Developing a prediction model in a large case-control study for the early detection of bladder cancer

Abstract

Background

Bladder cancer (BCa) is a lethal cancer, but early-detection offers an opportunity to improve prognosis. Our objective was to develop a urine-based multi-marker panel for BCa detection across multiple longitudinal cohort studies in a nested case-control study.

Methods

Longitudinal cohorts included healthy participants enrolled in the Southern Community Cohort Study (SCCS), Singapore Chinese Health Study (SCHS), Shanghai Women/Men Health Study (SWMHS), and Multiethnic Cohort (MEC). We measured the levels of 10 protein biomarkers (A1AT, ANG, APOE, CA9, IL8, MMP9, MMP10, PAI1, SDC1, and VEGF) in spot-voided urine samples using the multiplex immunoassay Oncuria. Single urine specimens collected from 274 participants who would go on to develop BCa in the ensuing 3‒60 months (i.e., cases) were age/sex-matched to 274 cancer-free controls. We used generalized estimating equation models, logistic regression analysis, and random forest algorithms to analyze the data.

Results

Differences in the individual biomarker levels between cases and controls were noted for ANG at 12 months (p = 0.046), APOE at 12 months (p < 0.001), MMP10 at 12 months (p = 0.009), PAI1 at 12 months (p = 0.005), SDC1 at 12 months (p = 0.003), 48 months (p = 0.029) and 60 months (p = 0.002), and VEGF at 12 months (p < 0.001). Lastly, the best preliminary model to predict subsequent BCa was IL8, CA9, PAI1, APOE and clinical features which had an AUC of 0.98, accuracy of 0.94 with a sensitivity of 0.88 and specificity of 1.00.

Conclusions

Additional testing is needed; however preliminary results demonstrate that a multiplex immunoassay may be able to facilitate the early detection of BCa in at-risk patients. Identification of BCa at an early stage may lead to improved patient outcomes.

Prevention relevance

Using large multinational patient populations, we tested the performance of the Oncuria multiplex assay to accurately predict the risk of developing bladder cancer by simultaneously analyzing the concentrations of 10 protein biomarkers in urine samples.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-025-07511-1.

Introduction

An estimated 84,870 newly diagnosed cases of bladder cancer (BCa) and 17,420 deaths from BCa will occur in 2025 in the United States (U.S.) alone [1]. Both the absolute numbers of cases and deaths from BCa have increased by 57% and 41%, respectively, since 2000 [1, 2]. When detected early (i.e., non-muscle invasive bladder cancer (NMIBC) or ≤ stage 1), the 5-year survival rate is approximately 94%, compared to at best a 50% 5-year survival rate when the disease is diagnosed as muscle invasive bladder cancer (MIBC) (stage 2) and < 20% 5-year survival when the disease is metastatic (stages 3 and 4) [3–6]. Thus, the prevailing idea is that early detection of BCa (i.e., detection before symptoms, hematuria) in at-risk individuals is the best modality to address and combat bladder cancer’s alarming and increasing death rates.

A literature review regarding early detection markers for BCa identified scant reporting. First, Messing et al.. utilized a chemical reagent strip for hemoglobin detection (urinary dipstick testing) to screen a high-risk population harboring BCa (i.e., men aged 50 years and older with a significant tobacco history; ≥40 pack years) [7]. The urinary dipstick test for hematuria was noted to possess a sensitivity of 50% at a specificity of 54% for BCa detection and led to a lower proportion of high-grade invasive bladder cancers (late stage) in screened (10%) vs. unscreened men (60%; p = 0.002). At 14 years of follow-up, no men with screen-detected BCa had died of disease, whereas 20.4% of men with unscreened BCa had died of disease (p = 0.02) [7]. In addition, the Bladder Cancer Urine Marker Project (BLU-P) study assesses the feasibility of a population-based screening for BCa and at the same time, evaluates a screening algorithm, which included NMP22, FGFR3, microsatellite analysis (MA) and MLPa (a custom methylation-specific test) in an attempt to circumvent the high number of cystoscopies being performed. To date, 1,611 men have been included, and 23.5% have tested positive for hematuria. The additional molecular-based screening tests before referring to cystoscopy decreased the number of cystoscopies from 378 to 66 (↓82.5%) [8]. However, to date, no data related to the test’s overall sensitivity and specificity, as well as patient outcomes (i.e., bladder cancer detection, bladder cancer stage, and bladder cancer death) have been reported. Based on the lack of early detection assays in the clinic, an urgent clinical need exists to identify and validate a more robust non-invasive, urine-based assay for the early detection of BCa.

