Recurrence in Eyelid Sebaceous Carcinoma: A Multicentric Study of 418 Patients

Abstract

Purpose: Local recurrence predicts dismal prognosis in eyelid sebaceous carcinoma (SC). Recurrence predictors vary across studies. Accurate recurrence estimation is essential for individualized therapy in eyelid SC. This study aims to identify recurrence predictors and develop a nomogram for personalized prediction in eyelid SC.

Methods: We conducted a multicenter retrospective cohort study. Chart reviews were performed in 418 consecutive patients with eyelid SC. All patients were followed up after their initial surgery. Multivariate Cox regression was used to explore the independent predictors of recurrence. A nomogram for recurrence prediction was developed and validated with bootstrap resampling. The predictive accuracy and discriminative ability were compared with the Tumor, Node, Metastasis (TNM) staging system.

Results: Over a median of 60-month follow-up, 167 patients (40%) had local recurrence. The median time from diagnosis to recurrence was 14 months. The 1-year cumulative recurrence rate was 18%. Diagnostic delay (hazard ratio [HR] = 1.01, 95% confidence interval [CI] = 1.00–1.01, P = 0.001), orbital involvement (HR = 4.47, 95% CI = 3.04–6.58, P < 0.001), Ki67 (HR = 1.01, 95% CI = 1.00–1.02, P = 0.008) and initial surgery of Mohs micrographic surgery with intraoperative frozen section control (HR = 0.53, 95% CI = 0.35–0.80, P = 0.003) were independent influencing factors of recurrence. A nomogram integrating these four factors combined with pagetoid spread displayed satisfactory discriminative ability (C-index = 0.80–0.83; area under the curve [AUC] = 0.82–0.84), which compared favorably than TNM staging (all P < 0.05).

Conclusions: The recurrence rate is high in eyelid SC. Early detection and primary resection with Mohs micrographic surgery are recommended in controlling recurrence. Patients with orbital involvement, high Ki67 expression, and pagetoid spread may require adjuvant measures. This nomogram offers more accurate recurrence estimates, aiding in therapeutic decision making.

Eyelid sebaceous carcinoma (SC), an aggressive malignancy originating from sebaceous glands, comprises 0.2% to 4.7% of all eyelid malignancies in the United States.^1–3 However, it is more prevalent in Asia,^4–7 accounting for 33% to 42% of all eyelid malignancies in Chinese people.^6,7 Eyelid SC at early stage is insidious and often presents with inflammation signs, leading to delayed diagnosis and managements.^8,9 Complete resection is the primary treatment for eyelid SC.¹⁰ Unfortunately, recurrence remains a great challenge after surgery, which varies among different regions, ranging from 6.3% to 37%.^{1,2,5,9,11–16} 

The reported predisposing factors for recurrence include the involvement of both upper and lower eyelids, multicentric origin, large tumor size, diffuse growth pattern, conjunctival involvement, and advanced T category (>T3a).^11,17 Risk factors for recurrence are multifaceted, therefore, identifying key predictors to develop an individualized risk assessment model is crucial to optimize treatment. 

The American Joint Committee on Cancer (AJCC) staging system is the most widely used system for eyelid SC risk stratification.¹⁸ However, several important clinical parameters are not considered in this system, such as pathological parameters, including pagetoid intraepithelial neoplasia and histology differentiation,^15,16,19 demographic factors, medical history, and surgical modalities. Given the need to weigh risks and benefits for each patient, to develop an individualized scoring system that accurately estimate recurrence is essential to guide therapeutic decisions in clinical practice. 

In this multicenter retrospective study, we aim to explore the independent predictors of recurrence in eyelid SC. Then, we develop and validate a nomogram risk scoring system for individualized prediction of recurrent SC. 

