A Nomogram for Preoperatively Predicting the Ki-67 Index of a Pituitary Tumor: A Retrospective Cohort Study

Background The Ki-67 index is an indicator of proliferation and aggressive behavior in pituitary adenomas (PAs). This study aims to develop and validate a predictive nomogram for forecasting Ki-67 index levels preoperatively in PAs. Methods A total of 439 patients with PAs underwent PA resection at the Department of Neurosurgery in Jinling Hospital between January 2018 and October 2020; they were enrolled in this retrospective study and were classified randomly into a training cohort (n = 300) and a validation cohort (n = 139). A range of clinical, radiological, and laboratory characteristics were collected. The Ki-67 index was classified into the low Ki-67 index (<3%) and the high Ki-67 index (≥3%). Least absolute shrinkage and selection operator algorithm and uni- and multivariate logistic regression analyses were applied to identify independent risk factors associated with Ki-67. A nomogram was constructed to visualize these risk factors. The receiver operation characteristic curve and calibration curve were computed to evaluate the predictive performance of the nomogram model. Results Age, primary-recurrence subtype, maximum dimension, and prolactin were included in the nomogram model. The areas under the curve (AUCs) of the nomogram model were 0.694 in the training cohort and 0.658 in the validation cohort. A well-fitted calibration curve was also generated for the nomogram model. A subgroup analysis revealed stable predictive performance for the nomogram model. A correlation analysis revealed that age (R = −0.23; p < 0.01), maximum dimension (R = 0.17; p < 0.01), and prolactin (R = 0.16; p < 0.01) were all significantly correlated with the Ki-67 index level. Conclusions Age, primary-recurrence subtype, maximum dimension, and prolactin are independent predictors for the Ki-67 index level. The current study provides a novel and feasible nomogram, which can further assist neurosurgeons to develop better, more individualized treatment strategies for patients with PAs by predicting the Ki-67 index level preoperatively.


INTRODUCTION
Pituitary adenomas (PAs), which account for approximately 15% of primary intracranial neoplasms, are the most common tumors in the sellar region (1). PAs were recently renamed pituitary neuroendocrine tumors (PitNET), because although they are mostly benign, PAs could present invasive, aggressive, and metastatic behaviors (2). The Ki-67 labeling index is a proliferative marker for PAs (3,4). A Ki-67 ≥3% is an indicator for invasive growth according to the World Health Organization 2004 classification (5). The French five-tiered classification is another approach for classifying PAs with validated prognostic value, and Ki-67 is used as one of the proliferative markers (2).
The preoperative prediction for the Ki-67 index is valuable and may affect the patient's surgical approach and postoperative management. A high Ki-67 index requires neurosurgeons to obtain gross-total resection and prepare for higher probability of compression and tissue adhesion in the surrounding structures (6). Furthermore, a high Ki-67 index may prompt a closer follow-up or early radiation therapy (7). Ugga et al. proposed a machine learning model based on T2W magnetic resonance imaging (MRI) for the prediction of the Ki-67 proliferation index class (8). However, how to integrate machine learning model into the routine practices have not been assessed so far (9). Obviously, applying this machine learning model in an actual clinical practice requires supporting software and platform, which must be developed in the future. Conficoni et al. suggested that quantitative measures of apparent diffusion coefficient (ADC) values could predict the Ki67 value (10). However, this value was not routinely assessed in clinical practice.
A nomogram, which is an easy-to-use and graphical predictive tool, has been widely applied to predict numerous binary and prognostic outcomes (11). However, a nomogram for preoperatively forecasting the Ki-67 level in patients with PAs remains unavailable. Therefore, in this retrospective study, we aimed to develop and validate the first predictive nomogram for preoperatively predicting Ki-67 levels in patients with PAs.