Oncuria is the first and only multiplex immunoassay designed to detect and manage BCa. Briefly, we applied two complementary techniques to profile urine samples from patients with and without BCa: gene expression (mRNA) of shed urothelial [9, 10] and glycoproteomics profiling of urine supernatant [11, 12]. Using advanced bioinformatics, we integrated the data from both datasets and identified a cancer-associated signature consisting of 19 candidate biomarkers. The clinical potential of the 19 candidate protein biomarkers was assessed using voided urine samples from an independent cohort of 127 patients (64 with BCa), employing commercial ELISA kits. Ten of the 19 biomarkers (SerpinA1 = A1AT, Angiogenin = ANG, Apolipoprotein = APOE, Carbonic Anhydrase 9 = CA9, Interleukin 8 = IL8, Matrix Metalloproteinase 9 = MMP9, Matrix Metalloproteinase 10 = MMP10, Plasminogen Activator Inhibitor 1/SerpinE1 = PAI1, Syndecan 1 = SDC1, and Vascular Endothelial Growth Factor = VEGF) were validated and incorporated into a diagnostic algorithm demonstrating a diagnostic sensitivity of 92% and specificity of 97% [13] The multiplex immunoassay has subsequently been validated in 7 independent cohorts [14–20]. Recently, Oncuria-Detect used to evaluate patients with hematuria (gross or microscopic) was tested in a real-world cohort of 931 patients with hematuria to identify de novo BCa resulting in an AUC of 0.84, sensitivity of 85%, specificity 72% and a negative predictive value (NPV) of 95% [21].

Several large population-based longitudinal cancer cohort studies [22–26] have prospectively collected voided urine samples in healthy individuals who went on to develop various disease states, including BCa. Such cohort studies can be leveraged to refine, test and validate novel assays for the early detection of BCa based on risk profile. We retrospectively and blindly evaluated Oncuria-Detect in these samples per PRoBE biomarker study design and reporting criteria [27] with the express purpose of testing an established diagnostic signature for the early detection of BCa.

Materials and methods

Study population

This nested case-control study pooled data from four large prospective cohort studies: Southern Community Cohort Study (SCCS; n = 84,797 participants), Singapore Chinese Health Study (SCHS; n = 63,257), Shanghai Women/Men Health Study (SWMHS; n = 136,480), and Multiethnic Cohort (MEC; n = 215,000), which recruited healthy participants. Across all cohorts there were 274 histopathologically-confirmed BCa cases (ICD-9 codes C670–C679) with at least a single spot overnight or first-morning voided urine sample from its participants. These cases were 1:1 age-, sex-, race-matched to individuals with no history of BCa over the same time period (i.e., controls), who also had at least a single spot over-night or first-morning voided urine sample [22–26]. All controls were alive at time of their matched case’s diagnosis. A total of 548 participants were included in the study.

All participants provided written informed consent and ethics committees at each study site approved human subject research for their cohorts. This pooled study was performed in accordance with the Declaration of Helsinki and approved by the Institutional Review Board at Cedars-Sinai (IRB # Study00001459).

Sample collection, storage, and multiplex immunoassay

Voided urine sample collection and storage follow standard protocols outlined in SCCS [22], SCHS [23], SWHS [24], SMHS [25], and MEC [26] studies. A single urine specimen per participant was available from 3 to 60 months [3‒12 months (n = 62), 13‒24 months (n = 50), 25‒36 months (n = 53), 37‒48 months (n = 57) and 49‒60 months (n = 52)] before the BCa diagnosis in cases, with approximately the same study enrollment time for the control participants. Voided urine was stored at -80 °C before biomarker measurement. The Oncuria^®-Detect immunoassay (Nonagen Bioscience, Los Angeles, CA) is a urine test developed to identify de novo BCa according to a 10-protein biomarker signature [20, 27]. With Oncuria-Detect, 10 distinct capture bead sets (plate #1 MMP9, IL8, VEGF, CA9; plate #2 A1AT, ANG, APOE, PAI1, SDC1; plate #3 MMP10) allow the 10 target analytes to be concurrently isolated and analyzed by incubation with a single urine sample. The test was performed on 300 µL of urine at a central laboratory (Cedars-Sinai Medical Center, Los Angeles, CA). Briefly, aliquots of frozen urine were passively thawed and handled on ice. Urine specimens were diluted 2-fold with assay diluent. Samples, standards, and controls (50 µL) were added to the assay kit’s 96-well plates (Plates #1‒3) in duplicate. The multiplex immunoassay was conducted according to the manufacturer’s instructions. A seven-point standard curve across the 4-log dynamic range was included in the current assay design. Plates were read on the Luminex^® 100/200 plate reader (Luminex Corp, Austin, TX).

Our study employed a population-based nested case–control design conforming to the PRoBE (Prospective Specimen Collection, Retrospective Blinded Evaluation) [28], STARD (Standards for Reporting of Diagnostic Accuracy Studies) [29], and TRIPOD [30] framework.

Data analysis

Baseline patient characteristics and biomarker concentrations were compared between cases and controls using Pearson’s Chi-square test for categorical variables and the Wilcoxon rank-sum test for continuous variables. Fisher’s exact test was used for categorical variables with small sample sizes. We evaluated the diagnostic properties of each individual biomarker for BCa detection using logistic regression models, with the biomarker as the predictor and cancer status as the outcome. Individual biomarker performance was assessed using 1,000 bootstrap resamples to evaluate the area under the receiver operating characteristic curve (AUC). The 95% confidence intervals for these metrics were calculated using the standard error of the mean.

