Prediction models for CAR-T Cell therapy-related adverse events in hematologic malignancies: a systematic review and meta-analysis

Abstract

Background

Identifying patients at high risk of adverse events is crucial in chimeric antigen receptor T-cell (CAR-T) therapy to enable early intervention. Despite the development of numerous prediction models, their methodological quality and performance remain systematically unassessed.

Methods

We conducted a systematic search across seven databases and Google Scholar, covering all records up to October 29, 2025. The risk of bias and applicability were assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST). Only the Area Under the Receiver Operating Characteristic Curve (AUC) results from external validation cohorts were included in the meta-analysis. Subgroup and sensitivity analyses were conducted to explore potential sources of heterogeneity.

Results

Twenty-nine studies were included, reporting 26 distinct prediction models. The most common predictors were platelet count, C-reactive protein, and interleukin-6. Ten models underwent external validation, and six reported calibration. The PROBAST assessment indicated a high risk of bias across all studies. Pooled AUCs from external validation cohorts ranged from 0.60 to 0.79.

Conclusion

Pooled AUC estimates indicate moderate discrimination of the included models. However, these estimates reflect discriminative ability under uncertainty rather than direct clinical readiness.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-026-15877-8.

Introduction

Chimeric antigen receptor T-cell (CAR-T) therapy is an immunotherapy that genetically modifies T cells to express tumor-specific chimeric antigen receptors, enabling precise targeting of cancer cells [1]. It is associated with high disease remission rates in relapsed or refractory B-cell leukemia, lymphoma, and multiple myeloma (MM), making it a promising or breakthrough treatment for hematologic malignancies [2].

While CAR-T cell therapy demonstrates significant efficacy in treating hematologic malignancies, its therapeutic outcomes are often limited by adverse events. It is estimated that 60% to 90% of patients with hematologic malignancies experience adverse events following CAR-T cell therapy [3, 4]. The most frequently predicted adverse events included cytokine release syndrome (CRS), immune effector cell-associated neurotoxicity syndrome (ICANS), and immune effector cell-associated hematotoxicity (ICAHT), infections, cardiovascular adverse events, hemorrhage, and immune effector cell-associated hemophagocytic lymphohistiocytosis-like syndrome (IEC-HS). CRS, the most common toxicity, results from immune activation and cytokine surge, typically manifesting as high-grade fever and potentially progressing to hypotension, hypoxemia, and multi-organ failure [5]. Similarly, severe ICANS may progress to cerebral edema, transient coma, or seizures [6, 7], also posing a risk to patient survival.

Prediction models play a significant role in forecasting and managing adverse events in CAR-T cell therapy. While multiple models have been developed recently, most exhibit limited clinical generalizability owing to small sample sizes and lack of external validation [8]. Existing reviews have primarily focused on the incidence and clinical characteristics of CAR-T cell therapy-related adverse events, with no systematic evaluation of the quality and performance of published models [9, 10]. Therefore, this study aims to systematically synthesize prediction models for CAR-T cell therapy-related adverse events in hematologic malignancies, evaluate their methodological quality, and assess reported discrimination through meta-analysis to identify candidates suitable for clinical application.

Methods

The review protocol was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO, registration number: CRD420251062713).

Literature search

We systematically searched PubMed, the Cochrane Library, Embase, Web of Science, Scopus, China National Knowledge Infrastructure (CNKI), China Biology Medicine disc (CBM), and Google Scholar for prediction models for CAR-T cell therapy-related adverse events. Additionally, we screened the reference lists of included studies to identify any other relevant literature that might have been overlooked in the search strategy. The search period ranged from the inception of each database to October 29, 2025. The following keywords were used for retrieval: “Hematologic neoplasm”, “Hematologic malignancy”, “non-Hodgkin lymphoma”, “multiple myeloma”, “acute lymphoblastic leukemia”, “Receptors, Chimeric Antigen”, “CAR-T”, “Tisagenlecleucel”, “Axicabtagene ciloleucel”, “Lisocabtagene maraleucel”, “Brexucabtagene autoleucel”, “Idecabtagene vicleucel”, “Ciltacabtagene autoleucel”, “Drug-Related Side Effects and Adverse Reactions”, “Cytokine Release Syndrome”, “immune effector cell-associated neurotoxicity syndrome”, “immune effector cell-associated hematotoxicity”, “Infection”, “cardiovascular toxicities”, “Tumor Lysis Syndrome”, “assessment”, “machine learning”, “deep learning”, “risk”, “predict”, “nomogram”, “model”, “score”. (The complete search strategy is provided in the Supplementary Table S2) The literature search was conducted independently by two investigators. The results from all databases were merged and deduplicated using reference management software to determine the total number of records requiring further screening.