Study Design and Patients
We reviewed the clinical records of patients with PAs who underwent PA resection at the Department of Neurosurgery in Jinling Hospital between January 2018 and October 2020. The inclusion criteria were as follows (1): histopathologically confirmed PA (2); patients who underwent PA resection via transcranial or transsphenoidal approaches using microscope or endoscope (3); patients with a Ki-67 index identified from histopathological examination results; and (4) patients who had at least one collected variable. The exclusion criteria included the following (1): patients who had no histopathological examination; and (2) patients who had no collected variables. Our institutional research ethics board approved this retrospective study (2016NZKY-008-02). Because of the nature of the retrospective cohort study and the anonymization of data prior to analysis, informed consent was waived.

Data Collection
Overall, there were 93 variables collected in this study. Baseline characteristics included age, gender, primary-recurrence subtype, treatment history for PAs (e.g., medication, surgery, and radiotherapy), and preoperative signs and symptoms (e.g., moon face, acromegalia, headache, visual impairment, and visual field defect). Radiological features that were collected included maximum dimension of tumor, Knosp grade, Hardy grade, multiple lesions, optic nerve compression, and pituitary apoplexy. Grades 0 to 2 and grades 3 to 4 were classified into noninvasive and invasive classes, respectively, for Knosp grade (12) and Hardy grade for sellar invasion (13). We also collected 74 preoperative laboratory tests, including pituitary hormones, routine blood, coagulation, renal and hepatic functions, and electrolytes, which were based on peripheral blood samples (Table S2). Included samples were also diagnosed with clinical subtypes including nonfunctioning, growth hormone (GH) secreting (14,15), prolactin (PRL) secreting (15), and adrenocorticotropic hormone (ACTH) secreting (16) PAs. Ki-67 index was extracted as postoperatively assessed outcome from the histopathological examination and classified into a low Ki-67 index (<3%) and a high Ki-67 index (≥3%).

Development and Validation of the Nomogram
First, the patients were randomly divided into a training cohort and a validation cohort. Second, least absolute shrinkage and selection operator (LASSO) algorithm was applied to filter features with a missing data percentage of <60% using glmnet R package (version 4.1). During the LASSO analysis, mean imputation for missing data was applied. Missing data were not imputed in the following analysis to simulate the model performance in real-world conditions. Uni-and multivariate logistic regressions were utilized to determine independent risk factors associated with the Ki-67 level by use of an rms R package (version 6.1.0). In this step, variables have to meet at least one of the following criteria to be included in multivariate logistic regression analysis: displaying a significant difference between the low and high Ki-67 index groups; avoiding being filtered out in the LASSO analysis; or being shown to be a significant predictor in univariate logistic regression analysis. Finally, a nomogram was constructed to visualize the risk factors. The predictions for the validation cohort were calculated using rms R package. The receiver operation characteristic (ROC) curve and the calibration curve were computed separately using pROC (version 1.17.0.1) and rms R packages to evaluate the predictive performance of the nomogram model. We assessed the model performance in the subgroups according to age, gender, primaryrecurrence subtype, maximum dimension, clinical subtype, Knosp and Hardy grades, and variables included in the final nomogram. The cutoff values of the continuous variables in the subgroup analysis were the mean values in the validation cohort.

Sample Size
There were no generally accepted approaches for sample size estimation in the development and validation studies of risk prediction models. However, based on the events per variable (EVP) = 10 criteria (17), the event number in the training cohort should exceed 10 the number of variables included in the multivariate logistic regression analysis. After the filtering process, although there were 21 variables included in the following analysis, we only analyzed all possible combinations up to seven variables in the multivariate logistic regression analysis. Because the current study included 128 events in the training cohort, the sample size was sufficient for analysis in this research.

Statistical Analysis
Model development and validation were performed according to "Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis" (TRIPOD) guidance (Table  S1) (18). The R software (version 3.6.0) was applied for statistical analysis, and statistical significance was set at p < 0.05. Continuous variables were presented as the mean ± standard deviation (SD). Comparisons between two continuous variables were evaluated using Student's t-test. The chi-squared test or Fisher's exact test was used for comparisons of categorized variables. Spearman correlation analysis was applied to evaluate two continuous variables, and the data were visualized using "ggplot" (version 3.3.3) R packages.