Additionally, we examined the differences in each biomarker concentration by urine collection months using a generalized estimating equation (GEE) model with a Gaussian link function [31, 32]. The GEE model accounts for the correlation structure inherent in the matched case-control study design. The dependent variable is the biomarker concentration, and the independent variables included cancer status, urine collection months, and the biomarker, as well as a three-way-interaction between those terms. An exchangeable correlation structure was specified. Months between urine collection and the case’s diagnosis were modeled using restricted cubic splines with three knots to allow for non-linear relationships. We estimated the marginal mean difference in biomarker concentration for each biomarker separately over a period of 3 to 60 months between cases and controls. The Holm correction [33] was applied to control for the family wise error rate within each biomarker for 6 comparisons at the various timepoints.

To evaluate the predictive value of various biomarkers and demographic factors (age, sex, race, smoking status, pack years, and urine collection year) for cancer detection at different time intervals, we conducted a series of machine learning analyses at the following predefined time points: 0–2, 3–12, 13–24, 25–36, 37–48, 49–60, and 61 + months. For each time interval, the dataset was split into training (75%) and testing (25%) subsets, stratified by cancer status to maintain class balance. We systematically assessed all possible combinations of the 10 Oncuria biomarkers in conjunction with the demographic variables listed above, at each time point. This resulted in a total of 2,046 unique model combinations per time interval and 14,352 models across all seven time points for each machine learning algorithm evaluated, including logistic regression (glm), elastic net regression (glmnet), random forest (ranger), and decision trees (rpart). Prior to model fitting, the training data underwent centering and scaling of continuous variables, removal of zero-variance predictors, and creation of dummy variables for categorical predictors. Missing values in predictor variables were imputed using a bagged tree-based imputation method. Hyperparameter tuning was performed using 10-fold cross-validation within the training set, with a grid search of 10 hyperparameter value combinations per model. The optimal hyperparameter set, as determined by cross-validation performance, was then used to fit the final model, which was subsequently evaluated on the held-out test set. For classification, a probability threshold of 0.50 was applied, such that samples with predicted probabilities ≥ 0.5 were classified as cases, and those < 0.5 as controls. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.

All statistical analyses were conducted using R software [34] with two-sided tests and a significance level set at 0.05.

Data availability

Original data sets used in this report are available from the Corresponding Author upon reasonable request.

Results

Characteristics of the study sample

Of the 548 participants, the majority of the study participants were enrolled in the MEC (52%). The mean age at diagnosis for both cases and controls was 70 years (SD = 10). As expected, 80% were male given the higher prevalence of BCa in men compared to women, and most (64%) self-identified as Asian due to the enrollment locations of the cohorts included (Table 1). Cases were somewhat more likely to report a history of current or former smoking (68% vs. 60%, p = 0.058) with the median pack-year of 12 (interquartile range (IQR) = 0‒33) for cases and 7 (IQR = 0‒25) for controls (p = 0.018).

Table 1.

Participants demographic and clinical characteristics

Characteristic	Overall	Control	Case	p-value2
	N = 5481	N = 2741	N = 2741
Cohort				> 0.999
SCHS	70 (13%)	35 (13%)	35 (13%)
SWMHS	142 (26%)	71 (26%)	71 (26%)
MEC	286 (52%)	143 (52%)	143 (52%)
SCCS	50 (9.1%)	25 (9.1%)	25 (9.1%)
Age at Case Diagnosis*				> 0.999
Mean (SD)	70 (10)	70 (10)	70 (10)
Median [Q1, Q3]	71 [64, 78]	71 [64, 78]	71 [64, 78]
Sex*				> 0.999
Male	436 (80%)	218 (80%)	218 (80%)
Female	112 (20%)	56 (20%)	56 (20%)
Race*				> 0.999
White	134 (24%)	67 (24%)	67 (24%)
Black	32 (5.8%)	16 (5.8%)	16 (5.8%)
Asian	352 (64%)	176 (64%)	176 (64%)
Hispanic	27 (4.9%)	13 (4.7%)	14 (5.1%)
Other	3 (0.5%)	2 (0.7%)	1 (0.4%)
Smoking Status at baseline				0.054
Never	196 (36%)	109 (40%)	87 (32%)
Former/Current	351 (64%)	165 (60%)	186 (68%)
Unknown	1	0	1
Pack Years				0.018
Mean (SD)	118 (877)	130 (936)	106 (815)
Median [Q1, Q3]	11 [0, 31]	7 [0, 25]	12 [0, 33]
Unknown	6	2	4
Years between urine collection and case’s diagnosis				0.957
Mean (SD)	2.62 (1.66)	2.65 (1.87)	2.60 (1.43)
Median [Q1, Q3]	2.59 [1.02, 4.00]	2.42 [1.00, 4.00]	2.66 [1.18, 4.00]

