fever, WBC >15,000/μL or clinical infectious diagnosis) were excluded to avoid confounding (n = 23, 1.0%).

Surveillance variables: First post-LEEP cytology was recorded as initial result at 6–9 months, classified per Bethesda 2014 (normal/NILM, ASC-US, LSIL, HSIL, or higher). Adjuvant management was categorized as observation only (surveillance without additional treatment) versus repeat excision (repeat LEEP, conization, or hysterectomy during follow-up). Immunosuppression status was defined as solid organ transplant with maintenance immunosuppression, autoimmune disease treated with chronic immunosuppressives (corticosteroids ≥20 mg prednisone-equivalent daily ≥3 months, DMARDs, or biologics), hematologic malignancy, or immunodeficiency disorders other than HIV, as HIV-seropositive individuals were excluded from this cohort per the exclusion criteria (binary: present/absent).

Outcome variables: Total follow-up duration was calculated from LEEP to last encounter, recurrence, death, or study end (months). Time to recurrence was measured from LEEP to histological CIN2+ confirmation for recurrent cases (months). Recurrence was defined as CIN2+ detected ≥6 months post-LEEP; events <6 months were classified as residual disease and excluded from primary analysis. Loss to follow-up was defined as failure to attend surveillance >18 months with unsuccessful contact attempts.

Continuous variables were evaluated for distributional assumptions using Shapiro–Wilk test alongside graphical inspection; because several measures were non-normally distributed, results are summarized as medians with interquartile ranges (IQRs). Categorical variables are reported as frequencies and proportions. Baseline comparisons between participants with and without recurrent CIN2+ were undertaken using Mann–Whitney U-test for continuous variables and χ² test or Fisher’s exact test for categorical variables as appropriate. Effect sizes were reported as rank-biserial correlation for continuous comparisons and Cramér’s V for categorical comparisons. All tests were two-sided, with statistical significance defined as p < 0.05. Sample size was prespecified for time-to-event prediction modeling. Assuming a 15% recurrence proportion and 16 candidate predictors in the final multivariable model, a minimum of 160 recurrent events was required consistent with contemporary recommendations for events-per-parameter ratios in time-to-event prediction modeling (45). The observed cohort accrued 334 CIN2+ recurrences, providing approximately 20.9 events per predictor and supporting stable multivariable estimation. On this basis, the available sample was considered sufficient to detect moderate associations compatible with clinical prediction modeling within a single-center cohort.

Time-to-event analyses used Cox proportional hazards regression with recurrent CIN2+ as the event. Time zero was the date of index LEEP, and participants were censored at the last documented follow-up date. Ties were handled using the Efron method. The proportional hazards assumption was evaluated using tests based on Schoenfeld residuals and inspection of log–log survival plots; no material departures were carried forward into the primary modeling strategy. Univariable Cox models were fitted for each candidate predictor to quantify crude associations, reporting hazard ratios (HRs) with 95% confidence intervals (CIs). Continuous predictors were parameterized on clinically interpretable scales to enhance interpretability and comparability: age per 10 years, body mass index per 5 kg/m², excision depth per 5 mm, and neutrophil-to-lymphocyte ratio (NLR) per one-unit increment. Predictors meeting a liberal screening threshold (p < 0.10) were prioritized for multivariable modeling, while clinically essential variables were retained irrespective of univariable significance to preserve face validity and alignment with established post-treatment risk frameworks.

Multivariable modeling followed a prespecified hierarchical strategy reflecting the clinical sequence of information acquisition and the conceptual pathway from host factors to lesion biology and viral persistence. Model 1 included demographics and lifestyle characteristics (age, body mass index, menopausal status, parity, smoking status, and HPV vaccination status). Model 2 added pre-LEEP disease and clinical history measures (previous cervical treatment, pre-LEEP biopsy grade, and transformation zone type). Model 3 incorporated index LEEP pathology and surgical-pathologic variables (cone depth, highest CIN grade in the specimen, glandular involvement, and multifocal margin involvement operationalized as none, single-margin positive, or multiple-margins positive). Model 4, the final nomogram model, further included HPV-related molecular variables and inflammatory biomarker information (HPV genotype grouping, persistent HPV infection defined by genotype-concordant persistence, and NLR). Incremental predictive contribution across models was assessed using likelihood ratio χ² tests and the Akaike information criterion (AIC). Multicollinearity was evaluated using variance inflation factors, with values >5 considered potentially problematic; all predictors retained in the final model remained below this threshold. Adjusted hazard ratios (aHRs) with 95% CIs were reported for multivariable models, and p < 0.05 denoted statistical significance in the final model.

