A Nomogram Derived by Combination of Demographic and Biomarker Data Improves the Noninvasive Evaluation of Patients at Risk for Bladder Cancer

With an estimated 70,980 newly diagnosed cases of bladder cancer and 14,330 deaths from bladder cancer in 2015, cancer of the urinary bladder is the second most common genitourinary malignancy in the United States and among the five most common malignancies worldwide (1, 2). When detected early (i.e., non-muscle invasive), the 5-year survival rate of bladder cancer is >90%; however, at later stages (i.e., muscle invasive and beyond), the 5-year survival rate is <50%. Thus, early bladder cancer identification, both at the initial diagnosis and at recurrence, can dramatically affect outcomes (3). Urine-based assays that can noninvasively detect bladder cancer have the potential to improve the rapid diagnosis of bladder cancer. As such, several urine-based commercial molecular tests have been FDA approved for bladder cancer detection and surveillance. These tests include the measurement of soluble proteins, such as bladder tumor antigen, and nuclear matrix protein 22 (NMP22), or proteins detected on fixed urothelial cells (ImmunoCyt), and chromosomal aberrations detected by FISH (UroVysion; ref. 4). Because of their marginal detection performance, these urine-based assays have a limited role in the management of patients at risk for, or with, bladder cancer, and thus, the search for noninvasive urine-based tests with clinical utility for bladder cancer continues.

We and others have described the diagnostic capabilities of urine-based molecular signatures to noninvasively detect bladder cancer (5–12). We have refined and validated a multiplex protein biomarker panel (MMP9, MMP10, IL8, VEGFA, SERPINE1, SERPINA1, CA9, APOE, ANG, and SCD1) in a series of independent cohorts (13–15). Given the utility of key demographic features (e.g., age, race, sex, and tobacco history) in stratifying patients, we investigated the potential utility of a hybrid nomogram that incorporates key demographic features with the results of the bladder cancer–associated diagnostic signature in hopes of improving the evaluation of risk for harboring bladder cancer. If proven accurate and reliable, the application of such a nomogram may guide the decision to perform invasive diagnostic procedures.

Demographic, clinical, and biomarker data from 686 subjects [394 incident bladder cancer subjects and 292 subjects with benign urologic conditions (e.g., erectile dysfunction, voiding symptoms, microscopic hematuria, kidney stones)] from 10 clinical sites were extracted from our series of independent cohorts previously published (Supplementary Table S1; refs. 13–15). Subjects with urinary tract infection and gross hematuria were excluded. All molecular data were normalized to creatinine. On the basis of the total distribution of each biomarker's concentration, cut-off points were identified deriving low/high expression status (13–15).

The primary endpoint of the study was to predict the histologic presence of urothelial carcinoma of the bladder, which was confirmed by biopsy. Tumor grade (2002 WHO classification; ref. 16) and tumor stage (2002 TNM classification; ref. 17) were noted for each case. No central pathology review was obtained.

The distributions of the key demographic data as well as molecular data were examined. Multivariate logistic regression analysis was used to examine the association between these predictor variables and detection of bladder cancer. All decisions with respect to the coding of the nomogram variables were made prior to modeling, as making these decisions afterwards can have deleterious effects on the predictive ability of the model (18). A logistic regression model based on disease status was the basis for our nomograms, which included only key demographic data, only key biomarker data, and the combination of key demographic data and biomarker data. Stepdown method was employed to obtain the reduced model with highest concordance index (C-index).

Nomogram validation contained two components. The nomogram was subjected to bootstrapping as a means of calculating a relatively unbiased measure of its ability to discriminate among subjects. Briefly, we compared the predicted probability of diagnosis with actual diagnosis (i.e., nomogram calibration) on the 686 subjects, using 1,000 bootstraps to reduce overfit bias, which would otherwise overstate the accuracy of the nomogram. We quantified the discrimination ability of the risk calculator by calculating the C-index, which is a surrogate of the nonparametric area under the receiver operating characteristic curve (AUC; ref. 19). C-index gives the probability that in a randomly selected pair of subjects in which one has bladder cancer and the other does not, the subject with bladder cancer will be assigned the worse predicted risk (20). C-index ranges from 0.5 (no discrimination) to 1.0 (perfect discrimination). To test the significance between the AUCs of the three nomograms (demographic data only, biomarker data only, and combination of key demographic data and biomarker data), we created 1,000 C-indices for each model by using bootstrapping analysis and then calculated the differences between the paired C-indices. Finally, nonparametric bootstrap test (21) was used to calculate the P value for each pair of the nomograms.

The calibration of the three nomograms was compared by plotting the prediction on the x-axis and the observed outcomes on the y-axis in the same plot (22). In the calibration plot, the 45° line represents the perfect predictions. Because of binary outcomes, a smoothing technique was used to generate the observed probabilities of bladder cancer on the x-axis. We also applied the decision curve analysis (23) on our proposed nomograms and compared the net benefits of different examined actions. All statistical analyses were performed using S-Plus software (PC Version 3.3) and R software version 3.2.3 with additional functions. All P values were calculated by two-sided statistical tests, unless notified otherwise.