^*, these demographic factors were matched

¹n (%)

²Pearson’s Chi-squared test; Wilcoxon rank sum test; Fisher’s exact test

Single urinary biomarkers

The median time of urine collection was 2.66 years before diagnosis (IQR = 1.18, 4.00 years) for cases and 2.42 years before the reference date (IQR = 1.00‒4.00 years) for controls (p = 0.957). Median urinary concentrations of seven of the 10 biomarkers were statistically significantly elevated at 3‒12 months for cases compared with controls [MMP9: 159 pg/mL vs. 32 pg/mL; IL8: 20 vs. 3; VEGF: 95 vs. 52; CA9: 3 vs. 2; PAI1: 22 vs. 12; APOE: 481 vs. 387; and A1AT: 50,806 vs. 18,479], while five of the 10 biomarkers were statistically significantly elevated 37‒48 months for cases compared with controls [MMP9: 77 vs. 20; IL8: 3 vs. 2; CA9: 1.53 vs. 1.52; PAI1: 12 vs. 9; and APOE: 400 vs. 308] and three of the 10 biomarkers were statistically significantly elevated at 49‒60 months for cases compared with controls [VEGF: 84 vs. 45; APOE: 400 vs. 245; and MMP10: 12 vs. 11] (Table 2). Levels of SDC1 (12 vs. 10, p = 0.057) and ANG (274 vs. 112, p = 0.055) at 3‒12 months, IL8 (4 vs. 2, p = 0.080) at 13‒24 months, A1AT (21,305 vs. 17,131, p = 0.058), ANG (127 vs. 56, p = 0.083) and MMP10 (12 vs. 10, p = 0.055) at 37‒48 months and ANG (382 vs. 164, p = 0.055) at 49‒60 months were non-statistically significantly elevated for cases and controls.

Table 2.

Mean and median urinary concentration of the 10 urinary biomarkers (pg/mL)