Effect heterogeneity for persistent HPV infection—the dominant biological predictor—was explored across prespecified strata (age <40 vs. ≥40 years, menopausal status, highest CIN grade, margin status, vaccination status, and smoking status). Stratified Cox models were fitted within each subgroup, and multiplicative interaction terms were tested using likelihood ratio tests; non-significant interactions were interpreted as evidence of broadly consistent effects. A clinical nomogram was derived from the final multivariable Cox model by translating regression coefficients to a point-based system proportional to the β coefficients. Total points were mapped to predicted recurrence probabilities at 24 and 36 months using the baseline survival function estimated via the Breslow approach. Discrimination was quantified using Harrell’s concordance index (C-index) and time-dependent discrimination metrics at clinically relevant time horizons (6, 12, 24, 36, and 60 months), accounting for censoring. The incremental value of HPV-related predictors was assessed by comparing the discrimination of the full model with that of the model excluding HPV variables. Calibration was evaluated at 24 months by comparing predicted probabilities with observed event rates estimated by Kaplan–Meier methods across deciles of predicted risk, supplemented by the calibration slope derived from regressing observed outcomes on predicted log-risk. Overall calibration error was summarized using the Brier score. Clinical utility was examined using decision curve analysis across clinically plausible threshold probabilities for intensified surveillance, demonstrating a net benefit of nomogram-guided strategies over default “treat-all” or “treat-none” approaches across relevant risk ranges. For clinical risk stratification, the nomogram score was categorized into three pragmatic groups (low, moderate, and high) using prespecified point ranges intended to reflect actionability. Kaplan–Meier curves were constructed for these strata and compared using log-rank test, and stratum-specific cumulative recurrence estimates were reported with 95% CIs. Internal validation used bootstrap resampling (500 iterations) to estimate optimism-corrected performance, including C-index and calibration slope. Missingness was described and assessed for plausibility of missing completely at random using Little’s test. The primary analysis used complete-case modeling for variables with minimal missingness; for predictors with ≥5% missingness, sensitivity analyses used multiple imputation by chained equations (50 imputations; predictive mean matching for continuous variables and logistic or multinomial models for categorical variables), with coefficients pooled using Rubin’s rules. Additional sensitivity analyses examined robustness to alternate endpoint definitions (CIN3+ only; inclusion of early residual disease), competing-risk specification (Fine–Gray sub-distribution hazards with death as a competing event), and temporal stability by comparing model performance across calendar periods within the study window. All analyses were conducted using R (version 4.3.1) with established packages for survival modeling, prediction, calibration, decision curve analysis, and multiple imputation; confidence intervals were two-sided at the 95% level, and univariable analyses were treated as exploratory while inference focused on the prespecified final multivariable model.

A total of 2,230 women who underwent loop electrosurgical excision for CIN2+ were included, among whom 334 developed histologically confirmed CIN2+ recurrence (15.0%) (Figure 1A). The median follow-up duration was 31.8 months (IQR 19.6–43.5), and the median time to recurrence was 15.6 months (IQR 8.2–24.3) (Figure 1A). Baseline characteristics stratified by recurrence status are summarized in Table 1. Women who developed recurrence were more frequently postmenopausal (17.1% vs. 13.4%; p = 0.04) and multiparous (≥3 children: 37.4% vs. 28.6%; p = 0.008) and had a higher prevalence of current or former smoking (12.6% vs. 9.0%; p = 0.02) (Table 1). HPV vaccination coverage differed substantially between groups (p < 0.001), with a higher proportion unvaccinated among women with recurrence (89.2% vs. 79.2%) and a lower proportion fully vaccinated (2.7% vs. 7.8%) (Table 1; Figure 1C). Prior cervical treatment was more common among women with recurrence (18.9% vs. 12.5%; p = 0.006) (Table 1). In contrast, age at LEEP (median 39.0 vs. 38.0 years; p = 0.12) and body mass index (median 23.8 vs. 23.6 kg/m²; p = 0.31) did not differ materially (Table 1; Figure 1B).

