A Machine Learning Model Optimized for Local Data Stratifies Patients for the Adoptive Cell Therapy with Tumor Infiltrating Lymphocytes in Bladder Tumors

Adoptive cell therapy with tumor-infiltrating lymphocytes (ACT-TILs) involves autologous TILs that are expanded ex vivo and then reinfused into the patient as a personalized form of therapy. For bladder cancer patients, there is a unique opportunity to deliver TILs locally, through intravesical administration, without systemic cytotoxic chemotherapy to induce lymphodepletion. Improved selection of patients for ACT-TILs by predicting TIL expansion may prevent excessive treatment-related costs and delays in treatment of patients whose bladder cancer specimens would not yield successful TIL growth. In this study, we developed a machine-learning (ML) model optimized for local data of a medium size to determine a minimal robust feature combination (demographic, clinical, and biological tumor specimen-based) that is predictive of whether TILs can be expanded from a resected bladder cancer. The study uses a retrospectively identified set of data from bladder cancer patients at Moffitt Cancer Center that were assigned to internal training/testing or blinded validation cohorts. Using random forest method (RF) to identify a combination of robust predictive features, support vector machine (SVM) model to determine the optimal classification hyperparameters, and Matthews correlation coefficient (MCC) method to adjust the decision-boundary threshold for imbalanced data, our model yielded AUC=0.771 for the internal testing cohort and AUC=0.828 for the blinded validation cohort. Thus, our ML model optimized for local data of medium size has favorable performance metrics for predicting TIL expansion from a given tumor. This computational predictor can serve as a clinical supportive tool to determine which patients are candidates for ACT-TILs.