	3–12 months (n = 124)			13–24 months (n = 100)			25–36 months (n = 106)			37–48 months (n = 113)			49 + months (n = 93)
Characteristic	Control N = 62	Case N = 62	p-value 1	Control N = 50	Case N = 50	p-value 1	Control N = 53	Case N = 53	p-value 1	Control N = 57	Case N = 57	p-value 1	Control N = 52	Case N = 52	p-value 1
MMP9			0.014			0.15			0.6			0.028			0.2
Mean (SD)	680 (2,511)	2,415 (5,252)		1,183 (4,282)	997 (3,513)		1,153 (3,276)	1,164 (3,453)		456 (1,744)	1,178 (3,858)		523 (1,086)	4,190 (12,053)
Median (Q1, Q3)	32 (20, 280)	159 (20, 1,590)		20 (20, 159)	83 (20, 525)		78 (20, 248)	60 (19, 238)		20 (18, 127)	77 (20, 590)		67 (20, 403)	168 (20, 836)
IL8			< 0.001			0.080			0.9			0.048			0.15
Mean (SD)	46 (241)	632 (4,355)		7 (15)	165 (593)		16 (41)	14 (28)		15 (42)	33 (143)		13 (26)	66 (194)
Median (Q1, Q3)	3 (1, 15)	20 (3, 83)		2 (1, 6)	4 (1, 18)		2 (1, 6)	4 (1, 14)		2 (1, 5)	3 (1, 14)		3 (2, 13)	5 (2, 18)
VEGF			0.004			0.3			0.8			0.13			0.033
Mean (SD)	76 (93)	278 (763)		75 (66)	248 (922)		116 (147)	103 (115)		69 (85)	98 (103)		110 (246)	174 (345)
Median (Q1, Q3)	52 (12, 108)	95 (34, 161)		58 (28, 109)	65 (24, 191)		63 (12, 172)	71 (17, 134)		34 (14, 85)	65 (21, 121)		45 (15, 116)	84 (37, 168)
CA9			0.003			0.2			0.8			< 0.001			0.11
Mean (SD)	2 (2)	47 (180)		3 (5)	27 (151)		2.4 (1.9)	7.2 (27.6)		1.80 (1.40)	8.19 (31.55)		1.85 (1.41)	32.70 (182.04)
Median (Q1, Q3)	2 (2, 2)	3 (2, 15)		2 (2, 2)	2 (2, 5)		1.5 (1.5, 2.6)	1.5 (1.5, 3.4)		1.52 (1.26, 1.52)	1.53 (1.52, 3.30)		1.52 (1.43, 1.74)	1.52 (1.52, 3.62)
SDC1			0.057			0.8			0.5			0.091			0.12
Mean (SD)	8,082 (6,254)	10,508 (8,949)		8,833 (5,395)	9,910 (7,598)		10,151 (6,055)	9,353 (5,683)		7,710 (4,486)	9,850 (5,852)		8,782 (5,103)	10,782 (6,324)
Median (Q1, Q3)	6,656 (3,716, 10,314)	8,190 (5,160, 12,786)		7,214 (4,488, 11,966)	8,085 (3,943, 11,359)		9,379 (5,380, 14,291)	9,422 (4,146, 12,870)		7,794 (3,397, 11,468)	7,699 (5,747, 13,972)		7,699 (4,795, 11,718)	10,600 (6,464, 15,269)
PAI1			0.010			0.2			0.2			0.020			0.3
Mean (SD)	34 (93)	722 (3,798)		16 (17)	595 (4,154)		17 (20)	41 (90)		16 (18)	38 (106)		16 (17)	47 (135)
Median (Q1, Q3)	12 (7, 27)	22 (12, 59)		12 (7, 17)	15 (7, 32)		12 (7, 16)	13 (7, 37)		9 (4, 22)	12 (8, 31)		10 (6, 19)	12 (6, 27)
APOE			0.008			0.3			> 0.9			0.034			0.038
Mean (SD)	468 (463)	972 (1,559)		402 (358)	931 (3,106)		391 (334)	419 (434)		341 (238)	583 (745)		342 (386)	588 (894)
Median (Q1, Q3)	387 (179, 509)	481 (277, 1,054)		326 (153, 400)	400 (178, 780)		331 (121, 585)	294 (119, 530)		308 (131, 400)	400 (198, 672)		245 (122, 400)	400 (147, 604)
A1AT			0.001			0.3			0.7			0.058			0.3
Mean (SD)	121,673 (491,047)	123,710 (197,501)		42,005 (69,944)	88,376 (189,194)		75,302 (135,503)	88,594 (196,331)		26,295 (28,081)	124,547 (357,358)		73,110 (163,116)	63,754 (83,088)
Median (Q1, Q3)	18,479 (7,474, 51,891)	50,806 (14,773, 129,047)		17,619 (9,868, 53,013)	27,414 (9,370, 70,291)		31,030 (10,522, 65,690)	20,960 (8,808, 69,328)		17,131 (9,605, 25,704)	21,305 (12,679, 68,693)		26,995 (12,556, 64,300)	40,451 (16,932, 67,998)
ANG			0.055			0.6			> 0.9			0.083			0.055
Mean (SD)	383 (729)	1,610 (6,294)		373 (585)	839 (2,926)		436 (585)	493 (730)		265 (454)	396 (577)		439 (782)	945 (1,868)
Median (Q1, Q3)	112 (31, 481)	274 (52, 886)		119 (18, 517)	169 (15, 974)		100 (16, 663)	99 (14, 681)		56 (6, 293)	127 (19, 514)		164 (21, 488)	382 (93, 914)
MMP10			0.8			> 0.9			0.6			0.055			0.017
Mean (SD)	20 (54)	90 (364)		14 (19)	55 (313)		20 (67)	13 (16)		18 (56)	25 (55)		36 (182)	23 (41)
Median (Q1, Q3)	12 (7, 12)	12 (6, 28)		10 (6, 12)	12 (6, 12)		12 (7, 12)	11 (6, 12)		10 (6, 12)	12 (7, 14)		11 (7, 12)	12 (9, 18)

¹Wilcoxon rank sum test; bolded values highlight p < 0.05

We estimated performance metrics of each individual biomarker via a series of ROC curves (Supplemental Fig. 1). Across all biomarkers, CA9 had the best accuracy (0.59, 95%CI = 0.59‒0.59) and AUC (0.61, 95% CI = 0.60‒0.61).

Accuracy of single urinary biomarkers over time

In the GEE model using the full dataset, we were able to more accurately determine a difference in the individual biomarker levels between cases and controls at specific time periods (Fig. 1; Table 3). Based on the GEE model, while holding the parameters for biomarker and time constant, we estimated the marginal mean difference in biomarker concentration between Cases vs. Control. We showed statistical difference for ANG at 12 months (p = 0.046), APOE at 12 months (p < 0.001), MMP10 at 12 months (p = 0.009), PAI1 at 12 months (p = 0.005), SDC1 at 12 months (p = 0.003), 48 months (p = 0.029) and 60 months (p = 0.002), and VEGF at 12 months (p < 0.001).

Fig. 1.

Generalized Estimating Equation (GEE) model for each individual biomarker. Differences in each biomarker concentration were evaluated by urine collection months using a GEE model with a Gaussian link function. The GEE model accounts for the correlation structure inherent in the matched case-control study design. The dependent variable is the biomarker concentration, and independent variables included cancer status, urine collection months, and the biomarker, as well as a three-way-interaction between those terms

Table 3.

