This study defines a clinical tool that offers clinicians and physicians the ability to effectively triage-out patients presenting with hematuria from the need to have a full urological work-up for the detection of UC. The study presents an internally validated genotypic-phenotypic model, G + P INDEX, with bootstrap-based CI estimates, that offers a combination of high sensitivity and high NPV (i.e. a low probability of an individual patient with UC providing a false-negative result and a high proportion of all negative results being true) that is not offered by models derived exclusively from genotypic or phenotypic data alone. This provides clinicians and physicians with a unique opportunity to triage out patients with both micro- and macrohematuria, in particular by identifying patients with a low risk of having UC who do not require a full urological work-up.

A high test-negative rate in the context of high sensitivity is an important consideration for an effective triage-out test that aims to direct patients with a low probability of UC away from a full clinical work-up [37]. Accordingly, at a test-negative rate of 0.4 the sensitivity of the G + P INDEX presented here maximizes both the sensitivity and NPV (0.95 and 0.98, respectively). This can be compared with the best fit selected from the genotypic model published in [33] (sensitivity = 0.82; NPV = 0.97) and is also comparable with the sensitivity and NPV of both cystoscopy (sensitivity = 0.89–0.98; NPV = 0.99) and virtual cystoscopy using computed tomography (CT) scans or magnetic resonance imaging (MRI) (sensitivity = 0.94 and 0.91, respectively) [38-40].

It is acknowledged that the population used to derive the G INDEX, P INDEX and G + P INDEX in this instance consisted of patients with macrohematuria. Therefore, the derived NPV values are only applicable to the macrohematuria population. The sensitivity of 0.95 (95% CI: 0.86–0.98) may, however, be applied across populations. Presuming that patients with and without UC are similarly distributed amongst the micro- and macrohematuria patient populations as depicted in [5,23-25], and that expected UC prevalence is 4% in the microhematuria population, a higher NPV and test-negative rate can be expected in the target microhematuria population.

By applying the G + P INDEX to the sample population of patients with microhematuria who do not have UC it was shown that 80% of the patients would have been triaged out on the basis of the result. Only 20% would be referred for a full urological work-up. This compares with conventional guidelines that would currently see all of the patients (100%) with microhematuria that cannot be attributed to a benign cause undergoing a full urological work-up, incurring significant unnecessary costs and negatively impacting patient QoL.

Severity of hematuria is correlated with the probability of a patient having UC, but not the stage or grade of any tumor, and an estimated 96% and 77–88% of patients with micro- and macrohematuria, respectively, referred to a urologist will not have UC [2-6]. Consequently, avoiding potentially unnecessary urological work-ups for patients with hematuria has several benefits. Cystoscopy may be associated with adverse effects, such as pain on voiding, bleeding, UTIs, male sexual dysfunction and the anxiety that accompanies an inconclusive or unconfirmed UC diagnosis [9-11]. Most notably, this novel approach has the potential to reduce the burden on resources and the financial cost associated with a full urological work-up on UC-negative patients. For example, in the UK, avoiding cystoscopy in patients with hematuria with an initial negative cytology and/or tumor biomarker test has been estimated to save approximately US$770 per patient (£483 per patient) evaluated [13]. The G + P INDEX described here provides an effective alternative to the use of urine cytology when used in a primary evaluation setting. This is particularly relevant in settings where primary evaluation is carried out by primary care physicians.

On this basis, if we assign an arbitrary ‘nominal cost’ of US$4,500 for each full urological work up, the total cost for working up 1,000 patients with microhematuria would approach US$4.5 million. In contrast, if 80% of patients with microhematuria are triaged out using the G + P INDEX at an arbitrary nominal cost of US$2,500, the total direct cost of testing and full urological work-ups for the remaining 20% of patients would total US$3.4 million. This provides a notional net saving in direct costs of approximately US$1.1 million per 1,000 patients with microhematuria.

The G + P INDEX and G INDEX, developed in this study, use the same genotypic biomarkers used in the genotypic model described in O’Sullivan et al. [33]. However, the G + P INDEX adds a further four phenotypic variables to enhance the ability to segregate patients who have a low probability of UC. The G + P INDEX uses a combinatorial method with a high sensitivity and a high NPV. By contrast, the genotypic model described in O’Sullivan et al. [33] optimizes the balance between sensitivity and specificity to calibrate the model calibrated for the optimal primary detection of UC in symptomatic patients (i.e. presenting with hematuria) who were undergoing a full urological work-up for suspected UC. The G + P INDEX has a significantly different clinical endpoint as no attempt is made to define or select patients with UC. Instead the aim is to confidently rule out those who do not have UC, and as such, all patients not segregated out would progress for a full urological work-up.

