Oncology Development of a Clinical Prediction Model for Assessment of Malignancy Risk in Bosniak III Renal Lesions Ajit H. Goenka, Erick M. Remer, Andrew D. Smith, Nancy A. Obuchowski, Joseph Klink, and Steven C. Campbell OBJECTIVE METHODS
RESULTS
CONCLUSION
To identify independent predictors of malignancy in Bosniak III (BIII) renal lesions and to build a prediction model based on readily identifiable clinical variables. In this institutional review board-approved, Health Insurance Portability and Accountability Act (HIPAA)-compliant retrospective study, radiology, and hospital information systems containing data from January 1, 1994, to August 31, 2009, were queried for adult patients (age >18 years) with surgically excised BIII lesions. Clinical variables and results of histopathology were noted. Univariate and multiple-variable logistic regression analyses were performed to identify potential predictors and to build a prediction model. Cross-validation was used to assess generalizability of the model’s performance, as characterized by concordance (c) index. Of the 107 lesions in 101 patients, 59 were malignant and 48 benign. On univariate analyses, the strongest potential predictors of malignancy were African American race (P ¼ .043), history of renal cell carcinoma (RCC; P ¼ .026), coexisting BIII lesions (P ¼ .032), coexisting Bosniak IV (BIV) lesions (P ¼ .104), body mass index (BMI; P ¼ .078), and lesion size (P 80%, the positive predictive value of the model is 92% (CI 78%-100%). Clinical risk factors offer modest but definite predictive ability for malignancy in BIII lesions. In particular, a prediction model encompassing lesion size, BMI, and history of RCC seems promising. Further refinements with possible inclusion of imaging biomarkers and validation on an independent dataset are desirable. UROLOGY 82: 630e635, 2013. 2013 Elsevier Inc.
C
omplex renal cysts are common, they can be benign or malignant, and are being detected with increasing frequency because of an escalating use of cross-sectional imaging.1 A system of classifying the risk of malignancy based on computed tomography (CT) scan findings was developed by Bosniak.2 A particularly problematic category is the Bosniak III (BIII) complex cyst, which is characterized by thickened, irregular or smooth walls, or septa with measurable enhancement. These lesions are most commonly managed with surgical excision
Preliminary findings presented at the 2012 meetings of the American Roentgen Ray Society (ARRS), Society of Abdominal Radiology (SAR), and the Genitourinary Meeting of the American Society of Clinical Oncology (ASCO). Financial Disclosure: The authors declare that they have no relevant financial interests. From the Section of Abdominal Imaging, Imaging Institute, Cleveland Clinic e Hb6, Cleveland, OH; the Department of Radiology, University of Mississippi Medical Center, Jackson, MS; Quantitative Health Sciences, Cleveland Clinic e JJN3, Cleveland, OH; and the Section of Urologic Oncology, Glickman Urological and Kidney Institute, Cleveland Clinic - Q10, Cleveland, OH Reprint requests: Erick M. Remer, M.D., Section of Abdominal Imaging, Imaging Institute, Cleveland Clinic e A21, 9500 Euclid Avenue, Cleveland, OH 44195. E-mail:
[email protected] Submitted: February 13, 2013, accepted (with revisions): May 14, 2013
630
ª 2013 Elsevier Inc. All Rights Reserved
because the malignancy risk is approximately 50%.3 Because of our limited ability to predict the malignant potential of a BIII lesion in a given patient, some patients undergo unnecessary surgery. Moreover, the survival benefit of resection of a BIII lesion in operable candidates has not been demonstrated and the long-term incidence of metastases in cancers that arise in BIII lesions is not known.3 Thus, the identification of a BIII lesion on crosssectional imaging often generates clinical uncertainty. Previous work has suggested that patient-related factors that predispose to renal cell carcinoma (RCC) may also be at work in malignant complex renal cysts.3 Preoperative risk assessment and predicting the likelihood of malignancy in BIII lesions would be useful for patient counseling, risk stratification, and clinical assessment of the need for surgical treatment or imaging surveillance. We hypothesize that easily obtainable clinical factors and patient history are associated with an increased risk of malignancy in BIII lesions. The objective of this study was to identify preoperative clinical factors that are independently associated with malignancy on surgical pathology in patients with BIII cystic renal lesions. 0090-4295/13/$36.00 http://dx.doi.org/10.1016/j.urology.2013.05.016
