We conducted a single-center retrospective study of 302 patients with R/R MM treated with CAR-T therapy at the Affiliated Hospital of Xuzhou Medical University from June 2018 to October 2024 (Figure 1). The administered CAR-T regimens included: (1) monospecific BCMA-targeted CAR-T cells, (2) monospecific GPRC5D-targeted CAR-T cells, (3) BCMA/CD19 bispecific CAR-T cells, or (4) BCMA/GPRC5D bispecific CAR-T cells. The study was approved by the hospital's Ethics Committee and conducted in accordance with the Declaration of Helsinki. This retrospective analysis utilized data from patients at our center who were enrolled in prospective CAR-T clinical trials registered on Chictr.org.cn (ChiCTR2000033567, ChiCTR-OIC-17011272, ChiCTR1900026219) and ClinicalTrials.gov [NCT05509530; a phase 1 trial of anti-BCMA/GPRC5D bispecific CAR-T cells (8)]. All participants provided written informed consent for the original trials, which included authorization for subsequent data analysis. Patient inclusion and exclusion criteria were consistent with previously published studies (8, 21). Prior to CAR-T infusion, all patients underwent lymphodepletion with fludarabine (30 mg/m² per day on days −5 to −3) and cyclophosphamide (750 mg/m² on day −5).

Based on clinical experience, baseline demographic and clinicopathological variables were collected for patients with R/R MM prior to lymphodepletion. The variables considered for analysis included age, gender, myeloma subtype, prior treatment history, tumor burden, cytogenetic abnormalities, extramedullary disease status, body mass index (BMI) and the following laboratory values: absolute neutrophil count (ANC), hemoglobin (Hb), platelets (PLT), ferritin, C-reactive protein (CRP), interleukin-6 (IL-6), interleukin-8 (IL-8), interleukin-10 (IL-10), interferon-α (IFN-α), interferon-γ (IFN-γ), and lactate dehydrogenase (LDH). The CONUT score was calculated from serum albumin, serum cholesterol, and total peripheral lymphocyte count. Patients were stratified into one of four CONUT score groups, as defined in Table S1.

Cytopenia was graded according to the Common Terminology Criteria for Adverse Events (CTCAE) version 5.0 (22). PHT was defined as grade 3 or 4 cytopenia persisting beyond day 28 (D28+). We defined two specific endpoints: PHT_ANC (grade 3/4 neutropenia lasting > 28 days) and a PHT_Composite (grade 3/4 neutropenia, anemia, or thrombocytopenia lasting > 28 days). CAR-HEMATOTOX score was calculated prior to lymphodepletion using the online CAR-HEMATOTOX calculator from the German Lymphoma Alliance (GLA) (14).

All data were retrospectively extracted from the electronic medical record system of the Affiliated Hospital of Xuzhou Medical University between June 2018 and October 2024. For each included patient, the collected dataset comprised baseline demographic, disease history, pre-lymphodepletion laboratory results, CAR-T product details, and post-infusion toxicity records. All baseline variables were obtained prior to lymphodepletion and were available for the entire cohort. To ensure accuracy, data extraction was performed independently by two investigators, with any discrepancies resolved through consensus.

Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR), depending on their normality distribution, and were compared using independent samples t-tests or non-parametric tests, respectively. Categorical variables are expressed as frequencies and percentages (n, %) and were compared using the Chi-square test. In this retrospective cohort, patients were randomly split into a training cohort and an internal validation cohort (7:3). Within the training cohort, univariate logistic regression was first used to identify baseline variables and CONUT scores associated with PHT_ANC. Variables significant at the p < 0.05 level were subsequently entered into a stepwise multivariate logistic regression model. The final multivariate model served as the basis for constructing a predictive nomogram for PHT_ANC.

The discriminatory capability of the nomogram for predicting PHT_ANC after CAR-T therapy was evaluated employing bootstrap validation. This process involved 200 repetitions of random sampling with replacement from the training set, with the mean area under the receiver operating characteristic curve (AUC) and its 95% confidence interval (CI) subsequently derived and compared to those of other predictors. The agreement between predictions and observations (calibration) was analyzed using a calibration curve constructed from 1,000 bootstrap resamples. Decision curve analysis (DCA) was utilized to quantify the clinical utility by estimating the net benefit across different threshold probabilities. To further assess the model's stability, the nomogram was deployed on the validation cohort, and its effectiveness was verified through the AUC, calibration curve, and DCA. The entire statistical analysis was performed using R software (version 4.3.2): R Foundation for Statistical Computing, Vienna, Austria.