Methods

Patients and Design

The requirement for informed consent was waived by the institutional review board (IRB) for its retrospective, noninterventional nature. This research followed the tenets of the Declaration of Helsinki. A centralized IRB Review Process was used in this multicenter study. This study was approved by the IRB of the lead unit, Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine, and the other two sites accepted the review of the central IRB (No. SH9H-2019-T185-2, Supplementary File S1). We identified 496 patients pathologically diagnosed with eyelid SC who were admitted to Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine, Xiangya Hospital affiliated with Central South University, and the Second Affiliated Hospital of Zhejiang University School of Medicine between January 1991 and September 2017. The detailed procedures of this study were described previously by our group (Fig. 1).^19–21 We contacted the patients or their relatives and explained the purpose of the study. They participated in this study voluntarily without any additional compensation. They were inquired about the conditions after discharge. The vital status was also confirmed via mandatory Chinese resident registry. For patients who died, the primary cause of death was documented by checking the death certificate in Chinese Center for Disease Control. Among the 496 patients, 39 patients were not reached, 19 declined participation, and 20 were ruled out due to incomplete data collection, resulting in a final sample size of 418 patients. The number of patients recruited from Shanghai Ninth People's Hospital, the Second Affiliated Hospital of Zhejiang University School of Medicine, and Xiangya Hospital were 246, 96, and 76, respectively. The recruited patients were divided into 2 groups at a ratio of 7:3 using computer-generated random numbers; 293 patients were allocated to the training cohort, and 125 patients were allocated to the validation cohort. 

Figure 1.

 

Patient flow chart. The procedures of follow ups.

Data Collection

Medical charts and pathologic data of each patient were reviewed. The demographics, clinical characteristics, treatment modalities, and final outcomes at the follow-up were collected (Supplementary File S2). The demographics included gender and age at the diagnosis. Clinical parameters consisted of diagnostic delay (the duration from the onset of symptoms to pathological diagnosis of eyelid SC), the presence of second/third primary tumor, initial referral diagnosis, tumor location, greatest tumor basal diameter, the presence of pagetoid spread (confirmed by multiple map biopsy), perineural invasion and muscle infiltration, degree of histology differentiation, TNM staging (8th edition of the AJCC staging system for eyelid SC¹⁸), surgical approaches, and adjuvant treatments. The degree of differentiation was subdivided as described elsewhere,²² and well-differentiated tumors were characterized by lobules with sebaceous differentiation. Moderate differentiation was mainly constituted by several areas of highly differentiated sebaceous cells and anaplastic cells. Poor differentiated tumors were formed of prominent nucleoli, pleomorphic nuclei, and amphophilic-positive cytoplasm. The primary surgical approaches included wide local excision (WLE) without frozen section control and Mohs micrographic surgery (MMS) with frozen section control. A total of 195 patient had received treatment elsewhere after their initial diagnosis, the prior clinical details and pathological sections were retrieved for review. Clinical outcomes included local recurrence, metastasis, and death. Local recurrence was defined as new SC at the original tumor site or at any ocular site, and the time interval to the last tumor resection ≥3 months. The time of recurrence and metastasis was the date on which the dissemination was confirmed by clinical examination, imaging, or biopsy. The interval from initial diagnosis to the first recurrence was calculated. Moreover, the locations of metastasis were recorded. 

Statistical Analysis

All analyses were performed using SPSS 22.0 (IBM Corp, Armonk, NY, USA) and the “rms” and “glmnet” packages in R 3.4.1 (Vienna, Austria; http://www.R-project.org/). Frequency (percentage) and median (interquartile range) were applied to describe categorical and continuous variables, respectively. Medians and proportions were compared using the nonparametric Mann-Whitney U test and the chi-square test (or Fisher's exact test, if appropriate), respectively. 

The independent risk factors for recurrence were explored. First, to identify the possible correlates of recurrence, we compared the demographic and clinical indicators between the patients with and without recurrence using univariate Cox regression. The significant factors (P < 0.05) were put into multivariate Cox regression models as independent variables after adjusting for the collinearity among variables with a correlation matrix. The regression coefficients and hazard ratios (HRs) with 95% confidence intervals (CIs) were calculated. The cumulative incidence of recurrence was investigated by Kaplan-Meier survival curve and compared with a log rank test. All tests were two-sided, and a P value < 0.05 was considered statistically significant. 