This review is based on the Prediction model Risk of Bias Assessment Tool (PROBAST) framework, as outlined in the Transparent Reporting of multivariable prediction models for Individual Prognosis Or Diagnosis: checklist for Systematic Reviews and Meta-Analyses (TRIPOD-SRMA) checklist [11], and follows methodological guidance for systematic reviews of prediction models [12]. The PICOTS elements are defined as follows:

P (Population): Patients with hematologic malignancies receiving CAR-T cell therapy, including non-Hodgkin lymphoma (NHL), acute lymphoblastic leukemia (ALL), and MM.

I (Prediction Model): Any prediction model for adverse events related to CAR-T cell therapy, comprising at least two predictors.

C (Comparator): Not applicable.

O (Outcome): CAR-T cell therapy-related adverse events, including CRS, ICANS, ICAHT, IEC-HS, infections, hemorrhage, and cardiovascular adverse events, among others.

T (Timing): Any time point before or after CAR-T cell infusion, including long-term follow-up.

S (Setting): Outpatient or inpatient settings (including transplant units, hematology/oncology wards). No restrictions on CAR-T cell targets (e.g., B-cell maturation antigen (BCMA), Cluster of differentiation 19 (CD19), CD22).

Inclusion and exclusion criteria

Inclusion criteria were as follows: (1) Study population comprised patients with hematologic malignancies receiving CAR-T cell therapy, including NHL, ALL, MM, with no age restrictions; (2) CAR-T cell targets were specified, including approved or investigational targets such as CD19, CD22, or BCMA; (3) Clinical settings included outpatient clinics, hematopoietic stem cell transplantation units, hematology or oncology wards; (4) Both retrospective and prospective study designs were eligible; (5) Studies included at least one prediction model for CAR-T cell therapy-related adverse events, involving model development, validation, or updating of existing models; (6) Models contained two or more predictors; (7) The predicted outcomes included CAR-T cell therapy-related adverse events such as CRS, ICANS, ICAHT, IEC-HS, infections, bleeding, and cardiovascular events. These outcomes were defined according to standardized criteria: CRS and ICANS based on the American Society for Transplantation and Cellular Therapy (ASTCT) 2019 consensus [13]; ICAHT following the consensus by Rejeski et al. [14]; and all other adverse events graded using the Common Terminology Criteria for Adverse Events, version 5.0 (CTCAE v5.0). Exclusion criteria were: (1) Publications not written in English or Chinese; (2) Unavailable full texts, or publications in the form of abstracts, letters, protocols, reviews, or case reports; (3) Studies with prediction outcomes focused on prognostic endpoints such as overall survival (OS), progression-free survival (PFS), or response rate (RR); (4) Studies developing prediction models based primarily on electrophysiological or radiological findings as predictors.

Study selection

The study selection process was conducted independently by two authors (LLQ and YLR). First, duplicate records were removed using EndNote (X9.3.3 Clarivate; Philadelphia, Pennsylvania, USA), and then the remaining studies were screened based on their titles and abstracts. Full-text articles were retrieved for studies that appeared eligible based on title and abstract screening, and these were then assessed against the full inclusion and exclusion criteria. In addition, the reference lists of included studies were manually searched to identify any additional relevant studies. Any discrepancies were resolved through discussion and consensus with a third investigator (XXH).

Data extraction

Data was extracted using a standardized data extraction form based on the TRIPOD-SRMA checklist. The extracted data included the following items: author, country, data source, sample size, model type, predictors, modeling approach, handling of missing data, model validation, model performance, presentation format, and study results. For studies reporting multiple models, data was extracted for each model. Data extraction was performed independently by two trained reviewers (LLQ and YLR) using a standardized data extraction form, with cross-checking of the extracted data. Any discrepancies were resolved by consensus, with input from a third investigator (XXH).

Bias risk assessment and critical appraisal

Study quality was assessed using the PROBAST [15] to evaluate the risk of bias and applicability at the model level. The PROBAST comprises 20 signaling questions across four domains: participants, predictors, outcomes, and analysis. Responses to each question can be rated as “yes”, “probably yes”, “probably no”, “no”, or “no information”. The overall risk of bias is categorized as low, high, or unclear. A domain is considered to have a high risk of bias if any signaling question receives a “no” or “probably no” response. The overall risk of bias is rated as low only if all four domains are assessed as low risk. The PROBAST assessment was conducted independently by two investigators (ZXJ and WLL), who evaluated the risk of bias and applicability of each model in the included studies to our review question. Any uncertainties encountered during the process were resolved through consultation with a third author (XXH).

