1773P Prediction of chemotherapy response in muscle-invasive bladder cancer: A machine learning approach (vol 32, S1348, 2022)

The authors regret that there are certain differences in cohort characteristics which need to be clarified in the title, methods and conclusion of this abstract. The correct title of the abstract is now given above. The abstract itself should read as follows: Background: The response to cisplatin-based chemotherapy in patients with muscle-invasive bladder cancer (MIBC) is variable, highlighting the need for biomarkers to drive patient selection in order to minimize treatment-related morbidity and optimize oncologic outcomes. We aimed to develop an artificial intelligence-based platform leveraging routine pretreatment histopathology specimens to predict chemotherapy response. Methods: Using the Cancer Genome Atlas bladder cancer dataset, 59 patients with MIBC who received gemcitabine and cisplatin (GC) from were identified. A deep learning-based algorithm was employed to segment nuclei from hematoxylin and eosin-stained histopathology specimens and subsequently to extract quantitative, cell-type-specific features falling in one of three categories: nuclear geometry, cellular spatial arrangement, and tissue heterogeneity. These features were subsequently correlated with cancer-specific survival (CSS) utilizing a multivariable Cox proportional hazards (CPH) model in order to construct a signature associated with chemotherapy response. Lastly, univariate CPH regression was conducted to identify the features most associated with response. Results: The multivariable CPH model incorporating features from all three aforementioned categories was predictive of response with a concordance index of 0.70 [95% CI 0.53-0.87]. Kaplan-Meier analysis demonstrated the model was able to separate the cohort robustly with a statistically significant hazard ratio of 0.265 [95% CI 0.07, 0.98] (p1/40.03) for predicted responders as compared to predicted non-responders. Features most correlated with CSS were those related to tissue heterogeneity and diversity of nuclear axis measurements, and increased heterogeneity was associated with worse CSS. Conclusions: An artificial intelligence-based platform utilizing routine histopathology specimens may help identify patients The authors would like to apologise for any inconvenience caused.