Machine learning–based identification of patients at risk for early infectious complications after CAR-T in multiple myeloma.

Abstract

e19518 Background: Infections occur in 50–70% of Multiple Myeloma(MM) patients after CAR T therapy, with the highest risk in the first few months during the period of profound and recurrent cytopenias. Although the incidence and timing of infections have been well characterized, prospective identification of patients at highest risk remains a challenge. Existing risk models have focused on adverse events across heterogeneous B-cell malignancies, and no MM-specific, infection-focused tool exists for prospective risk stratification. We developed a MM-specific, Day 0, clinically based machine-learning model to enrich for early (≤60 day) severe infection risk after anti BCMA CAR T. Methods: We conducted a retrospective analysis of 51 MM patients treated with CAR T (both ide-cel and cilta-cel) at our institution between September 2024 - May 2025. The primary assessment was grade ≥3 infection within 60 days of infusion. Baseline clinical variables used to develop a Day 0 infection risk model were age, ECOG status, comorbidity burden, and high-risk myeloma features. An elastic-net logistic regression approach was employed to estimate the probability of early severe infection. This modeling framework allows multiple correlated clinical variables to contribute to risk estimation while applying regularization to reduce overfitting in modestly sized datasets. Model performance was evaluated with an emphasis on clinical risk enrichment rather than individual level prediction, including assessment of discrimination, calibration, and observed infection rates across strata of predicted risk. Results: Fourteen patients (27%) developed grade ≥3 infections within 60 days, with most occurring within the first 30 days. Greater than 50% required readmissions post initial CAR T discharge. Among grade ≥3 infections, 65% were viral and 36% bacterial, challenging the assumption that early events are predominantly bacterial. The model demonstrated modest discrimination (apparent AUC 0.67) with meaningful risk stratification. When patients were grouped into tertiles of predicted risk, observed 60-day infection rates increased stepwise from 18% in the low-risk group to 30% in the intermediate-risk and 35% in the high-risk group. Calibration was generally reasonable, with underestimation of risk among the highest-risk patients. Conclusions: Using readily available Day 0 clinical variables, the study demonstrates proof of concept that early severe infection risk after CAR T therapy can be stratified in real world practice. While not intended for precise individual prediction, it provides a clinically actionable risk enrichment tool to inform post infusion management. The model may help identify patients who could benefit from intensified early surveillance or a lower threshold for admission and may support risk-based design of future infection prevention trials after CAR T therapy.