Changes in urinary biomarker concentrations in cases vs. Controls according to time before cancer diagnosis

Urine Collection Months	Estimate (95% CI)	p	Adjusted p*
ANG
0	2,026.06 (324.90, 3,727.22)	0.020	0.098
12	1,065.69 (281.33, 1,850.06)	0.008	0.046
24	246.23 (-303.09, 795.55)	0.380	> 0.999
36	-136.00 (-788.13, 516.12)	0.683	> 0.999
48	-19.45 (-281.27, 242.37)	0.884	> 0.999
60	345.42 (-233.45, 924.29)	0.242	0.969
APOE
0	1,080.40 (436.03, 1,724.77)	0.001	0.005
12	600.75 (362.80, 838.70)	< 0.001	< 0.001
24	185.81 (-387.14, 758.77)	0.525	> 0.999
36	-28.31 (-659.57, 602.95)	0.930	> 0.999
48	-13.36 (-264.44, 237.72)	0.917	> 0.999
60	115.65 (-363.98, 595.27)	0.637	> 0.999
CA9
0	509.57 (-100.54, 1,119.67)	0.102	0.508
12	145.74 (35.55, 255.94)	0.010	0.057
24	-157.00 (-546.70, 232.70)	0.430	0.860
36	-270.19 (-749.90, 209.52)	0.270	0.809
48	-167.15 (-375.01, 40.71)	0.115	0.508
60	43.54 (-320.85, 407.93)	0.815	0.860
IL8
0	1,329.28 (-88.90, 2,747.46)	0.066	0.397
12	574.02 (-39.61, 1,187.65)	0.067	0.397
24	-63.05 (-446.25, 320.15)	0.747	0.747
36	-333.37 (-862.32, 195.59)	0.217	0.650
48	-185.30 (-432.64, 62.04)	0.142	0.568
60	171.05 (-198.55, 540.66)	0.364	0.729
MMP10
0	547.01 (-46.83, 1,140.84)	0.071	0.355
12	166.56 (63.61, 269.50)	0.002	0.009
24	-151.72 (-537.90, 234.46)	0.441	0.883
36	-277.06 (-750.37, 196.26)	0.251	0.754
48	-182.30 (-389.12, 24.52)	0.084	0.355
60	22.03 (-347.23, 391.29)	0.907	0.907
MMP9
0	2,767.43 (878.43, 4,656.42)	0.004	0.025
12	1,204.32 (278.77, 2,129.87)	0.011	0.054
24	0.05 (-527.08, 527.19)	> 0.999	> 0.999
36	-90.70 (-706.69, 525.30)	0.773	> 0.999
48	1,088.82 (5.74, 2,171.91)	0.049	0.146
60	2,900.74 (672.10, 5,129.37)	0.011	0.054
A1AT
0	22,202.23 (-63,259.97, 107,664.42)	0.611	0.665
12	44,382.06 (-18,227.02, 106,991.14)	0.165	0.659
24	59,444.67 (-4,927.57, 123,816.92)	0.070	0.422
36	52,422.02 (-9,872.67, 114,716.72)	0.099	0.495
48	20,205.14 (-20,644.06, 61,054.34)	0.332	0.665
60	-24,554.63 (-60,846.00, 11,736.75)	0.185	0.659
PAI1
0	1,291.73 (681.78, 1,901.69)	< 0.001	< 0.001
12	697.52 (215.23, 1,179.80)	0.005	0.023
24	177.46 (-510.88, 865.80)	0.613	> 0.999
36	-112.48 (-764.12, 539.17)	0.735	> 0.999
48	-139.90 (-373.36, 93.56)	0.240	0.961
60	-36.62 (-526.80, 453.55)	0.884	> 0.999
SDC1
0	2,596.22 (805.14, 4,387.30)	0.004	0.018
12	1,523.40 (508.39, 2,538.42)	0.003	0.016
24	670.55 (-518.11, 1,859.21)	0.269	0.538
36	500.23 (-806.31, 1,806.77)	0.453	0.538
48	1,108.53 (115.82, 2,101.25)	0.029	0.086
60	2,104.48 (752.54, 3,456.41)	0.002	0.014
VEGF
0	715.83 (98.15, 1,333.50)	0.023	0.116
12	287.02 (155.28, 418.75)	< 0.001	< 0.001
24	-75.29 (-481.23, 330.65)	0.716	> 0.999
36	-231.25 (-723.58, 261.07)	0.357	> 0.999
48	-151.81 (-364.19, 60.57)	0.161	0.645
60	44.82 (-329.84, 419.49)	0.815	> 0.999

*Holm correction applied to account for multiple comparisons across time intervals