Statistical analysis

We summarized the characteristics of all included models, presenting categorical variables as counts and percentages. To avoid heterogeneity arising from mixing different outcomes and model structures, we performed meta-analyses separately for each “Outcome–Model” combination, and a quantitative synthesis was conducted only when a combination had at least three external validation cohorts (k ≥ 3). To be included in the meta-analysis, studies were required to report the Concordance Statistic (C-statistic) / Area Under the Receiver Operating Characteristic Curve (AUC) along with their 95% confidence interval (95% CI) or standard error. If these measures of precision were not reported, they were estimated using the sample size and number of events, following the method described by Debray et al. [16]. Heterogeneity across studies was quantified using the I² and τ² statistics. Exploratory subgroup analyses were performed based on various study characteristics, such as disease type, event rate, and the timing of predictor measurement, to identify potential sources of heterogeneity and assess the robustness of the findings. Additionally, a leave-one-out sensitivity analysis was conducted to examine the stability of the pooled results. All statistical analyses were performed using R software, version 4.4.1 (R Core Team, Vienna, Austria), with the metafor and metamisc packages utilized for meta-analysis and the calculation of prediction model performance. A two-sided P value of less than 0.05 was considered statistically significant.

Results

Screening process

Figure 1 shows the study selection process. The initial database searches identified 3,730 records from PubMed (n = 1,343), Web of Science (n = 652), the Cochrane Library (n = 68), Embase (n = 1,202), Scopus (n = 461), CNKI (n = 0), and CBM (n = 4). After removing duplicates, 3,034 records were screened based on their titles and abstracts. Of these, 2,827 did not meet the inclusion criteria, and 5 had unavailable full texts, and were therefore excluded. Full-text review of the remaining 202 articles yielded 27 studies that met the eligibility criteria. An initial search of Google Scholar and the reference lists of included studies identified 26 potentially relevant articles. After screening, two of these articles from the reference lists met the inclusion criteria. Ultimately, 20 studies were identified for model development and 9 for external validation.

Fig. 1.

Literature screening flow diagram

Characteristics of the included studies

Table 1 lists the main characteristics of the included studies. Of the 29 studies, 20 involved model development and 9 involved external validation of prediction models. A total of 26 prediction models were reported. The age range across all included studies spanned from 0 to 89 years. CD19 was the most frequent CAR-T cell therapy target among the included studies (66%), followed by BCMA (10%). The remaining studies employed multi-target or combination approaches involving CD19, CD22, CD20, and BCMA. Among the included studies, ALL was involved in 12 studies (41.4%), predominantly B-cell ALL (B-ALL) (Supplementary Table S4). Regarding validation status, 12 models (47%) underwent internal validation only, 4 (15%) underwent external validation only [14, 17–23], 6 (23%) underwent both [24–32], and 4 (15%) had no validation [8, 33–35]. Classified by predicted outcome, there were 12 models for CRS, 10 for ICANS, 10 for ICAHT, 2 for infection [22, 33], and one each for serious cardiovascular adverse events [36], disseminated intravascular coagulation (DIC) [31], and the aplastic phenotype [32]. The majority of models (n = 16, 62%) were derived from retrospective studies, 7 (27%) from prospective studies [27, 37–41], and 3 (11%) from studies with a mixed retrospective-prospective design [28, 42]. Thirteen models (50%) originated from the United States, 8 (30%) from China, 2 (8%) from France, with the remainder developed in South Korea, Japan, and Germany (Table 1).

Table 1.

Characteristics of reviewed prediction models

Model characteristic	n (%) Model (n = 26)
Outcome*
CRS	12
ICANS	10
ICAHT	10
Infection	2
DIC	1
Serious cardiovascular adverse events	1
Aplastic phenotype	1
Location
United States	13(50)
China	8(30)
France	2(8)
South Korea	1(4)
Japan	1(4)
Germany	1(4)
Study type
Developed and internally validated study	12(47)
Developed and externally validated study	4(15)
Developed, internally, and externally validated study	6(23)
Developed study without validation	4(15)
Study design
Retrospective study	16(62)
Prospective study	7(27)
Retrospective and prospective study	3(11)
Handling of missing data
Multiple imputation	4(15)
Complete case analysis	1(4)
Exclusion	6(23)
Not reported	15(58)
Modelling method
Logistic regression model	15(57)
Machine learning	7(27)
Fine and Gray regression analysis	2(8)
Cox regression	1(4)
None	1(4)
Validation
Bootstrapping	12(46)
Cross validation	5(19)
Random split	1(4)
None	8(31)
Model presentation
Sum score	12(46)
Nomogram	4(15)
Website	3(12)
Decision tree	2(8)
None	5(19)

