AUTHORS
Sapir Israeli, PhD, Jennifer Liu2, Mohammad Alhomoud, MD3, Elan Baskir, MD4, Candela Ceballos5, Alessandro Criscimanna6, Sigrun Einarsdottir, MD, PhD7, Silvia Escribano Serrat, MD8, Marina Gomez-Llobell, MD9, Richard J. Lin, MD, PhD10, Maria Lia Palomba, MD11, Jae H Park, MD12, Miguel-Angel Perales, MD13, Gilles Salles I, MD, PhD14, Craig Sauter, MD15, Michael Scordo, MD16, Gunjan L. Shah, MD, MS17, Yannis K. Valtis I, MD18, Ioanna Kosmidou, MD19, Anthony Yu, MD20, Roni Shouval Jr, MD, PhD21, Brian C Shaffer, MD, MS22
AFFILIATIONS
1 Memorial Sloan Kettering Cancer Center, New york, United States
2 Memorial Sloan Kettering Cancer Center, New york, United States
3 Memorial Sloan Kettering Cancer Center, New york, United States
4 Memorial Sloan Kettering Cancer Center, New york, United States
5 Memorial Sloan Kettering Cancer Center, New york, United States
6 IRCCS San Raffaele Scientific Institute, Milan, Italy
7 Memorial Sloan Kettering Cancer Center, New york, United States
8 Memorial Sloan Kettering Cancer Center, New york, United States
9 Memorial Sloan Kettering Cancer Center, New york, United States
10 Memorial Sloan Kettering Cancer Center, New york, United States
11 Memorial Sloan Kettering Cancer Center, New york, United States
12 Memorial Sloan Kettering Cancer Center, New york, United States
13 Memorial Sloan Kettering Cancer Center, New york, United States
14 Memorial Sloan Kettering Cancer Center, New york, United States
15 Memorial Sloan Kettering Cancer Center, New york, United States
16 Memorial Sloan Kettering Cancer Center, New york, United States
17 Memorial Sloan Kettering Cancer Center, New york, United States
18 Memorial Sloan Kettering Cancer Center, New york, United States
19 Memorial Sloan Kettering Cancer Center, New york, United States
20 Memorial Sloan Kettering Cancer Center, New york, United States
21 Memorial Sloan Kettering Cancer Center, New york, United States
22 Memorial Sloan Kettering Cancer Center, New york, United States
ABSTRACT
Background: Cardiovascular complications are increasingly recognized after chimeric antigen receptor T-cell (CAR T-cell) therapy, yet preinfusion risk stratification remains limited. Electrocardiograms (ECGs) are routinely obtained before treatment, but whether artificial intelligence (AI)-derived ECG features can identify patients at increased risk for cardiovascular toxicity and adverse outcomes is unknown.
Methods: We retrospectively studied consecutive patients with non-Hodgkin lymphoma treated with commercially available CD19-directed CAR T-cell therapy at a single center between 2016 and 2025. A pre-trained Compact Convolutional Transformer (CCT) deep-learning model, previously developed in a large pan-cancer cohort to estimate future atrial fibrillation (AF) from raw 12-lead ECG waveforms, was applied without retraining or fine-tuning to the preinfusion ECG obtained closest to CAR T-cell infusion. The continuous model output was termed the AI-ECG score. Associations with baseline cardiovascular and inflammatory characteristics, cytokine release syndrome (CRS), 30-day cardiovascular events (arrhythmias, heart failure, acute coronary syndrome, and cardiogenic shock), and survival were evaluated using correlation, logistic, Cox proportional hazards, and Fine-Gray regression analyses. Multivariable models adjusted for age, ejection fraction, performance status, LDH, CAR-T product, bridging therapy, and cardiovascular comorbidity burden.
Results: A total of 563 patients with an evaluable preinfusion ECG were included. Median age was 65.8 years; 81.9% had large B-cell lymphoma, 10.5% mantle cell lymphoma, and 7.7% indolent B-cell lymphoma. CAR T-cell products included axi-cel (37.5%), liso-cel (43.3%), tisa-cel (14.0%), and brexu-cel (5.2%). Within 30 days after infusion, approximately 8% experienced a composite cardiovascular event, primarily atrial arrhythmias, and 36.05% developed grade ≥2 CRS. Median follow-up was 30.5 months, with 1-year overall survival (OS), progression-free survival (PFS), and non-relapse mortality (NRM) of 72.9%, 55.1%, and 4.4%, respectively.
Preinfusion ECGs were obtained a median of 4 days (IQR, 1–28.5) before CAR-T infusion. AI-ECG scores ranged from 0 to 1 (median 0.29; IQR, 0.13–0.53) and were not associated with ECG timing (Spearman ρ=0.06; P=0.16). Higher scores were associated with baseline cardiovascular comorbidities, including prior arrhythmia, heart failure, hypertension, coronary artery disease, diabetes mellitus, and reduced left ventricular ejection fraction, and correlated with higher baseline CRP, IL-6, ferritin, TNF-α, LDH, and IL-10 (all FDR q<0.05).
Each 0.1-point increase in AI-ECG score independently predicted grade ≥2 CRS (adjusted OR [aOR], 1.08; 95% CI, 1.02–1.17; P=0.016), and the composite 30-day cardiovascular outcome (aOR, 1.14; 95% CI, 1.01–1.28; P=0.037). Composite cardiovascular events occurred in 5% versus 11%, including atrial arrhythmias in 4.3% versus 10.3%, among patients with AI-ECG scores below versus above the median, respectively.
Higher AI-ECG scores also independently predicted inferior OS (adjusted HR [aHR], 1.10 per 0.1-point increase; 95% CI, 1.06–1.16; P<0.001), inferior PFS (aHR, 1.06; 95% CI, 1.01–1.10; P=0.010), and increased NRM (aHR, 1.22; 95% CI, 1.09–1.35; P<0.001), but not relapse. Patients with AI-ECG scores above the median had lower 1-year OS (66.9%vs. 78.7%).
The AI-ECG score provided incremental prognostic value beyond HCT-CI, improving prediction of the composite cardiovascular outcome (AUROC, 0.62 to 0.68; likelihood-ratio P=0.003) and OS (C-index, 0.54 to 0.62; P<0.001). Addition of the score to a multivariable clinical model also significantly improved model fit for cardiovascular events (likelihood-ratio P=0.038) and OS (P<0.001).
Conclusions: To our knowledge, this is the first study demonstrating that a pre-trained AI-ECG model, originally developed to estimate future atrial fibrillation, can inform risk stratification before CAR T-cell therapy. The resulting AI-ECG score captures host susceptibility to cardiovascular complications and systemic inflammation while independently predicting treatment-related toxicities, non-relapse mortality, and overall survival. Because ECGs are routinely obtained before CAR T-cell infusion, AI-enhanced ECG analysis represents a scalable strategy for pre-treatment risk stratification and may help identify patients who could benefit from intensified monitoring and individualized supportive care.
DISCLOSURES
Brian C Shaffer, MD, MS: No relevant conflicts of interest to declare.