Machine learning–driven prognostication of overall survival at 6 months after CAR-T cell therapy

Abstract

Abstract Background Chimeric-antigen-receptor (CAR) T-cell therapy has transformed outcomes for relapsed or refractory hematologic malignancies, yet early identification of patients at highest risk for death remains elusive. A prognostic model that leverages only the information routinely available at the time of CAR-T evaluationcould guide pre-infusion counseling, intensity of follow-up, and resource allocation. Methods We analyzed a prospectively maintained cohort of CAR-T recipients treated at Mayo Clinic to develop and test our model. For each patient, data (vital signs, lab results, cardiac assessments [ECG and echocardiography], and PET/CT images) from initial CAR-T evaluation (up to 60 days prior to CAR-T infusion) was considered. Vital signs and labs were left un-imputed, while ECG, echocardiographic parameters, and PET/CT features were forward-filled from the preceding 60 days. Each patient-day was labeled as “outcome present” or “no outcome,” enabling supervised training of gradient-boosting models to predict overall survival (OS) at 6 months post CAR-T infusion. Model performance was evaluated using the area under the ROC curve (AUC) with 10,000-fold bootstrapping with data shuffling between each iteration to derive confidence intervals for stability. To enhance interpretability, we also developed a simplified model using the top 19 features (determined by Shapley value-based importance), aiming to preserve predictive accuracy with a smaller feature set. Results A total of 570 CAR-T cell treated patients were included in this study (476 patients in the training cohort and 94 patients in the validation cohort). The median age was 64 years and 61% males. The primary diagnoses were B-cell lymphoma (324 patients) and multiple myeloma (246 patients). A total of 4,851 individual vital-sign readings, 11,781 laboratory measurements were included. PET/CT images were available in 318 patients. The ML model demonstrated high discriminative ability for predicting OS at 6 months post CAR-T infusion. Vital signs alone produced a mean AUC of 0.63; radiomics features alone produced a mean AUC of 0.71, while laboratory tests alone produced a mean AUC of 0.84. The features most contributory to each single-modality model were then combined into a simplified multimodal model with a total of 19 features, which produced a mean AUC of 0.87, with a 95% CI of 0.73 - 0.97. The 19 features included in the simplified model were: LDH, ferritin, uric acid, eosinophils, fibrinogen, D-dimer, heart rate, mean arterial blood pressure, QTc, QT interval, and the radiomics features of sphericity, first-order Kurtosis, first-order mean, first-order energy, first-order total energy, shape maximum2D diameter (sagittal plane), SUV CoV, SUVp90, shape flatness. This multimodal predictive model was then applied to the validation cohort (uninvolved with model training or testing), which produced an AUC of 0.93, with a 95% CI of 0.84 - 0.98. Conclusions Our ML-driven approach successfully predicts overall survival 6 months post CAR-T infusion, highlighting those individuals at most risk of poor outcomes at the time of CAR-T evaluation. The strong performance of the multimodal version of the model suggests that each data domain contributes complementary resolving power, and by including novel data types, such as PET CT radiomics features, the most accurate predictions could be realized. By providing accurate, near-real-time risk stratification for poor outcomes within 6 months of CAR-T infusion, the model could facilitate proactive patient management. Prospective studies of the integration of this ML model into CAR-T treatment workflows will be crucial to confirm its clinical benefits and to facilitate its adoption in practice.