Predictive model at various timepoints

We assessed urine samples using various types of algorithms and among those various algorithms we assessed various combinations of biomarkers and demographic factors in predicting cancer status. Of the time frames analyzed (3‒12, 13‒24, 25‒36, 37‒48 and 49‒60 months), the timeframe that harbored the numerically highest AUCs was the 49‒60 months. Within this timeframe, the performance of Oncuria-Detect’s 10 biomarkers, along with the top 3 combinations of the 10 biomarker models for the early detection of BCa for each time frame are reported in Supplemental Table. The best model was not Oncuria-Detect, which has 10 biomarkers (49‒60 months before diagnosis – accuracy 0.77 with sensitivity 0.63 and specificity of 0.89). Notably, the best combinatorial model to predict subsequent BCa was IL8, CA9, PAI1, APOE and clinical features (Decision Tree), which had an AUC of 0.99, accuracy of 0.94 with a with a sensitivity of 0.88 and a specificity of 1.00. Interestingly, the addition of clinical factors to the model slightly improved its performance (AUC went from 0.94 to 0.99). Additional examples of the models being improved by the addition of the clinical features included: CA9, PAI1, APOE (Random Forest) at 37–48 months and VEGF, SDC1, A1AT, ANG, MMP10 (Random Forest) 25–36 months. While some examples where the addition of the clinical features reduced the performance of the models included PAI1, APOE (Random Forest) and CA9, PAI1, APOE (Random Forest) at 49–60 months.

Discussion

Bladder cancer remains a significant public health challenge, particularly due to late-stage diagnosis in a substantial subset of patients [35]. Although hematuria is a hallmark symptom, it is neither sensitive nor specific for malignancy, and early-stage disease may present asymptomatically. As such, there is a pressing need for reliable, non-invasive biomarkers capable of detecting disease earlier in its clinical course, particularly among high-risk populations. Earlier and ongoing BCa screening studies identified individual urinary proteins (e.g., NMP22 [36, 37] and bladder tumor antigen [38]), and chromosomal [39] and DNA characteristics [40] of shed urothelial cells as biomarkers of BCa. Several of these approaches have gained regulatory approval for diagnostic use [41]. A recent study from Germany compared various BCa rapid diagnostic tests and concluded that they appear best-suited for identifying high-grade disease [42]. Additionally, the low overall BCa prevalence in the general populace may restrict the positive predictive value of screening tests. The Oncuria-Detect multiplex assay was previously demonstrated to reliably detect low-grade BCa [20, 21], with better sensitivity than current single-protein assays.

In the current large-scale study pooling data across four independent prospective cohorts, we evaluated the performance of Oncuria-Detect, a multiplex immunoassay measuring a 10-biomarker signature associated with BCa. This analysis confirmed the assay’s reliability for early detection, with individual analytes showing statistically significant elevations well before clinical diagnosis for the following biomarkers ANG at 12 months, APOE at 12 months, CA9 at 12 months, MMP10 at 12 months, MMP9 at 12 months, 48 months and 60 months, PAI1 at 12 months, SDC1 at 12 months, 48 months and 60 months, and VEGF at 12 months. The AUC of individual markers was lower than the full Oncuria-Detect 10-protein panel and other biomarker combinations, highlighting the molecular complexity of BCa that needs to be considered with any urine-based diagnostic test. Most importantly, we noted that a smaller subset of biomarkers from Oncuria-Detect performed better at the early detection of BCa than the full multi-marker panel of Oncuria-Detect. Specifically, when molecular biomarkers (IL8, CA9, PAI1 and APOE) were assessed in voided urine at 49‒60 months before diagnosis of BCa, the combined model achieved an AUC of 0.99, accuracy of 0.94 with a sensitivity of 88% and specificity of 100%. This simplified panel offers notable advantages: reduced cost of manufacture, easier integration into clinical workflows, and enhanced scalability for large population screening initiatives. Therefore, these early preliminary findings hint at the assay’s potential to identify at-risk individuals, such as smokers over 50 years of age with >20 pack-years of tobacco exposure, prior to the onset of symptoms. BCa may take many years to develop into frank disease. The potential utility of identifying BCa risk with Oncuria-Detect long before clinical symptoms appear is supported by the large Golestan Cohort Study, which detected BCa-associated telomerase reverse transcriptase (TERT) promoter mutations in urine samples up to 10 years ahead of BCa diagnosis [43]. The apparent bimodal elevation of some biomarkers in the current study (high at 0‒12 months, receding at 24‒48 months, elevated at 60 months pre-diagnosis) highlights the need to better understand the temporal nature of BCa biomarker expression throughout BCa development.

To date, Oncuria-Detect has been analytically [24] and clinically validated [20] and launched commercially as a laboratory-developed test (LDT) for assessing risk in patients with hematuria or upper tract abnormalities on imaging for urothelial carcinoma. The Oncuria platform has also been extended for additional clinical use cases: Oncuria–Monitor for surveillance of BCa recurrence [14] and Oncuria–Predict for evaluating the risk of recurrence following intravesical BCG therapy in patients with NMIBC [44].

Efforts to develop effective screening approaches for BCa remain limited. In one pivotal randomized study, men ≥ 50 years with ≥ 40 pack-years of smoking history were randomized to screening with urinary dipstick testing versus no monitoring [7]. Among the screened group, the proportion of muscle-invasive disease was significantly reduced (10% vs. 60%; p = 0.002), and at 14 years of follow-up, no deaths from BCa were reported in the screened arm, compared with a 20.4% disease-specific mortality in the control arm (p = 0.02). Despite these promising results, urinary dipstick testing demonstrated limited performance, with a sensitivity of only 50% and a specificity of 54% for BCa detection [45].

