Machine learning-based prediction of cytokine release syndrome post CAR-T cell therapy.

Abstract

e13622 Background: Chimeric antigen receptor T-cell (CAR-T) therapy has revolutionized treatment of hematologic malignancies; however, severe toxicities like cytokine release syndrome (CRS) remain a limitation to broader utilization. Using a large dataset of patients treated with CAR-T, we developed a machine learning (ML) model leveraging multi-model patient characteristics to predict the risk of CRS onset with clinically meaningful lead times. Methods: A prospectively maintained dataset of CAR-T patients treated at Mayo Clinic was used. Baseline disease/patients characteristics, labs, ECG, echocardiogram, and vital signs (including data from remote patient monitoring program) were collected. A 24-hour rolling window was used to generate daily averages, maximum, minimum, and deltas for each parameter. ECG and echocardiogram data were forward-imputed; vital signs and labs were not imputed. Each day-post-infusion served as a potential event (CRS onset vs. no CRS). A gradient boosting classifier was trained and tested on a subset of patients (n = 351) using 1,000 bootstrap iterations, evaluating 24-, 48-, and 72-hour lead times. Shapley-based feature importance guided simplification to an 11-feature model, externally validated on an independent cohort (n = 51). Models’ performance assessed using the area under the receiver operating characteristic curve (AUC). Results: A total of 402 patients treated with CAR-T were examined. Median age was 64 years and 36% were female. Primary diagnosis was B-cell lymphoma (n = 240) or multiple myeloma (n = 162). A total of 226,332 vital sign measurements and 372,871 lab values, as well as ECG and echocardiogram data, were collected for 30 days post-infusion. Overall, 79.8% of patients experienced CRS within 30 days. The full ML model demonstrated AUCs of 0.90 (95% CI 0.86–0.93), 0.83 (95% CI 0.77–0.88), and 0.82 (95% CI 0.77–0.87) for 24-, 48-, and 72-hour predictions of CRS, respectively. Including “day of CRS onset post CAR-T” as a feature improved the 72-hour AUC to 0.87 (95% CI 0.83–0.90). A model restricted to vital signs alone showed AUCs of 0.88, 0.75, and 0.71 for 24-, 48-, and 72-hour horizons. An 11-feature model—focusing on temperature, blood pressure, fibrinogen, CRP, serum sodium level, QRS interval, and lymphocyte counts—maintained high accuracy at 24 hours (AUC 0.90, 95% CI 0.86–0.93) and was validated externally using an independent dataset with an AUC of 0.92. Conclusions: Our ML model provides a reliable approach for predicting if/when CRS will occur, with up to three days of lead time before clinical manifestation, offering an opportunity for proactive management. The robust performance of a parsimonious 11-feature model supports practical adoption across clinical settings and real-time risk stratification to enhance patient safety. Prospective validation and integration into clinical workflows will be critical to translating these findings into practice.