With respect to index LEEP pathology, the distribution of the highest CIN grade differed between groups (p = 0.03), including a higher proportion of adenocarcinoma in situ among women with recurrence (28.4% vs. 23.6%) (Table 1; Figure 1E). Endocervical margin positivity was more frequent among women with recurrence (40.7% vs. 35.3%; p = 0.01) (Table 1). Multifocal margin involvement also differed (p = 0.04), with multiple positive margins observed in 31.7% of women with recurrence versus 28.4% without recurrence (Table 1; Figure 1C).

Several procedural and pathological characteristics were broadly comparable between groups, supporting the specificity of associations observed for margin burden and virological factors. Transformation zone type showed modest distributional differences that did not reach conventional statistical significance (p = 0.09), with type 1 observed in 41.3% versus 47.0%, type 2 in 43.1% versus 39.2%, and type 3 in 15.6% versus 13.8% among women with and without recurrence, respectively (Table 1). Excision technique was similar (single-pass: 78.1% vs. 80.4%; p = 0.42), and cone depth did not differ meaningfully (median 15.1 mm [IQR 12.1–18.3] vs. 14.6 mm [11.5–17.9]; p = 0.22) (Table 1). Glandular involvement was more frequent in the recurrence group (32.3% vs. 27.5%), although the difference was not statistically significant (p = 0.08), and neither ectocervical margin positivity (29.6% vs. 27.7%; p = 0.18) nor deep/stromal margin positivity (13.2% vs. 12.8%; p = 0.25) differed materially (Table 1).

HPV-related measures exhibited the largest between-group separations. Pre-LEEP high-risk HPV positivity was high in both groups (94.3% vs. 93.0%; p = 0.31), whereas post-LEEP high-risk HPV positivity at first follow-up was markedly more common among women with recurrence (51.5% vs. 27.8%; p < 0.001) (Table 1; Figure 1C). Persistent HPV infection (same genotype pre-/post-LEEP) occurred in 50.6% of women with recurrence compared with 23.1% without recurrence (p < 0.001) (Table 1; Figure 1C). HPV genotype distribution among HPV-positive women was similar between groups (p = 0.15), with HPV 16 detected in 48.5% versus 47.3% and HPV 18 in 19.5% versus 17.0% among women with and without recurrence, respectively (Table 1). Inflammatory status also differed, with a higher pre-LEEP neutrophil-to-lymphocyte ratio in the recurrence group (median 2.4 vs. 2.2; p = 0.03) (Table 1; Figure 1D). First post-LEEP cytology differed strongly by outcome (p < 0.001), including higher frequencies of ASC-US (22.8% vs. 15.0%), LSIL (14.4% vs. 6.5%), and HSIL or higher (9.6% vs. 3.0%) among women with recurrence (Table 1).

In univariable Cox regression (Table 2; Figure 2A), the strongest association was observed for persistent HPV infection (HR 2.87, 95% CI 2.31–3.57; p < 0.001). Additional predictors included abnormal first post-LEEP cytology (HR 2.28, 95% CI 1.84–2.82; p < 0.001), post-LEEP HPV positivity (HR 2.13, 95% CI 1.72–2.64; p < 0.001), and HPV unvaccinated status (HR 1.89, 95% CI 1.33–2.68; p < 0.001) (Table 2). Procedural and clinical history variables were also associated, including repeat excision (HR 1.64, 95% CI 1.14–2.36; p = 0.008) and previous cervical treatment (HR 1.58, 95% CI 1.19–2.09; p = 0.002) (Table 2). Pathology and margin-related predictors included multiple margins positive versus all margins negative (HR 1.56, 95% CI 1.13–2.16; p = 0.007), single margin positive versus all margins negative (HR 1.45, 95% CI 1.08–1.94; p = 0.01), and highest CIN grade AIS versus CIN2 (HR 1.51, 95% CI 1.08–2.12; p = 0.02) (Table 2). The neutrophil-to-lymphocyte ratio was also associated with recurrence risk (per one-unit increase: HR 1.24, 95% CI 1.05–1.46; p = 0.01) (Table 2). In contrast, several clinically relevant features showed weaker or non-significant univariable associations, including transformation zone type 3 versus type 1/2 (HR 1.21, 95% CI 0.90–1.64; p = 0.21), cone depth per 5-mm increase (HR 0.92, 95% CI 0.78–1.09; p = 0.33), and HPV 16/HPV 18 genotypes versus other high-risk types (Table 2).