Baseline Clinical Characteristics of Participants
A total of 439 eligible patients, who were included in the study, were randomly divided into a training cohort (n = 300) and a validation cohort (n = 139). Detailed characteristics for these two cohorts showed homogeneity in these cohorts ( Table S2). The clinical information of patients with PAs in the high Ki-67 cohort and in the low Ki-67 cohort, which was summarized in Table 1 and Table S3, revealed remarkable differences in the following variables: age, clinical subtype, Hardy grade for suprasellar extension, history of pituitary surgery, multiple lesions, maximum dimension, luteinizing hormone (LH), folliclestimulating hormone (FSH), free triiodothyronine (FT3), Creactive protein, red blood cell (RBC) count, creatinine, potassium, procalcitonin (PCT), lymphocyte percentage, urea, mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), fibrin/fibrinogen degradation products, thrombocytocrit, and platelet counts.

Filtering Process for Collected Variables
Age, primary-recurrence subtype, history of pituitary surgery, clinical subtype, Hardy grade for suprasellar extension, maximum dimension, prolactin, LH, FSH, T3, FT3, RBC count, potassium, PCT, MCV, MCH, thrombocytocrit, and platelet counts were obtained as the primary predictive factors with p < 0.05 in a univariate logistic regression analysis ( Table  S4). The results showed that age, LH, FSH, PCT, MCV, and MCH were protective factors for high Ki-67 level, whereas the other primary predictive factors were risk factors. As shown in

Determination of Independent Predictors for the Ki-67 Level
The variables filtered using univariate logistic regression analysis and LASSO analysis were combined and further analyzed using multivariable logistic regression analysis. Finally, as shown in

Construction and Validation of Nomogram Prediction Model
The multivariate model was visualized as a nomogram ( Figure 2). To apply the nomogram, users should draw a virtual vertical line from each variable to the "Points" axis to identify the points attributed by each variable. Then, users need to compare the summed points with the bottom scale to assess the probability of a high Ki-67 index. The areas under the curve (AUCs) of the training and validation cohorts were 0.694 and 0.658, respectively ( Figures 3A, B). A well-fitted calibration curve was also generated, which reflected adequate prediction accuracy using the nomogram model ( Figure 3C). In the subgroup analysis, the nomogram model obtained stable predictive performance in the following subgroups: age (AUC = 0.574 and 0.722 in subgroups >51 years and ≤51 years, respectively), gender (AUC = 0.610 and 0.709 in male and female subgroups, respectively), primary-recurrence subtype (AUC = 0.706 and 0.533 in primary and recurrence subgroups, respectively), maximum dimension (AUC = 0.500, 0.664, and 0.567 in <10, 10-40, and ≥40 mm subgroups, respectively), clinical subtype (AUC = 0.635, 0.667, 0.556, and 0.667 in nonfunctioning PAs, GH secreting PAs, PRL secreting PAs, and ACTH secreting PAs subgroups, respectively), Knosp grade (AUC = 0.692 and 0.679 in grades 0-2 and grades 3-4 subgroups, respectively), Hardy grade (AUC = 0.679 and 0.713 in grades 0-2 and grades 3-4 subgroups, respectively) and prolactin (AUC = 0.667 and 0.590 in >815 mIU/L and ≤815 mIU/L subgroups, respectively) ( Figure S1 and Table 3).