Of the 686 subjects available for analysis, 394 had bladder cancer, whereas 292 were healthy volunteers/benign controls. More than 84% of the bladder cancer subjects were >55 years (60% of controls), 92% of the bladder cancer subjects were Caucasian (66% of controls), and 83% of the bladder cancer subjects were male (79% of controls). Nineteen percent of bladder cancer subjects denied tobacco history, while 37% of controls denied tobacco use (Table 1). Of the subjects with bladder cancer, 240 of the tumors were noninvasive (Ta, Tis, T1) and 147 were muscle invasive and 7 did not have a stage reported. In addition, 134 were low grade, 251 were high grade, and 9 did not have a grade reported.

Logistic regression analysis identified key demographic risk factors (e.g., age, race, and tobacco use) and molecular biomarkers (MMP9, MMP10, IL8, VEGFA, SERPINE1, SERPINA1, CA9, APOE, ANG, and SCD1) associated with bladder cancer. The key demographic factors were used to generate a demographic only model with AUC of 0.81 [95% confidence interval (CI), 0.76–0.85] and with C-index of 0.806. The key biomarker data were used to generate a biomarker only model with AUC of 0.84 (95% CI, 0.80–0.87) and with C-index of 0.835. Under the likelihood ratio test, the biomarker model performed better than the demographic model (P = 6.745e−4). Subsequently, these two nomograms were combined to create a hybrid nomogram that incorporated key demographic and biomarker data (Fig. 1). The AUC of the hybrid nomogram was 0.89 (95% CI, 0.86–0.92), which, based on the nonparametric bootstrap test, was significantly improved from the demographic model (0.81; 95% CI, 0.76–0.85; P < 0.0001) and the biomarker model (0.84; 95% CI, 0.80–0.87; P < 0.0001; Fig. 2). The hybrid model possessed a C-index of 0.891. Using the hybrid nomogram, we were able to calculate the sensitivity and specificity for a range of probability for bladder cancer (Table 2).

Figure 3 illustrates how the predictions from the hybrid nomogram compare with actual outcomes for the 686 subjects. The x-axis is the prediction calculated with use of the hybrid nomogram, and the y-axis is the actual freedom from cancer for our subjects. The dashed line represents the performance of an ideal nomogram, in which predicted outcome perfectly corresponds with actual outcome. Our hybrid nomogram performance after adjusting the overfitting bias with bootstrap is plotted as the solid line. Note that, because the solid line is relatively close to the dashed reference line, the predictions calculated with the use of our hybrid nomogram approximate the actual outcomes. In general, the performance of the hybrid nomogram appears to be within 10% of actual outcome, and possibly slightly more accurate at very high levels of predicted probability.

We also applied decision curve analysis to measure the performance of our hybrid nomogram for bladder cancer (Fig. 4). We tested the theoretical net benefits of all actions in a range of threshold probabilities for bladder cancer. Basically, the net benefit is measuring how our action can affect the examined relative value of false positives and false negatives, which is when our hybrid nomogram is compared with cystoscopy and biopsy. The decision curve analyses showed that the hybrid nomogram performed better than demographic data alone above the risk threshold of 6% as well as biomarker data alone above 24% to 88%.

Predictive and prognostic nomograms in bladder cancer have been published in both non-muscle–invasive (24) and muscle-invasive bladder cancer (25, 26). Specifically, non-muscle–invasive nomograms of precystoscopy urinary levels of NMP22 improved the ability of age, gender, and voided urinary cytology (VUC) to predict tumor stage and grade as well as tumor recurrence (24), while in muscle-invasive nomograms, precystectomy clinical and pathologic factors pT and pN stages at the time of cystectomy (25) and to estimate the probabilities of recurrence and all-cause and bladder cancer–specific survival (25, 26) after cystectomy.

To the best of our knowledge, this is the first study to evaluate and internally validate a bladder cancer diagnostic nomogram composed of pertinent demographic features and our bladder cancer–associated diagnostic signature. Previously, we have reported and confirmed in voided urines our bladder cancer–associated diagnostic signature composed of 10 biomarkers in three separate studies (13–15). The first study was a case–control study of 127 patients (64 tumor-bearing subjects), in which we reported a sensitivity of sensitivity 92% and specificity 97%, significantly outperforming voided urinary cytology (13). Subsequently, in another case–control study, we tested the bladder cancer diagnostic signature in 308 patients (102 tumor-bearing subjects and 206 subjects with varying urological disorders, for example, urolithiasis, gross hematuria, urinary tract infection, moderate-to-severe voiding symptoms), recording a sensitivity of 74% and specificity of 90%, which outperformed VUC and the UroVysion cytogenetic test (14). Recently, we published a multicenter, international case–control study of 320 patients (183 tumor-bearing subjects) and demonstrated continued diagnostic performance with a sensitivity of 79% and a specificity of 79% (15).