In general, approximately 20 observations per variable are required to produce reasonably stable estimates.²³ The training sample size was 293; thus, the number of selected variables included in the model should be no greater than 14. Least absolute shrinkage and selection operator (LASSO) regression was applied to select the parameters for recurrent SC from the primary data set. Lambda was selected via 10-fold cross-validation to minimize the mean cross-validation partial likelihood error rate. 

A nomogram was developed based on the LASSO regressions for recurrence. The discriminative ability of this model was evaluated by C-index and the area under the curve (AUC) in receiver operating characteristic (ROC) curves. Calibration curves of the nomogram were examined to assess the agreement between the predicted and observed outcomes. Comparisons between the nomogram model and the TNM staging systems were performed with the “rcorrcens” function in the “Hmisc” package in R. Furthermore, the nomogram was externally validated with bootstrap resampling in an independent validation cohort of 125 patients. A decision-curve analysis (DCA) was used to compare the clinical net benefit associated with the use of this nomogram and the TNM staging system.²⁴ 

Results

Descriptive Statistics

A total of 418 patients were included, among them, 188 (45%) were men, and 230 (55%) were women. The median age at diagnosis was 63 years old (interquartile range [IQR] = 53–73 years), ranging from 18 to 96 years. The median diagnostic delay was 12 months (IQR = 6.0–30 months). Of note, 17 (4.1%) patients had second or third primary tumors, including breast cancer, gastric carcinoma, bladder carcinoma, colon cancer, gingival carcinoma, laryngocarcinoma, tongue cancer, esophageal cancer, prostate cancer, and pancreatic adenocarcinoma. At initial diagnosis, as much as 159 (38%) patients were misdiagnosed as having other diseases, among which, blepharitis (46, 11%) was the second most common referral diagnosis. Eyelid SC could invade into adjacent ocular tissues, such as bulbar conjunctiva (36, 8.6%), caruncle (40, 9.6%), and orbital structures (58, 14%). In this cohort, 128 (31%) patients underwent MMS with intraoperative frozen section control for primary resection, and 290 (69%) patients underwent WLE without intraoperative frozen section control as the initial surgery. At the time of diagnosis, 24 (5.7%) patients exhibited regional nodal metastasis, and none of them had distant metastases. The demographic and clinical characteristics of these two groups were compared (Table 1). Significant differences were noted between these two groups regarding age at diagnosis (P = 0.014), diagnostic delay (P < 0.001), initial diagnosis (P = 0.006), bulbar conjunctival involvement (P < 0.001), caruncular involvement (P = 0.006), orbital involvement (P < 0.001), greatest tumor basal diameter (P < 0.001), pagetoid spread (P < 0.001), perineural invasion (P < 0.001), muscle infiltration (P < 0.001), Ki67 expression (P < 0.001), histology differentiation (P = 0.003), initial surgical modality (P < 0.001), and T categories (P < 0.001). 

Table 1.

 

Baseline Demographic and Clinical Characteristics for Patients With Eyelid Sebaceous Carcinoma

 

Recurrence

After a median follow-up period of 60 months (IQR = 30–98 months), 167 patients (40%) had local recurrence, of which 50 (12%) patients experienced 2 or more recurrences, and 55 (13%) participants were afflicted with orbital recurrence. The median duration from diagnosis to the initial recurrence was 14 months (IQR = 9.0–24 months). By Kaplan-Meier estimates, the 1-year cumulative recurrence rate was 18%. Out of the 167 relapsed cases, 58 (35%) had metastasis, including 55 (33%) patients that were afflicted with nodal metastasis and 31 (19%) who suffered from distant disseminations. In patients with recurrence, 59 (35%) patients died, with 39 (23%) patients that died of metastatic SC. Of note, nonrecurrent participants had significantly better prognosis (metastasis = 30 [12%]; nodal metastasis = 30 [12%]; distant metastasis = 13 [5.2%]; all-cause death = 44 [18%]; and tumor-related death = 17 [6.8%]). 

