887 USING ARTIFICIAL INTELLIGENCE AND MACHINE-LEARNING ALGORITHMS WITH GENE EXPRESSION PROFILING TO PREDICT SUPERFICIAL BLADDER CANCER RECURRENCE AT INITIAL PRESENTATION

INTRODUCTION AND OBJECTIVES: 50%-70% of patients with superficial (Ta/T1, N0M0) bladder cancer (SBC) recur within 5 years of initial presentation. Tumor grade and multifocality are routinely used indicators of recurrence. SBCs’ propensity to recur persistently necessitates intense follow up and invasive treatment. It is therefore crucial to objectively determine SBCs’ recurrence potential to identify patients at greatest risk and minimize invasive follow up schedules for those harboring relatively indolent disease. This study used a machinelearning algorithm to identify SBC genes at initial presentation that were most predictive of recurrence, and used them in a molecular signature to predict recurrence risk within 5 years after TURBT. METHODS: Whole genome profiling was performed on 112 frozen SBCs obtained at first presentation by TURBT on Illumina Human WG-6 BeadChips. A genetic programming (GP) algorithm was used to evolve classifier mathematical models for outcome prediction. Cross-validation-based resampling and gene usage frequencies were used to identify the most prognostic genes, which were combined into rules used in a voting algorithm to predict a sample’s target class. Key genes were validated by quantitative PCR. RESULTS: 88 (79%) patients recurred within 5 years of initial presentation. A GP algorithm was used to select a minimal set of markers grouped as a classifier for predicting recurrence, and crossvalidation estimated its robustness by analyzing its ability to generalize to unseen samples. The classifier set included 21 genes that could predict recurrence. Quantitative PCR was done on a subset of 100 patients for these genes. With amplicon sizes limited to 100 bases and Ct values 35 not being considered, a 4-fold cross-validation (n 83) resulted in a 5-gene combined rule that incorporated a voting algorithm to yield 77% sensitivity and 85% specificity in predicting recurrence in the training set. The respective values in the test set were 69% and 62%. A singular 3-gene rule was also constructed that predicted recurrence with 80% sensitivity and 90% specificity in the training set. The respective values in the test set were 71% and 67%. CONCLUSIONS: Using primary SBCs from initial occurrences, GP identified transcripts in a reproducible fashion that were predictive of recurrence. These findings could potentially impact SBC management, including surveillance frequency, administration of adjuvant therapy, and selection of candidates for an expectant approach.