The clinical rationale rests on the well-documented timing of infectious complications after CAR-T therapy. Infection is concentrated in the first month following infusion, a period when lymphodepletion chemotherapy, prolonged cytopenias, and immune effector cell toxicities such as cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS) overlap. Prior research has linked baseline health, prior infection, and treatment burden to infection outcomes, while CRS, ICANS, corticosteroid exposure, and delayed immune recovery add predictive information during follow-up. The authors emphasize that the meaning of any such variable depends on when it becomes available relative to the infection itself, which is why the study separates a descriptive, all-history analysis from a strictly timed prognostic analysis anchored at the Day 14 landmark.

Methodologically, the adult lymphoma dataset supplied 3,332 numeric patient records, of which 30 lacked a usable binary infection endpoint, leaving 3,302 for analysis. The endpoint was the source indicator of clinically significant infection requiring treatment through Day 100, with relapse or progression, non-infection death, and subsequent hematopoietic cell transplantation treated as competing events in the source time-to-first-infection framework. The original all-history candidate set contained 30 clinical variables: 18 pre-infusion measures, including age, performance status, comorbidity index, prior infection, prior transplantation, CAR-T product, bridging therapy, and the interval from lymphodepletion initiation to infusion, and 12 follow-up measures describing toxicity onset, supportive treatment, and hypogammaglobulinemia. The analysis used a stratified 70:30 development-test partition, yielding 2,311 development patients with 514 infection events and 991 test patients with 221 events.

The frozen all-history model was a six-member weighted ensemble built around XGBoost components, with ensemble weights proportional to performance on the precision-recall curve and a fixed score cutoff of 0.31. On the reserved test set it achieved an area under the receiver operating characteristic curve (AUROC) of 0.724 (95 percent confidence interval, 0.685 to 0.765) and an area under the precision-recall curve (AUPRC) of 0.532 (95 percent CI, 0.467 to 0.603). At the chosen cutoff, sensitivity was 91.4 percent, specificity 22.7 percent, positive predictive value 25.3 percent, and negative predictive value 90.2 percent. The cutoff placed 194 of 991 test patients below the boundary, and 19 of the 221 recorded infections fell in that low-score group, meaning the rule missed roughly one infection in twelve while flagging more than four-fifths of all patients as above the threshold.

The descriptive analysis also quantified which clinical features traveled with recorded infection. Bacterial, viral, and fungal infections were recorded in 453 (13.7 percent), 335 (10.1 percent), and 88 patients (2.7 percent) respectively, with overlap permitted: 608 infected patients had one subtype, 113 had two, and 14 had all three. Among infected patients, 59.9 percent had corticosteroid exposure compared with 39.9 percent of uninfected patients, tocilizumab use was 65.2 percent versus 52.0 percent, intravenous immunoglobulin use 35.4 percent versus 24.5 percent, and hypogammaglobulinemia 67.6 percent versus 57.2 percent. Severe toxicity was also more frequent in the infection group, with grade 3 to 5 ICANS in 26.9 percent versus 14.5 percent and grade 3 to 5 CRS in 13.6 percent versus 6.3 percent. The authors caution that SHAP-based model explanations describe associations shaped by treatment indications rather than causal drug effects, since corticosteroid and tocilizumab use can reflect the severity of CRS or ICANS itself.

The Day 14 landmark analysis required a corrected cohort definition. Eligibility demanded a valid source endpoint time greater than 14 days, and the team reconstructed membership in the original landmark partition before removing ineligible patients rather than drawing a new split, preserving partition provenance. The corrected cohort contained 2,876 patients with 375 subsequent infections, split into 2,014 development patients with 262 events and 862 test patients with 113 events, a test infection rate of 13.1 percent. The baseline model used the 18 pre-infusion variables; the dynamic model added CRS and ICANS occurrence and onset information established by Day 14, reaching 22 variables. In the test set, the baseline model achieved an AUROC of 0.650 (95 percent CI, 0.590 to 0.701) and the dynamic model 0.669 (95 percent CI, 0.616 to 0.716), with AUPRCs of 0.192 and 0.217 respectively; the paired AUROC difference of +0.018 had a confidence interval crossing zero, from -0.008 to +0.047.