In univariate analysis, age at diagnosis (P = 0.041), diagnostic delay (P < 0.001), initial diagnosis other than eyelid SC (P < 0.001), bulbar conjunctival involvement (P < 0.001), caruncular involvement (P = 0.001), orbital involvement (P < 0.001), greatest tumor basal diameter (P < 0.001), pagetoid spread (P < 0.001), perineural invasion (P < 0.001), muscle infiltration (P < 0.001), Ki67 expression (P < 0.001), histology differentiation (P < 0.001), initial surgical modality (P < 0.001), T categories (P < 0.001), and lymph node metastasis (P = 0.003) were potential risk factors for recurrence (Table 2). The collinearity among baseline parameters was assessed with a correlation matrix (Supplementary Table S1); when a correlation was identified, only the most clinically relevant parammeter were put into subsequent multivariable model. By collinearity tests, six parameters, namely greatest tumor basal diameter, pagetoid spread, perineural invasion, muscle infiltration, histology differentiation, and T categories were excluded from the multivariable analysis. To determine the independent predictors, multivariate Cox regression analyses were performed with the remaining factors, and the results demonstrated that diagnostic delay (HR = 1.01, 95% CI = 1.00–1.01, P = 0.001), orbital involvement at initial diagnosis (HR = 4.47, 95% CI = 3.04–6.58, P < 0.001), Ki67 expression (HR = 1.01, 95% CI = 1.00–1.02, P = 0.008), and initial surgical modality with MMS (HR = 0.53, 95% CI = 0.35–0.80, P = 0.003) were independent influencing factors for recurrence (see Table 2). 

Table 2.

 

Uni- and Multivariable Cox Proportional Hazards Regression Analysis for the Predictors of Local Recurrence

 

Nomogram for Recurrence: Internal Validation

The demographic and clinical characteristics of the training and validation cohorts are summarized in Supplementary Table S2. No statistical significance was noted between these two cohorts regarding age, sex, medical history, baseline clinical characteristics, treatments, and final recurrence (P = 0.112–0.962), except that the validation cohort included more patients afflicted with second primary tumors than the training cohort (P = 0.008). After similar follow-up periods (training cohort = 60.0 months, and validation cohort = 55.0 months, P = 0.670), a total of 115 (39%) and 52 (42%) patients had relapsed eyelid SC. 

LASSO penalization was performed for feature selection, and the cross-validated error plot of the LASSO model is shown in Supplementary Figure S1A. A predictive nomogram integrating 5 potential predictors derived from 15 features using LASSO regression (Supplementary Fig. S1B) was constructed (Fig. 2A). Elements contained in this nomogram included diagnostic delay, orbital involvement, Ki67 expression, initial treatment modality, and the presence of pagetoid spread. The final model for recurrence showed strong internal validity, with a discrimination C-index of 0.83. The bias-corrected C-index generated by bootstrap resampling was 0.82, indicating the excellent discriminative ability of this model. The calibration plot displayed excellent agreement between the predictions and actual observations for the training cohort (Fig. 2B). The standardized total scores generated by the nomogram for each patient in the training cohort are shown in Supplementary Figure S2. 

Figure 2.

 

Nomogram for recurrence. (A) Nomogram for predicting probability of recurrence after primary surgery. To use it, locate orbital involvement (yes/no), draw a vertical line up to the “Points” axis to obtain the score of this parameter. Repeat for the other 4 variables: diagnostic delay (months), pagetoid spread (yes/no), Ki 67(%), and initial treatment with Mohs surgery (yes/no). Then, the scores are summed and locate the total number on the line labeled “Total Points.” Draw a vertical line downward from the total point dot to determine the chances of relapsed sebaceous carcinoma at the intersection with the recurrence probability axis. Calibration of the nomogram for recurrence for the training (B) and validation (C) cohorts. Nomogram-predicted survival probability is plotted on the x-axis, with observed survival probability on the y-axis. Dashed line through the origin point represents the perfect calibration models in which the predicted probabilities are identical to the actual probabilities. Dotted line (without intersection with the origin point): predicted probabilities based on nomogram; Solid line: bootstrap corrected estimates. B = 300 repetitions for bootstrap.