The Bladder Cancer Urine Marker Project (BLU-P) introduced a more sophisticated approach, incorporating urinary NMP22, FGFR3 mutations, microsatellite analysis, and MLPa (a methylation-specific assay) [8]. However, this study’s performance metrics and outcome data remain pending, leaving its clinical utility unconfirmed.

In our study, voided urine samples were collected prospectively before clinical outcome ascertainment, with outcome verification performed through linkage to electronic health records and national cancer registries. This represents the largest longitudinal collection of urine specimens from the general population—spanning up to five years before BCa diagnosis—currently available. While these preliminary results of this study are encouraging, several limitations should be acknowledged. First, only a single urine specimen per participant was analyzed, thereby precluding temporal assessment of biomarker trajectories and intra-patient variability over time. Second, differences in urine collection, processing, and storage across cohorts may have introduced pre-analytical variability and potential batch effects. Although standardized protocols were followed where possible, complete harmonization across all samples could not be ensured. Third, detailed clinical data regarding hematuria type (microscopic versus gross) and tumor characteristics (grade and stage) were unavailable. Finally, the relatively small sample size and retrospective nature of the analysis may increase the risk of model overfitting and limit the generalizability of the findings.

In the early detection setting, several subsets of two to four analytes—IL8, CA9, PAI1, APOE; PAI1, APOE or CA9, PAI1, APOE—were associated with enhanced diagnostic accuracy, reaching 88–100% sensitivity and 89‒100% specificity. This observation is consistent with the biological rationale underlying the Oncuria-Detect panel, which comprises ten protein biomarkers implicated in distinct stages of tumorigenesis, including initiation, promotion, and progression. These findings highlight the potential of tailored analyte subsets within the Oncuria platform to improve early detection performance in BCa screening.

The ability to detect biomarker alterations three to five years prior to clinical diagnosis has significant potential clinical implications. Earlier identification of individuals with molecular evidence of disease could enable a shift in stage at presentation—from muscle-invasive (T2–T4) to non–muscle-invasive (Ta/T1) bladder cancer—thereby improving prognosis and expanding therapeutic options. In the future, such early detection could also provide a critical window for interventional strategies aimed at cancer prevention, should effective chemopreventive or lifestyle-based approaches become available. The most appropriate target population for such screening would likely include individuals over 50 years of age with a history of heavy tobacco exposure, given their elevated risk for bladder cancer. While the current findings are promising, comprehensive modeling of screening frequency, healthcare utilization, and cost-effectiveness will be required to determine the real-world feasibility of implementing such an approach. At this stage, these results should be viewed as an important proof of concept supporting the biological and clinical plausibility of preclinical bladder cancer detection.

These early results of our multiplex immunoassay demonstrates a potential as a non-invasive, urine-based diagnostic platform for early detection of BCa in high-risk populations. Additional studies are underway to confirm these results. The ability to detect early-stage tumors—when treatment is more effective—could translate into improved outcomes and reduced mortality.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

A1AT

Serpin A1

ANG

Angiogenin

APOE

Apolipoprotein

AUC

Area under the receiver operating characteristic curve

BCa

Bladder cancer

CA9

Carbonic Anhydrase 9

GEE

Generalized estimating equation

IL8

Interleukin 8

IQR

Interquartile range

LDT

Laboratory-developed test

MEC

Multiethnic Cohort

MIBC

Muscle-invasive bladder cancer

NMIBC

Non-muscle-invasive bladder cancer

MMP9

Matrix Metalloproteinase 9

MMP10

Matrix Metalloproteinase 10

NLST

National Lung Screening Trial

NPV

Negative predictive value

PAI1

Plasminogen Activator Inhibitor 1/Serpin E1

PRoBE

Prospective Specimen Collection, Retrospective Blinded Evaluation

SCCS

Southern Community Cohort Study

SCHS

Singapore Chinese Health Study

SDC1

Syndecan 1

SWMHS

Shanghai Women/Men Health Study

VEGF

Vascular Endothelial Growth Factor

Author contributions

Study concept and design – Furuya, Luu, Rosser. Acquisition of samples and data – Lynne Wilkens, Loïc Le Marchand, Gong Yang, Jian-Min Yuan, Woon-Puay Koh, Martha Shrubsole. Data analysis and interpretation – Luu and Pagano. Drafting of manuscript – Rosser, Furuya, Luu, Figueiredo. Critical revision of the manuscript for important technical content – All authors. Obtaining funding – Rosser, Furuya. Administrative technical and material support – Rosser. Supervision – Rosser.

Declarations

Ethics approval and consent to participate

Cedars Sinai Local ethics review board approved. Subjects gave written consent.

Consent for publication

Not applicable.

Competing interests

Dr. Charles Rosser is an officer of Nonagen Bioscience. All other authors declare that they have no competing interests.