Related Experiment Videos
Neural network using combined urine nuclear matrix protein-22, monocyte chemoattractant protein-1 and urinary
Sijo J Parekattil1, Hugh A G Fisher, Barry A Kogan
1Division of Urology, Albany Medical Center, New York, USA.
Purpose:
We developed a neural network to identify patients with bladder cancer more effectively than hematuria and cytology. The algorithm is based on combined urine levels of nuclear matrix protein-22, monocyte chemoattractant protein-1 and urinary intercellular adhesion molecule-1.
Materials And Methods:
A randomized double-blinded study of voided urine from 253 patients undergoing outpatient cystoscopy was performed. Of the patients 27 had bladder cancer on biopsy and 5 had muscle invasion. Urine tumor markers were measured using sandwich-enzyme-linked immunosorbent assay kits. Urine from patients with bladder cancer on cystoscopy was compared to urine from controls with negative cystoscopy results. An algorithm was created with 3 sets of cutoff values modeled to be 100% sensitive for superficial bladder cancer, 100% specific for superficial cancer and 100% specific for muscle invasive cancer, respectively. We compared our model to hematuria and cytology.
Results:
For the hematuria dipstick test sensitivity, specificity, positive and negative predictive values were 92.6%, 51.8%, 18.7% and 98.2%, respectively. For atypical cytology sensitivity, specificity, positive and negative predictive values were 66.7%, 81%, 29.5% and 95.3%, respectively. For the sensitive model set sensitivity, specificity, positive and negative predictive values were 100%, 75.7%, 32.9% and 100%, respectively. For the specific model set sensitivity, specificity, positive and negative predictive values were 22.2%, 100%, 100% and 91.5%, respectively. For the muscle invasive model set sensitivity, specificity, positive and negative predictive values were 80%, 100%, 100% and 99.6%, respectively. The standard bladder tumor evaluation of 253 patients costs 61,054 US dollars but 36,450 US dollars using our model.
Conclusions:
Our algorithm is superior to conventional screening tests for bladder cancer. The model identifies patients who require cystoscopy, those with bladder cancer and those with muscle invasive disease. It provides possible savings over current screening methods. The potential loss of other information by not performing cystoscopy was not evaluated in our study.