Nomogram for Recurrence: External Validation

The predictive nomogram for recurrence after primary resection was applied to an independent set of 125 patients for external validation. This recurrence prediction model, as tested in the validation cohort, showed an uncorrected C-index of 0.80 and a bootstrap-corrected C-index of 0.79 with 300 bootstrap resamplings. The calibration plot indicated good agreement between the nomogram predictions and actual observations for recurrence (Fig. 2C). 

Comparison of the Nomogram With the TNM Staging System

The discriminative accuracy of the TNM staging system for predicting recurrence was not satisfactory, with C-indices of 0.67 and 0.71 for the training and validation sets, respectively. The details of the comparisons are summarized in Table 3. This nomogram showed a significant improvement over the TNM staging system in terms of predicting recurrence (all P < 0.05). In the training cohort, the AUC of the nomogram was significantly higher than that of TNM (0.84 vs. 0.67, P < 0.001; Supplementary Fig. S3A), whereas in the validation cohort, this trend continued (0.82 vs. 0.71, P = 0.020; Supplementary Fig. S3B). Furthermore, patients were stratified into three risk groups (low, intermediate, and high) based on nomogram scores. Cumulative recurrence was significantly different across risk groups (Fig. 3A; log rank P < 0.001 (overall); P < 0.001 (low versus intermediate risk); and P < 0.001 (intermediate versus high risk). In addition, Kaplan-Meier curves were generated based on the T category of the TNM staging. For the combined cohort, the prognostic stratification was satisfactory, however, the overlap of the survival curves for patients at early stages was notable (Fig. 3B; log rank P < 0.001 [overall]; log rank P = 0.594 [T1 versus T2]; P = 0.883 [T2 versus T3]; and P < 0.001 [T3 versus T4]). 

Table 3.

 

The Predictive Discrimination Ability of the Nomogram Compared With the TNM Staging System in the Training and Validation Cohort

 

Figure 3.

 

(A) Kaplan-Meier curves of recurrence-free survival for low-, moderate-, and high-risk groups stratified by nomogram score (log rank P < 0.001 [overall]; P < 0.001 (low versus intermediate risk); and P < 0.001 [intermediate versus high risk]). (B) Kaplan-Meier curves of tumor-related survival according to the tumor (T) category of the Tumor, Node, Metastasis (TNM) staging system according to the 8th edition of the American Joint Committee on Cancer (AJCC; log rank P < 0.001 [overall]; log rank P = 0.594 [T1 versus T2]; P = 0.883 [T2 versus T3]; and P < 0.001 [T3 versus T4]). Log rank P value < 0.05 was considered statistically significant.

Clinical Benefit Analysis

The clinical benefit-centered accuracy generated by this nomogram and TNM staging system were compared by DCA. Clearly, the nomogram was superior to traditional TNM staging as well as assuming treating all patients (Supplementary Fig. S4). 

Discussion

The risk of recurrence remains high after surgical resection of eyelid SC. Independent predictors of recurrence included diagnostic delay, orbital involvement at initial diagnosis, Ki67 expression, and initial surgery of WLE without intraoperative frozen section control. Furthermore, we developed and validated a novel nomogram to estimate the probability of recurrence. The five variables included in this model are clinically relevant and easily available. This nomogram provided more accurate individualized risk estimates for recurrence than conventional TNM staging system, which could guide clinicians in their therapeutic decision making. 