Correlation Analysis
We further investigated the relationship between independent predictors and the Ki-67 index. The correlation analysis found that age (R = −0.23; p < 0.01), maximum dimension (R = 0.17; p < 0.01), and prolactin (R = 0.16; p < 0.01) were all significantly correlated with the Ki-67 index (Figures S2A-C and Table 4). Furthermore, the subgroup analysis revealed that the Ki-67 index was inversely correlated with age in nonfunctioning (R = −0.19; p < 0.01) and GH secreting (R = −0.31; p < 0.01) PA ( Figures S2D, F and Table 4). However, no significant correlation was found in PRL and ACTH secreting PA (Figures S2E, G and Table 4). Interestingly, the subgroup analysis according to maximum dimension found a negative correlation between Ki-67 index and a maximum dimension in microadenoma (R = −0.62; p = 0.02) ( Figure S2H and Table 4), and positive correlation in macroadenoma (R = 0.17; p < 0.01) ( Figure S2I and Table 4). No significant correlation between the Ki-67 index and maximum dimension was found in giant adenoma ( Figure S2J and Table 4).
In the subgroup analysis according to clinical types, we discovered a significant positive correlation between Ki-67 index and maximum dimension in GH (R = 0.30; p = 0.01) ( Figure S2M and Table 4) and PRL (R = 0.32; p = 0.05) ( Figure S2L and Table 4) secreting PA, whereas no significant correlation was found for nonfunctioning and ACTH secreting PA ( Figures S2K, N and Table 4). We conducted a subgroup analysis according to prolactinoma or no prolactinoma, and found that in the no prolactinoma subgroup, the Ki-67 index was significantly correlated with prolactin (R = 0.12; p = 0.02) ( Figure S2O and Table 4). However, significant correlation between the Ki-67 index and prolactin was not detected in the prolactinoma subgroup (R = 0.07; p = 0.65) ( Figure S2P and Table 4).