Progressive multivariable modeling demonstrated incremental improvement in discrimination across domains (Figure 2C), with the C-index increasing from 0.516 (model 1) to 0.538 (model 2), 0.562 (model 3), and 0.619 (model 4). The overall model fit for the final model was significant (likelihood ratio χ² = 73.0; df = 16; p < 0.001) (Table 3). In the fully adjusted model 4 (Table 3; Figure 2B), independent predictors were persistent HPV infection (aHR 2.51, 95% CI 1.99–3.16), HPV unvaccinated status (aHR 1.54, 95% CI 1.08–2.20), multiple margins positive (aHR 1.52, 95% CI 1.08–2.14), AIS versus CIN2 (aHR 1.48, 95% CI 1.03–2.12), previous cervical treatment (aHR 1.38, 95% CI 1.04–1.84), single margin positive (aHR 1.38, 95% CI 1.02–1.87), and neutrophil-to-lymphocyte ratio (per one-unit increase: aHR 1.21, 95% CI 1.02–1.44) (Table 3). In contrast, several covariates associated in univariable analyses did not retain independent statistical significance after full adjustment, including multiparity (aHR 1.14, 95% CI 0.92–1.42), postmenopausal status (aHR 1.15, 95% CI 0.86–1.54), and current/former smoking (aHR 1.35, 95% CI 0.99–1.84), as well as HPV 18 genotype (aHR 1.33, 95% CI 0.99–1.79) and HPV 16 genotype (aHR 1.14, 95% CI 0.89–1.46) (Table 3).

Subgroup analyses indicated that the association between persistent HPV infection and recurrence was consistent across clinically relevant strata (Table 4; Figure 2D). The adjusted hazard ratio was similar in women aged <40 years (2.46, 95% CI 1.84–3.28) and ≥40 years (2.58, 1.84–3.61) and remained robust across margin-burden strata (no positive margins: 2.82, 1.71–4.65; single positive: 2.41, 1.75–3.32; multiple positive: 2.48, 1.68–3.67) (Table 4). Effects were also comparable by vaccination status (unvaccinated: 2.56, 2.00–3.27; any vaccination: 2.21, 1.12–4.35) and smoking status (never: 2.48, 1.94–3.17; current/former: 2.66, 1.48–4.79), with no evidence of effect modification (all interaction p-values >0.05) (Table 4).

Kaplan–Meier curves demonstrated pronounced separation by persistent HPV status (HR 2.87, 95% CI 2.31–3.57; log-rank p < 0.001) (Figure 3B), as well as by HPV vaccination status (HR 1.89, 95% CI 1.33–2.68; log-rank p < 0.001) (Figure 3D). Recurrence-free survival also differed by margin involvement status (log-rank p = 0.007) (Figure 3C). Numbers at risk over follow-up for each stratum are presented in Figure 3. A clinical nomogram incorporating the final multivariable predictors was constructed (Figure 4A). Calibration assessment demonstrated an acceptable agreement between predicted and observed recurrence risks, with a calibration slope of 0.96 and a Brier score of 0.098 at 24 months, indicating limited overall prediction error (Figure 4B). Using nomogram-derived risk stratification, patients were categorized into low-, moderate-, and high-risk groups (n = 892, n = 856, and n = 482, respectively), with corresponding observed 24-month recurrence rates of 6.2%, 14.8%, and 31.5% (p for trend <0.001) (Figure 4E).