Local recurrence predicts dismal prognosis in eyelid SC, and presents a major challenge in clinical practice. In our study, after a median follow-up of 60.0 months, approximately 4 in 10 patients suffered from local recurrence. This recurrence rate observed in our study is higher than those reported in America (41 months = 18%¹¹; 5 years = 8.1%¹; 5 years = 33%¹⁵; 5.9 years = 36%²; and 80 months = 19%¹⁶), England (4.8 years = 9.7%¹⁴; and 40 months = 14%⁹), India (29 months = 24%¹²), and Japan (4.2 years = 6.3%¹³). The high recurrence rate in Chinese patients might be partially attributed to patient selection bias. The three tertiary medical centers in our study tended to include more patients with advanced SC, whereas, at initial presentation, as many as 58 (14%) patients presented with orbital involvement. Another important attribute to recurrence discrepancy is the different baseline of follow-ups. The baseline of this study was set as the initial diagnosis of eyelid SC, whereas some studies calculated follow-up time from the referral to their hospitals. Among all 418 patients, 195 of them had received primary tumor resection elsewhere, and 129 of them came for tumor recurrence. Whereas of the 223 patients who were admitted to these 3 medical settings for primary resection, 38 (17%) of them had recurrence, and this relapse rate was comparable to other studies (4.12 years = 17%²⁵ and 4 years = 18%²⁶). 

According to our analyses, the patients initially treated with MMS with intraoperative frozen section control had significantly lower chances of recurrence than their counterparts who received WLE without intraoperative frozen section control. A total of 128 patients received MMS as the primary surgery modality; among them, 28 (22%) exhibited recurrence. The 1-, 3-, and 5-year cumulative recurrence rates for the MMS group were 8.6%, 23%, and 24%, respectively. The median time interval from diagnosis to the initial recurrence was 15.5 months (IQR = 10.5–30.0 months). In contrast, of 290 patients initially treated with WLE, 139 (48%) had relapsed eyelid SC. The median duration to initial recurrence was 13.0 months (IQR = 9.0–24.0 months). By Kaplan-Meier estimates, the 1-, 3-, and 5-year cumulative recurrence rates for the WLE group were 21%, 44%, and 48%, respectively. These recurrence rates of the WLE group were significantly higher than those who received MMS (log rank P < 0.001). The Kaplan-Meier curves stratified by initial surgery modalities are displayed in Supplementary Figure S5. Similar findings were observed in patients without orbital involvement.²⁰ Spencer et al. reported a recurrence rate of 11% among 18 patients with eyelid SC treated with MMS after a mean follow-up of 3.1 years.²⁷ Tolkachjov et al. found 7 recurrences among 109 cases (6.4%) treated with MMS.²⁸ In another report, Sa H et al. reported a local recurrence rate of 6.0% among 100 patients after a median follow-up of 31.5 months.²⁹ Therefore, MMS with intraoperative frozen section control is recommended for primary tumor resection in eyelid SC. 

Delayed diagnosis is common in eyelid SC as it often mimics benign lesions that confound clinicians.¹⁰ Kaliki S et al. reported that the interval from symptoms to confirmed diagnosis longer than 6 months increased the chances of lymph node metastasis by 80%.¹² This finding is consistent with our study. In our cohort, as many as 159 (38%) patients were initially misdiagnosed with an average delay of 30.8 months. Blepharitis is the most common misdiagnosis, which were reported in 46 patients (11%). Consequently, physicians should distinguish SC from disguised blepharoconjunctivitis, especially for elderly patients who experienced routine treatment failure.³⁰ 

Orbital involvement was also an independent risk factor for SC recurrence. Rao et al. conducted a retrospective study of 104 patients with SC, and 34 (33%) patients had recurrence. Among these relapsed cases, 17 of them had orbital involvement.¹⁵ Ki67, a nuclear antigen highly relevant to advanced disease and metastasis, is commonly used as a proliferation marker of tumor.³¹ A recent study conducted by Sa H et al. analyzed 100 consecutive cases with eyelid SC, and 6 patients (6.0%) exhibited recurrence. According to their findings, orbital involvement and pagetoid intraepithelial neoplasia strongly correlated with an elevated risk of recurrence.²⁹ Takahashi Y et al. performed a retrospective study of 34 patients with eyelid SC. After an average follow-up time of 43.7 months, they found that the involvement of both upper and lower eyelids, topical treatments at other clinics, multicentric origin, diffuse pattern, stage T3a, large tumor size, and a nonlobular pattern were risk factors for recurrence.¹⁷ Potential explanations for the variations in risk factors may relate to different sample sizes, various follow-up periods, racial differences, and multiple surgeons with diverse experiences. 