DISCUSSION
PA is the most common tumor in the sellar region (1), and high Ki-67 suggests invasive growth and worse prognosis for patients     sensitivity and specificity of 90% and 85%, respectively. Although their prediction methods both showed considerable prediction performance, application of these methods was still inconvenient. Our nomogram is the first predictive model for Ki-67 levels based on age, primary-recurrence subtype, maximum dimension, and prolactin. These four parameters were routinely evaluated for patients with PAs in clinical practice, denoting strong practicability for the nomogram. The nomogram showed predictive value (AUC = 0.694 and 0.658 in the training and validation cohorts, respectively) and the performance was stable in various subgroups ( Figure S1 and Table 3). Obviously, this user-friendly nomogram was clinically useful and easy to expand in clinical practice.
Yonezawa et al. found that the Ki-67 index in patients under 30 years of age with nonfunctioning PA was significantly higher than that in patients over 40 years of age (p < 0.01) (19) (24). Conversely, few studies showed no statistical relationship between the Ki-67 index and age (25,26). However, these opposite results were based on mixed clinical subtypes, which could influence the results as a confounding factor. In the current study, we found that age was significantly higher in the low Ki-67 cohort compared with that in the high Ki-67 cohort (p < 0.001) ( Table 1). A multivariate analysis also suggested that age is an independent predictor for Ki-67 index level (OR, 95% CI: 0.9687, 0.9473-0.9906, p = 0.0052) ( Table 2). A correlation analysis revealed a significant inverse correlation between the Ki-67 index and age in all patents with PA (R = −0.23; p < 0.01), nonfunctioning PA (R = −0.19; p < 0.01), and GH secreting PA (R = −0.31; p < 0.01) subgroups ( Figure S2 and Table 4). Although a correlation analysis found no statistical significance in PRL and ACTH secreting PA, these results may be due to our relatively small PRL and ACTH secreting PA sample sizes. The predictive performance of the nomogram model was stable across different clinical subgroups, suggesting that this nomogram could be applied to patients with nonfunctioning, GH, PRL, and ACTH secreting PA.  Most of literature discussed whether the Ki-67 index could or could not predict recurrence, but only few studies analyzed the predictive value of primary-recurrence subtype for predicting Ki-67 index levels in PA. Doyle et al. proposed that Ki-67 was raised in recurrent (33%) than primary (11%) PA (p < 0.01) (27). Yao et al. found that the mean values of Ki-67 index for nonfunctioning PA were 2.76 and 4.09% for the primary and recurrent groups, respectively (28). Grimm et al. discovered a tendency that Ki-67 index were higher in recurrent ACTH and PRL secreting PA but did not reach statistical significance (29). In contrast, Selek et al. found that among the patients who received somatostatin analogue treatment between the first and second operations, the Ki-67 index was significantly lower at the time of the second operation compared to that of the first operation (p < 0.001) (30). However, this decrease in the Ki-67 index may be due to the cellular proliferation depression effect of somatostatin analogues. Based on the uni-and multivariate logistic regression analyses in the current study, the primary-recurrence subtype is an independent predictor for Ki-67 index level ( Table 2). These results suggested that recurrent PA has a higher probability for a high Ki-67 index compared with primary PA.
The relationship between Ki-67 and maximum dimension remains controversial. Some studies have found no significant correlation between Ki-67 and maximum dimension for mixed clinical subtypes (29,31,32), PRL secreting (33), GH secreting (34) and nonfunctioning (35) PAs. In contrast, positive correlations between Ki-67 and maximum dimension in macroadenoma were discovered for mixed clinical subtypes (23,36) and ACTH secreting (37) (43). It seems that maximum dimension subtypes and clinical subtypes should be considered while discussing Ki-67 and maximum dimension for PA. In our study, first, we found that maximum dimension was significantly lower in the low Ki-67 cohort than in the high Ki-67 cohort (low Ki-67 vs. high Ki-67 (mean ± SD): 26.23 ± 11.12 vs. 26.23 ± 11.12; p = 0.01) ( Table 1). The uni-and multivariate analyses revealed that the maximum dimension was an independent risk factor for a high Ki-67 index (OR, 95% CI: 1.0321, 1.0063-1.0585, p = 0.0144) ( Table 2). In the correlation analysis, a positive correlation between the Ki-67 index and maximum dimension was found for all of the included patients with PA (R = 0.17; p < 0.01) ( Figure S2B and Table 4). In the subgroup, according to the maximum dimension subtype and clinical subtype, we discovered a positive correlation and a negative correlation between the Ki-67 index and maximum dimension in the macroadenoma (R = 0.17; p < 0.01) and microadenoma (R = −0.62; p = 0.02) subgroups, respectively ( Figures S2H, I; Table 4), which was consistent with the results from the research by Turner et al. A significant positive correlation between Ki-67 index and maximum dimension was also revealed in GH (R = 0.30; p = 0.01) and PRL (R = 0.32; p = 0.05) secreting PA subgroups ( Figures S2L, M and Table 4). These results require further verification and we propose that additional discussion of Ki-67 and maximum dimension for PA should consider maximum dimension subtypes and clinical subtypes before drawing conclusions.
