A holistic comparative analysis of diagnostic tests for urothelial carcinoma: a study of Cxbladder Detect, UroVysion® FISH, NMP22® and cytology based on imputation of multiple datasets.

BACKGROUND: Comparing the relative utility of diagnostic tests is challenging when available datasets are small, partial or incomplete. The analytical leverage associated with a large sample size can be gained by integrating several small datasets to enable effective and accurate across-dataset comparisons. Accordingly, we propose a methodology for a holistic comparative analysis and ranking of cancer diagnostic tests through dataset integration and imputation of missing values, using urothelial carcinoma (UC) as a case study. METHODS: Five datasets comprising samples from 939 subjects, including 89 with UC, where up to four diagnostic tests (cytology, NMP22®, UroVysion® Fluorescence In-Situ Hybridization (FISH) and Cxbladder Detect) were integrated into a single dataset containing all measured records and missing values. The tests were firstly ranked using three criteria: sensitivity, specificity and a standard variable (feature) ranking method popularly known as signal-to-noise ratio (SNR) index derived from the mean values for all subjects clinically known to have UC versus healthy subjects. Secondly, step-wise unsupervised and supervised imputation (the latter accounting for the 'clinical truth' as determined by cystoscopy) was performed using personalized modelling, k-nearest-neighbour methods, multiple logistic regression and multilayer perceptron neural networks. All imputation models were cross-validated by comparing their post-imputation predictive accuracy for UC with their pre-imputation accuracy. Finally, the post-imputation tests were re-ranked using the same three criteria. RESULTS: In both measured and imputed data sets, Cxbladder Detect ranked higher for sensitivity, and urine cytology a higher specificity, when compared with other UC tests. Cxbladder Detect consistently ranked higher than FISH and all other tests when SNR analyses were performed on measured, unsupervised and supervised imputed datasets. Supervised imputation resulted in a smaller cross-validation error. Cxbladder Detect was robust to imputation showing a 2% difference in its predictive versus clinical accuracy, outperforming FISH, NMP22 and cytology. CONCLUSION: All data analysed, pre- and post-imputation showed that Cxbladder Detect had higher SNR and outperformed all other comparator tests, including FISH. The methodology developed and validated for comparative ranking of the diagnostic tests for detecting UC, may be further applied to other cancer diagnostic datasets across population groups and multiple datasets.

A segregation index combining phenotypic (clinical characteristics) and genotypic (gene expression) biomarkers from a urine sample to triage out patients presenting with hematuria who have a low probability of urothelial carcinoma.

BACKGROUND: Hematuria can be symptomatic of urothelial carcinoma (UC) and ruling out patients with benign causes during primary evaluation is challenging. Patients with hematuria undergoing urological work-ups place significant clinical and financial burdens on healthcare systems. Current clinical evaluation involves processes that individually lack the sensitivity for accurate determination of UC. Algorithms and nomograms combining genotypic and phenotypic variables have largely focused on cancer detection and failed to improve performance. This study aimed to develop and validate a model incorporating both genotypic and phenotypic variables with high sensitivity and a high negative predictive value (NPV) combined to triage out patients with hematuria who have a low probability of having UC and may not require urological work-up. METHODS: Expression of IGFBP5, HOXA13, MDK, CDK1 and CXCR2 genes in a voided urine sample (genotypic) and age, gender, frequency of macrohematuria and smoking history (phenotypic) data were collected from 587 patients with macrohematuria. Logistic regression was used to develop predictive models for UC. A combined genotypic-phenotypic model (G + P INDEX) was compared with genotypic (G INDEX) and phenotypic (P INDEX) models. Area under receiver operating characteristic curves (AUC) defined the performance of each INDEX: high sensitivity, NPV >0.97 and a high test-negative rate was considered optimal for triaging out patients. The robustness of the G + P INDEX was tested in 40 microhematuria patients without UC. RESULTS: The G + P INDEX offered a bias-corrected AUC of 0.86 compared with 0.61 and 0.83, for the P and G INDEXs respectively. When the test-negative rate was 0.4, the G + P INDEX (sensitivity = 0.95; NPV = 0.98) offered improved performance compared with the G INDEX (sensitivity = 0.86; NPV = 0.96). 80% of patients with microhematuria who did not have UC were correctly triaged out using the G + P INDEX, therefore not requiring a full urological work-up. CONCLUSION: The adoption of G + P INDEX enables a significant change in clinical utility. G + P INDEX can be used to segregate hematuria patients with a low probability of UC with a high degree of confidence in the primary evaluation. Triaging out low-probability patients early significantly reduces the need for expensive and invasive work-ups, thereby lowering diagnosis-related adverse events and costs.