A total of 302 patients with R/R MM were included in this study and randomly assigned to a training cohort (n = 211, 69.9%) and a validation cohort (n = 91, 30.1%) at a 7:3 ratio. In the training cohort, the median age was 57 years (range: 34–73), 113 patients (53.6%) were male, and 123 patients (58.3%) were classified as R-ISS stage III. High tumor burden was present in 26.5% (n = 56), high-risk cytogenetic features in 25.6% (n = 54), and extramedullary lesions in 36.5% (n = 77). The median number of previous therapy lines was 4 (range: 2–14), and a history of HCT was reported in 63 patients (29.9%). The IgG M-protein subtype was the most frequent (64.5%, n = 136). Regarding CAR-T cell therapy, 49 patients (23.2%) received BCMA, 54 patients (25.6%) received GPRC5D, 87 patients (41.2%) received BCMA+CD19, and 21 patients (9.9%) received BCMA+GPRC5D. The incidence of PHT_ANC after CAR-T therapy was 45.9% (n = 97) in this cohort. This incidence is consistent with recent clinical studies reporting PHT_ANC rates (40%−50%) in R/R MM patients after CAR-T therapy (13, 23), verifying the representativeness of the study sample.

No statistically significant differences were observed between the two cohorts across all variables (all p > 0.05), indicating that the random allocation effectively balanced confounding factors. Detailed baseline clinical characteristics of both the training and validation cohorts are summarized in Table 1.

To identify factors associated with PHT_ANC in patients with R/R MM after CAR-T therapy, we first performed univariate logistic regression analysis, with PHT_ANC treated as a binary endpoint. Results showed that seven variables significantly correlated with PHT_ANC (all p < 0.05): high tumor burden (p = 0.013), platelet count (p = 0.032), serum ferritin (p = 0.002), CRP (p = 0.025), IL-6 (p = 0.015), IFN-γ (p = 0.018), and CONUT (p = 0.011). These findings offered initial evidence supporting the potential utility of these clinical and laboratory indicators for predicting PHT_ANC in this patient population.

Notably, univariate analysis only captures the unadjusted association between individual variables and PHT_ANC, as it fails to account for confounding among variables. To address this limitation and ensure methodological consistency, we included all statistically significant variables from the univariate analysis in a multivariate logistic regression model. After adjusting for confounding effects, the multivariate analysis confirmed four variables as independent risk factors for PHT_ANC in R/R MM patients after CAR-T therapy: high tumor burden (p = 0.017), serum ferritin (p = 0.024), IFN-γ (p = 0.028), and CONUT score (p = 0.012; Table 2). This result underscored that these four factors exert a robust, independent effect on PHT_ANC, even after controlling for the influence of other related variables.

Subsequently, we developed a predictive nomogram to quantify the risk of PHT_ANC in R/R MM patients after CAR-T therapy, using the four independent risk factors. The nomogram was constructed based on the regression coefficients (β-values) of the multivariate logistic regression model: each independent risk factor was assigned a weighted score proportional to its β-value (higher β-values indicate greater predictive importance), and these scores were visually displayed on a dedicated “score axis” for intuitive interpretation (Figure 2).

In clinical practice, this nomogram enables personalized PHT_ANC risk assessment through three straightforward steps: (1) Map a patient's baseline values (e.g., serum ferritin level, CONUT score) to the corresponding score on each factor's scale; (2) Sum these individual scores to calculate a “total risk score”; (3) Convert the total risk score to a predicted probability of PHT_ANC via the nomogram's “risk axis.”

Eliminating the need to interpret complex raw regression outputs, this visual, user-friendly design makes PHT_ANC risk prediction more actionable in routine clinical practice.

In the training cohort, the bias-corrected area under the curve (AUC) derived from Bootstrap validation was 0.815 (95% CI: 0.809–0.822), which exceeded the AUC values of other indicators: CAR-HEMATOTOX (AUC: 0.706, 95% CI: 0.702–0.713), high tumor burden (AUC: 0.617, 95% CI: 0.608–0.619), ferritin (AUC: 0.672, 95% CI: 0.669–0.677), IFN-γ (AUC: 0.641, 95% CI: 0.638–0.642), and CONUT score (AUC: 0.692, 95% CI: 0.688–0.701). All pairwise comparisons yielded a p-value < 0.001 (Table 3). Additionally, we utilized the established nomogram to predict PHT_Composite in patients with R/R MM after CAR-T therapy. Application of the nomogram to this patient cohort yielded an AUC of 0.821 (95% CI: 0.811–0.826) for PHT_Composite prediction, with a non-significant p-value of 0.417, validating its excellent discriminative performance. Notably, the AUC remained stable in the validation cohort, reaching 0.824 (95% CI: 0.813–0.831).

In both the training and validation cohorts, the (actual and calibrated) calibration curves closely overlapped with the ideal line, indicating that the nomogram's predicted outcomes were in good agreement with the observed real-world results (Figures 3A, B). DCA results depicted in Figures 3C, D demonstrate that within a relatively wide range of clinical threshold probabilities, the application of this nomogram in clinical practice yields net clinical benefits.