Both landmark thresholds were selected with a development sensitivity target of 0.90. At the dynamic model’s cutoff of 0.25, the low-score group contained 160 of 862 test patients, or 18.6 percent, and just 5 infections, an event rate of 3.1 percent (exact 95 percent CI, 1.0 to 7.1 percent) representing 4.4 percent of all test infections. The 702 patients above the cutoff included 108 infections, giving a positive predictive value of 15.4 percent. Score intervals stratified risk further: the intermediate group recorded 55 infections among 449 patients (12.2 percent) and the high group 53 among 253 (20.9 percent). The authors frame this low-score subgroup not as a ready-made clinical rule but as a defined population for prospective evaluation of differentiated outpatient monitoring, in which group size determines recruitment potential and the observed event rate informs monitoring requirements.

Calibration results temper any enthusiasm about treating raw scores as probabilities. The all-history scores exceeded observed infection rates in all eight revised quantile groups; the raw Brier score of 0.185 was worse than the 0.173 achieved by a constant prediction equal to the development event rate, though logistic recalibration fitted to cross-fitted development scores improved it to 0.144. For the dynamic Day 14 model, raw and recalibrated Brier scores were 0.198 and 0.110 against a reference of 0.114. Decision curve analysis, which compared model-guided action with monitoring everyone or no one under a hypothetical additional-outpatient-monitoring action, returned negative raw net benefit for the all-history model at threshold 0.31 (-0.066) and the Day 14 model at 0.25 (-0.104), exposing the error of applying uncalibrated scores directly. The authors stress that a clinical action threshold requires prospective estimation of missed-event costs and monitoring burdens.

To test transportability, the team applied a common-feature classifier developed entirely on the adult lymphoma cohort to 465 children, adolescents, and young adults with B-cell acute lymphoblastic leukemia treated with tisagenlecleucel, in whom 102 infections were recorded (21.9 percent). Using 22 variables shared across the two settings, the model achieved an adult out-of-fold AUROC of 0.714 and a cross-disease AUROC of 0.663 (95 percent CI, 0.604 to 0.724) with an AUPRC of 0.356. At its fixed cutoff of 0.33, sensitivity in the leukemia cohort was 56.9 percent, specificity 68.3 percent, and 44 of 102 infections fell below the cutoff. Source-trained recalibration changed the Brier score only marginally, from 0.1723 to 0.1722, while log loss increased from 0.5277 to 0.5456, so the adult-derived calibration map did not consistently improve cross-disease probability error. Differences in disease, age, product, and predictor availability can alter both ranking and the score-to-event relationship, the authors note.

Stability and sensitivity analyses rounded out the picture. Outer five-fold cross-validation of the full development procedure produced AUROCs ranging from 0.651 to 0.728 for the all-history rule, 0.562 to 0.670 for the baseline landmark rule, and 0.572 to 0.687 for the dynamic rule, quantifying instability in the development process itself. Removing two toxicity status fields whose ascertainment window was defined by infection timing barely changed all-history performance (AUROC 0.727, paired change +0.002), and adding Day 14 neutrophil-recovery status shifted the dynamic model by +0.005. Key limitations include missing dates for corticosteroid, tocilizumab, and immunoglobulin exposure, absent longitudinal lymphocyte and immunoglobulin measurements, unspecified antimicrobial prophylaxis and viral surveillance data, and infection fields that did not distinguish suspected from microbiologically confirmed cases or capture severity. The authors conclude that prospective adult lymphoma studies can now test whether the Day 14 low-score subgroup supports different outpatient monitoring schedules with acceptable event detection and patient burden, within established prevention and treatment protocols.

Subject of Research: Infection risk stratification after CD19 CAR-T therapy using all-history and Day 14 landmark prediction models

Article Title: Temporal clinical trajectories and Day+14 landmark stratification of infection risk after CD19 CAR T-cell therapy

Article References: Yang, L., Shan, L., Feng, S., Zhao, R., Wen, Y., Gu, X., Li, Z., Yao, X., & Wang, Y. (2026). Temporal clinical trajectories and Day+14 landmark stratification of infection risk after CD19 CAR T-cell therapy. Frontiers in Cellular and Infection Microbiology, 16, Article 1933555. https://doi.org/10.3389/fcimb.2026.1933555

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1933555

Keywords: CAR-T therapy, CD19, large B-cell lymphoma, infection risk, machine learning, landmark analysis, CIBMTR, cytokine release syndrome, ICANS, calibration, cross-disease validation, pediatric ALL