MATERIAL AND METHODS Study Design and Study Group Informed consent was waived in this institutional review boardapproved, Health Insurance Portability and Accountability Act (HIPAA)-compliant retrospective study. The Hospital Information System (Epic Systems Corporation, Verona, WI) and Radiology Information System (Syngo Workflow, Siemens Medical USA, Inc., Malvern, PA) were searched for adult patients (age >18 years) with BIII lesions from January 1, 1994, to August 31, 2009. Included patients had 1 or more BIII lesions that were characterized by a renal CT without and with intravenous contrast and had available surgical histopathology data. All the CT scans on which the BIII lesions were identified were interpreted by fellowship-trained abdominal radiologists practicing at an academic medical center. Patients with Von HippelLindau disease were to be excluded from the study. However, there were no surgically excised BIII lesions from patients with Von Hippel-Lindau in the retrieved population. Thus, a total of 107 surgically excised BIII lesions in 101 patients (6 patients had 2 BIII lesions each) were included in the final analysis. This patient cohort has also been included in a previous study evaluating malignancy rate and outcomes.3
homogeneous groups based on a deviance score. A backward selection process was used, eliminating the variables with the largest P values first. The discriminant ability of the models was assessed by the concordance (c)-index. The c-index is a measure of discrimination (ie, model’s discriminant ability between malignant and benign lesions) and ranges in value from 0.0 (ie, the model incorrectly classifies malignant and benign) to 1.0 (ie, the model perfectly discriminates). The null value is 0.5 indicating that the model is no better than guesswork. To assess the fit of the final model, the Hosmer-Lemeshow test was performed. To assess how the models’ performance would generalize to an independent dataset, a leave-one-out cross-validation was performed in phase III. This method is based on using a single observation from the original sample as the validation data, and the remaining observations as the training data. This is repeated such that each observation in the sample is used once as the validation data. The c-index of the model with this cross-validation is reported. The calibration of the model was assessed by comparing observed and fitted proportions of patients with malignant lesions.
RESULTS Data Collection The following data were recorded from the electronic medical records and the imaging reports: patient gender; age at BIII identification; race; presenting symptoms (if any); body mass index (BMI); history of smoking; history of RCC; largest single diameter of the BIII lesion from the CT reports; presence and number of coexisting Bosniak IV (BIV) lesions or solid renal masses on the imaging study; and the number and side (right and/or left kidney) of BIII lesions. Moreover, the time interval from the diagnosis of a BIII lesion to that of surgical excision was noted. Histopathological specimen reports were reviewed, findings were noted, and results were classified as benign or malignant. The malignant lesions were further subclassified into aggressive and nonaggressive subtypes in an effort to correlate the variables to the grade of malignancy. Aggressive subtypes included Fuhrman grade 3 and 4 of clear cell, papillary, mixed, and unclassified RCC, whereas Fuhrman grade 1 and 2 of the clear cell, papillary, and cystic RCC were considered less aggressive variants.
Statistical Analyses Statistical analyses consisted of the following phases: assessment of potential predictors of malignancy (phase I); identification and assessment of potential prediction models using these variables (phase II); and cross-validation of the prediction models (phase III). In phase I, univariate analyses were performed to identify the strongest predictors of malignancy. For categorical variables, logistic regression analysis with generalized estimating equations (GEEs) and an exchangeable working correlation matrix was used to test each variable. For continuous variables, the variable was treated as both a continuous and categorical variable and tested in logistic regression analysis with GEEs. In phase II, potential predictors with univariate P values 30). Median duration between the imaging identification of a BIII lesion and its surgical excision was 0.1 years (range 0-2.1 years). Table 1 summarizes the results of univariate analyses. As depicted in the table, the strongest potential predictors of malignancy were race, history of RCC, coexisting BIII lesions, coexisting BIV lesions, BMI, and lesion size. The classification tree indicated that lesion size was the variable that had the greatest impact on reducing the variance. In the multiple-variable analysis, lesion size was the strongest independent predictor of malignancy followed by BMI and a history of RCC. The model suggests that the risk of a BIII lesion being malignant is higher for patients with smaller lesions, patients with larger BMIs, and patients with a history of RCC. No other main effects of the variables or the interactions between the variables had P values