In this study, we found that the preoperative prolactin concentration was significantly lower in the low Ki-67 cohort compared with that in the high Ki-67 cohort (low Ki-67 vs. high Ki-67 (mean ± SD): 551.82 ± 796.05 vs. 909.36 ± 1190.74; p = 0.001) ( Table 1). The prolactin concentration was an independent predictor for the Ki-67 index level (OR, 95% CI: 1.0004, 1.0000-1.0008, p = 0.0287) ( Table 2) (47). This theory may suggest a potential mechanism for the positive correlation between the Ki-67 index and prolactin, but additional high-quality research is warranted for validation. In the current work, we analyzed the relationship between the Ki-67 index and prolactin in the prolactinoma and no prolactinoma subgroups and found a significant correlation in the no prolactinoma subgroup (Figures S2O, P and Table 4). One possible reason for these interesting results is that in the no prolactinoma subgroup, the positive correlation between the Ki-67 index and prolactin is due to "stalk effect." Higher Ki-67 tends to show up in larger PAs and leads to a more severe "stalk effect." Because dopamine inhibits the basally high-secretory tone of lactotrophs, when the tumor mass blocks the infundibular dopamine release, hyperprolactinemia will occur (48).
Knosp classification, based on the relationship between PA and internal carotid artery, is used to evaluate invasion of the cavernous sinus. Luchi et al. found the Knosp grade was significantly correlated with the Ki-67 index (R = 0.73; p < 0.001) in GH secreting PA population (49). Recently, Li et al. also revealed that the Knosp grade positively correlated with the Ki-67 index (p <0.05) (41). However, another study (50) reported an inverse relationship between the Ki-67 index and Knosp grade for functional PA and nonfunctional PA (R= -0.59, p < 0.001 in functional PA and R = 0.367, p < 0.01 in nonfunctional PA). Das et al. found no significant correlation between Knosp grade and Ki-67 index (5). In the current research, PAs with the Knosp grades 0-2 and grades 3-4 were classified into noninvasive and invasive classes. PAs were also classified into a low Ki-67 index (<3%) and high Ki-67 index (≥3%). No significant correlation was found between the Knosp grade and Ki-67 index ( Table 4 and Supplementary Table 4). The difference between our results and previous studies' results may be due to our inclusion criteria, based on which we included both functional and nonfunctional PA patients.
This study had several limitations. First, as a single-institution retrospective study, potential selection bias derived from some unknown factors is inevitable and may decrease the reliability of the results. And because no thyrotropinoma, and gonadotropinoma was treated in our center, the current study only included nonfunctioning, GH secreting, PRL secreting, and ACTH secreting PAs. The lack of thyrotropinoma, and gonadotropinoma may lead to selection bias, which cannot be ignored. Therefore, additional investigations from external sources are warranted to comprehensively evaluate and validate the nomogram. Second, the predictive performance of the nomogram and that in some subgroups were unsatisfactory, e.g., microadenoma. However, this is the first nomogram to preoperatively predict the Ki-67 level for PAs, and the model did show predictive value to some extent. So, we recommended clinicians to apply this nomogram in clinical. But clinicians had to carefully select the appropriate patients while applying this nomogram. For example, patients with following characteristics are more suitable for this nomogram: under 51 years old (AUC = 0.722), female (AUC = 0.709), primary subtype (AUC = 0.706) or invasive with Hardy grade for sellar invasion (AUC = 0.713). In the future, additional nomogram studies are warranted to focus on the appropriate and eligible subgroups of patients. Third, several variables collected in this study had missing data. However, we checked all of the variables filtered by the LASSO and univariate regression analyses, and all of these analyses had missing data under 60%. Furthermore, in this study, the percentage of missing data in the final model was only 23.3%, which is a satisfactory and adequate sample size for a multivariate analysis. Finally, we found a different relationship between the Ki-67 index and its independent risk factors in the subgroup analysis; thus, a better nomogram should be based on some subgroup populations. Because the current work focused on all patients with PAs, we did not calculate nomograms for specific subgroup populations. Additional studies are warranted to analyze the prediction of the Ki-67 index in these subgroup populations.

CONCLUSIONS
This study showed that age, primary-recurrence subtype, maximum dimension, and prolactin are independent risk factors associated with the Ki-67 level. Among these risk factors, prolactin was found for the first time to be an independent predictor for the Ki-67 level. The novel nomogram developed in this study was feasible and stable for preoperative prediction of the Ki-67 index level, suggesting that the nomogram is of potential value for further assisting neurosurgeons to develop better, more individualized treatment strategies for patients with PAs.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the Ethics Committee of Jinling Hospital. Written informed consent for participation was not provided by the participants' legal guardians/next of kin because of the nature of the retrospective cohort study and the anonymization of data prior to analysis, informed consent was waived.

AUTHOR CONTRIBUTIONS
CM conceived and designed the investigation. XC analyzed the data and drafted the manuscript. JZ, JY, CT, FY, and ZC conducted statistical analyses. All authors contributed to the article and approved the submitted version.

ACKNOWLEDGMENTS
We thank Jinfang Sun for her assistance in statistical analyses.