The predictors of recurrence are multifactorial, consists of patient-, tumor-, and treatment-related factors. Our study covered these aspects and derived an individualized model to predict recurrence with a rigorous statistical methodology. A nomogram is a graphic scoring system based on statistical models to increase the predictive accuracy for individuals. This model, consisting of simple-to-collect clinical parameters, is beneficial to guide therapeutic decisions and optimize patient risk stratification for clinical trial design, which is of great significance to both patients and clinicians. In addition, this model displayed satisfactory discriminative ability, and the predictive score was externally replicated in another independent cohort. This reproducibility indicated the satisfactory generalizability of this nomogram. Notably, this predictive scale exhibited superior discrimination accuracy and achieved a higher net benefit than the traditional TNM staging system. Therefore, we propose this scoring system as a new standard to guide management for patients with eyelid SC. 

This study should be regarded as an initial attempt to apply the nomogram for recurrence prediction in ocular tumors. However, caution should be taken when interpreting the findings due to a number of limitations. First, all patients were recruited from tertiary hospitals, which may introduce selection bias and limit the generalization of our findings to a broader patient population. Another limitation may relate to the retrospective nature of our study. Nevertheless, the use of rigorous statistical methodology and adjustment of interactions enabled us to maximally eliminate bias. Additionally, to our knowledge, our cohort has the largest sample of eyelid SC to date, and the long follow ups made our results more reliable. 

In conclusion, the recurrence rate is high in eyelid SC, in which approximately 4 in 10 patients experienced local recurrence. Of note, diagnostic delay, orbital involvement at diagnosis, Ki 67 expression, and initial surgical modality of WLE without intraoperative frozen section control were independent risk factors for recurrence. Moreover, we established a novel, robust, and straightforward predictive model for recurrence after primary resection and demonstrated better discriminative ability than the traditional TNM staging system. To our knowledge, this is the first exploration to apply a nomogram for recurrence prediction in ocular diseases. Nevertheless, future studies are needed to fully validate our findings. 

Acknowledgments

Supported by the Shanghai Science and Technology Development Foundation (Grant No. 22QA1407500), and the Shanghai Rising Stars of Medical Talent Youth Development Program (Grant No. SHWSRS [2022-65]). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. 

Author Contributions: C.D.Z., J.Y., J.T., and H.Y.H. designed study. J.Y. and J.T. provided the data of patients. C.D.Z., M.P.X., Q.C., and Y.X.L. conducted the literature search. C.D.Z., M.P.X., and Q.C. designed the figures. M.P.X., Q.C., and Y.X.L. conducted the data collection and drafted the manuscript. P.W.C., H.Y.H., and X.Y.H. did data analysis and interpretation. M.P.X., Q.C. and Y.X.L. produced tables and figures. Q.C. conducted the proofreading of the manuscript. M.P.X., Q.C., and Y.X.L. contributed equally to this paper. C.D.Z., J.Y., J.T., and H.Y.H. are the co-corresponding authors. All authors approved this manuscript. 

Data Sharing Statements: The datasets used and analyzed during the current study are available from the corresponding authors on reasonable request. 

Ethics Approval and Consent to Participate: All procedures performed in studies involving human participants were in accordance with the ethical standards of the Ethics Committee of the Institutional Ethical Review Board of Shanghai Ninth People's Hospital. 

Disclosure: M. Xu, None; Q. Chen, None; Y. Luo, None; P. Chai, None; X. He, None; H. Huang, None; J. Tan, None; J. Ye, None; C. Zhou, None