The gold standard for initial clinical diagnosis and staging of bladder cancer involves cystoscopic examination of the bladder together with cytologic examination for malignant cells in the urine. Cystoscopy is an unpleasant invasive procedure, which may involve anesthetizing the patient and resection of biopsies for histopathologic diagnosis and staging. Cystoscopy may also have certain side effects, such as urinary tract infection, voiding symptoms, and stenosis of the urethra. VUC remains the method of choice for the noninvasive detection of bladder cancer, with its main use being to recognize the presence of recurrence and early progression in stage and grade. VUC can be used to diagnose new malignancy, yet although it has a specificity of >93%, its sensitivity is only 25% to 40%, especially for low-grade and low-stage tumors (4, 27, 28). Thus, current methods to noninvasively detect bladder cancer leave much to be desired. The inadequate power of these single markers must partly explain this. The concept that the presence or absence of one molecular marker will aid diagnostic or prognostic evaluation has not proved to be the case. A number of molecular signatures have been derived and are being made commercially available as clinical assays, especially in the breast cancer field (29, 30). We have employed a range of genomic (12, 31) and proteomic (11, 32) profiling approaches to study voided urine samples in hopes of identifying a unique, yet accurate, molecular signature associated with bladder cancer.

In biomarker research, although a variable maybe statistically significant in a multivariate model, it does not necessarily equate to the biomarker improving the model's predictive accuracy. For example, a biomarker with an OR of 3 may be a poor classifier and thus an OR of 10 or more may be required. In addition, a single measure of association, such as an OR, may not meaningfully describe a biomarker's ability to risk classify patients (33). Thus, it is critical to determine whether the addition of biomarker(s) to an existing clinical and pathologic model possesses the ability to improve the predictive accuracy of this model. The accuracy of the hybrid nomogram improved to 0.89 (95% CI, 0.86–0.92) compared with key demographic model (0.81; 95% CI, 0.76–0.85; P = 5.886e−8) and biomarker model (0.84; 95% CI, 0.80–0.87; P < 7.707e−5; Fig. 2). In general, the performance of the hybrid nomogram appears to be within 10% of actual outcome, and possibly slightly more accurate at very high levels of predicted probability. We also applied decision curve analysis to measure the performance of our hybrid nomogram for bladder cancer. The decision curve analyses showed that the hybrid nomogram performed better than demographic data alone above the risk threshold of 6% as well as biomarker data alone above 24% to 88%.

The main clinical utility of a hybrid nomogram in the described setting is to facilitate the decision on whether a patient requires cystoscopy with subsequent bladder biopsy. The hybrid nomogram would provide a probability of harboring bladder cancer. For example, if the probability of harboring bladder cancer is <10%, perhaps the patient and physician would forego an invasive procedure. However, if the risk was substantial (i.e., >70%), then mostly likely, the patient would be compelled to undergo confirmative diagnostic procedure. In the absence of definitive risk thresholds, it would be important to provide a range of threshold probabilities (Table 2).

We acknowledge that this study is limited due to its retrospective design, to the analysis of banked urine samples collected from high volume centers and so may not be representative of the general population at risk for bladder cancer and to limitations of available data (e.g., detailed tobacco history). Nevertheless, this cohort reflects a contemporary cohort of bladder cancer patients, which enabled the derivation of a hybrid nomogram for testing in larger, more diverse prospective studies.

In this study, we developed a hybrid nomogram that facilitates the accurate prediction of the probability of a patient harboring bladder cancer. The hybrid nomogram has been constructed by combining readily available key demographic factors with key biomarker data. If such a nomogram is proven to be reliable, adoption may assist the physician and patient in deciding whether or not further evaluation is needed.

S. Goodison is the Chief Executive Officer at Nonagen Bioscience Corp. and has ownership interest (including patents) in a patent issued. C.J. Rosser is the President at Nonagen Bioscience Corp., and has ownership interest (including patents) in Nonagen Bioscience Corp. and US Patent #9249467. No potential conflicts of interest were disclosed by the other authors.

Conception and design: S. Goodison, C.J. Rosser

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C.J. Rosser

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S. Huang, L. Kou, C. Yu, M.W. Kattan, L. Garmire

Writing, review, and/or revision of the manuscript: S. Huang, L. Kou, H. Furuya, C. Yu, S. Goodison, M.W. Kattan, L. Garmire, C.J. Rosser

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): H. Furuya, C.J. Rosser

Study supervision: C.J. Rosser

This work was supported by research grants from Weinman Foundation Fund (to C.J. Rosser), R01 CA198887 (to C.J. Rosser), 5P30CA0717890-6071 (to C.J. Rosser), R44 CA173921 (to S. Goodison), RO1 CA116161 (to S. Goodison), K01 ES025434 (to L.X. Garmire), and P20 GM103457 (to L.X. Garmire).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.