While several studies have previously sought to develop predictive models that consider phenotype when assessing the risk of UC in patients presenting with hematuria, the accuracy of phenotype-dependent models alone appears to be limited. For example, Loo et al. [21] prospectively investigated whether phenotypic parameters could be used to identify patients with microhematuria who may not have required a urological referral and full work-up and concluded that age, male gender and a recent diagnosis of macrohematuria were significant predictors of UC. A history of smoking and >25 RBCs/HPF in a recent urinalysis were not statistically significant predictors of UC, in isolation, but even when included in their ‘Hematuria Risk Index’ to improve predictive accuracy, this index resulted in an AUC of 0.809 [21]. Interestingly, the phenotypic ORs in this study and those identified by Loo et al. are comparable, with overlapping 95% CIs for smoking history and gender, and while age, gender and smoking history have similar weightings in each model, the influence of the genotypic component of the G + P INDEX presented here is likely to account for the higher AUC [21].

Likewise, Cha et al. [20] reported that age, smoking history and degree of hematuria, but not gender, were significantly correlated with the presence of UC in patients with asymptomatic hematuria and used a multivariate model to develop a nomogram comprised of phenotypic and urine cytology data for predicting UC. As with Loo et al. [21], the reported phenotypic ORs are comparable to those reported here, but even after incorporating urine cytology into the nomogram, the AUC of 0.831 reported in [20] was lower than that of the G + P INDEX.

In another study, Tan et al. [22] retrospectively stratified patients with hematuria who had been referred to a specialist urology clinic into high- and low-risk groups using a nomogram derived from patient age, gender, smoking history and the degree of hematuria. While comparisons with this study must be made with caution given the high proportion of patients who were excluded due to an absence of data (80 out of 405 patients), the AUC of 0.804, sensitivity of 0.900 and NPV of 0.953 were all lower than the G + P INDEX described here.

Several attempts have also been made to improve the accuracy of phenotypic models by supplementing them with the results of urinary biomarker tests. When the nuclear matrix protein NMP22 point of care proteomic assay is used in isolation to detect UC it has a sensitivity of 0.557 and NPV of 0.968 [17]. Lotan et al. [41] published a multivariable algorithm comprising phenotypic factors, NMP22 and urine cytology with an AUC for predicting UC of 0.826 that was then prospectively validated with an AUC of 0.802 [31]. However, it is important to note that this model attempted to discriminate between high-risk patients who did and did not have UC, as opposed to maximizing sensitivity and NPV to triage-out patients with a low probability of UC.

The improved accuracy obtained with algorithms comprising both genotypic and phenotypic data have previously been demonstrated in breast cancer, in particular [42-45]. Likewise, Mitra et al. [30] used a combination of molecular markers and smoking intensity to calculate a multivariate model that was superior to routine clinicopathological parameters in predicting survival in patients with UC. However, the present study is the first to demonstrate that phenotypic risk factors can be combined with genotypic data to increase the accuracy of a model for separating patients with hematuria into categories requiring differential levels of urological follow up and clinical care rather than survivorship prediction.

When phenotypic data are combined with genotypic data in a model, the resolution of data is likely to impact the accuracy of the model. For example, smoking is a well understood risk factor for UC and is included in most phenotypic models for detecting UC. In Cha et al. [20], Tan et al. [22], Lotan et al. [31,41] and the current study, the binary discriminants never smoked and current/ex-smoker were used, whereas Mitra et al. [30] calculated smoking intensity on the basis of years of smoking and number of cigarettes smoked each day and Loo et al. [21] categorized smokers into never smoked, passive smokers, smokers who had ceased and current smokers. While it is known that the risk of UC increases substantially with exposure to smoking [46], arbitrarily defining phenotypic variables may limit the overall accuracy and utility of phenotypic models. In contrast, an interaction between a patient’s genotypic and phenotypic variables would not be unexpected. However, combining the impact of phenotypic factors and genotypic variables in a single tool improved the accuracy of the model described in this study. A similar principle also applies to describing hematuria phenotype. Patients presenting with micro- or macrohematuria are essentially on a biological continuum and have different likelihoods of having UC [2-6,21]. Accordingly, despite all patients with microhematuria in this study having a Hfreq score of 0, the severity of their hematuria, in combination with other phenotypic factors, is likely to be indirectly accounted for in the genotypic component of the G + P INDEX.