The data presented reflect only the characteristics of model development studies; * The number of outcomes (n≠26) is due to the inclusion of models predicting two or more adverse events; CRS, Cytokine release syndrome; DIC, Disseminated intravascular coagulation; ICAHT, Immune effector cell-associated hematotoxicity; ICANS, Immune effector cell-associated neurotoxicity syndrome

Predictors in prediction models

A total of 80 predictors were employed in model development or validation. These predictors were categorized into five groups: baseline patient characteristics, tumor burden, disease status, laboratory parameters (including inflammatory/immune markers and hematologic/biochemical tests), and toxicity grading. Among models predicting CRS, ICANS, and hematologic adverse events, respectively, platelet count (PLT) was the most frequently utilized predictor, incorporated in 8 models for each outcome. This represented 7.9%, 22%, and 19.5% of all unique predictors used in models for each respective outcome (Fig. 2). It should be noted that due to the large total number of distinct predictors (n = 63) involved in CRS prediction models, the bar chart for CRS in Fig. 2 displays only those factors included in two or more models. The complete list of all predictors is available in the Supplementary Table S4.

Fig. 2.

Frequency of predictors included in the final models by outcome. ANC, Absolute neutrophil count; CRP, C-reactive protein; CRS, Cytokine release syndrome; Hb, Hemoglobin; ICANS, Immune effector cell-associated neurotoxicity syndrome; ICAHT, Immune effector cell-associated hematotoxicity; IFN-γ, Interferon-gamma; IL-2, Interleukin-2; IL-4, Interleukin-4; IL-6, Interleukin-6; IL-10, Interleukin-10; IL-17, Interleukin-17; LDH, Lactate dehydrogenase; PCT, Procalcitonin; PLT, Platelet count; PT, Prothrombin time; TNF-α, Tumor necrosis factor-alpha; WBC, White blood cell count

Modeling methods and missing data

The predominant modeling method was logistic regression (n = 15, 57%). Seven models (27%) were developed using machine learning [24, 25, 27, 36, 37, 40, 43], 2 models (8%) used Fine and Gray regression [39], 1 model (4%) used Cox regression [33], and 1 model (4%) lacked documentation on the development data [29]. Among the 11 models that reported missing data in the development dataset, 6 models (23%) excluded missing values directly [17, 18, 20, 33, 40, 41], 4 models (15%) used multiple imputation [24, 29, 42], and 1 model (4%) had no missing values [23]. The remaining models (n = 15, 58%) provided no missing data information.

Model presentation

The model presentation formats were as follows: 12 models (46%) used total scoring systems, 4 models (15%) used nomograms [18, 24, 26, 38], 3 models (12%) used web-based calculators [21, 25, 42], 2 models (8%) used decision trees [27, 37], and the remaining 5 models (19%) did not describe their presentation format.

Validation and performance

For internal validation, bootstrap resampling (12 models, 46%), cross-validation (5 models, 19%) [25, 27, 37, 40, 43], and holdout validation (1 model, 4%) were used [36], while 8 models (31%) did not undergo internal validation [17–20, 25, 33–35]. For model performance, discrimination was reported in 19 models (73%). The reported AUC ranged from 0.53 to 0.98. Only 6 models (23%) reported calibration [18, 20, 24, 26, 42]. Across models, sensitivity ranged from 43% to 98% and specificity from 39% to 99% (Supplementary Table S5).

Risk of bias and applicability assessment

The 26 included prediction models were evaluated for risk of bias and applicability using the PROBAST (Fig. 3). The assessment revealed that all models were rated as having a high overall risk of bias. In the domain of participants, 65% of the models had a high risk of bias, primarily because they were derived from retrospective cohorts. For the predictors domain, although no model was rated high risk, 54% were rated as having an unclear risk of bias due to unreported blinding of outcome data during predictor assessment. In the outcomes domain, all models were rated as low risk, as the definition and assessment of outcomes were sufficiently standardized. In the analysis domain, all models were rated as high risk of bias. The primary reasons were as follows: 88% had an events per variable (EPV) of less than 20; 69% did not avoid univariable pre-screening of predictors; 58% failed to specify the method for handling missing data; 31% did not undergo internal validation; and 77% did not report any measure of calibration (Supplementary Tables S4–S5). Regarding applicability, all models were judged to have low concerns.

Fig. 3.

Risk of bias across all included studies

Meta-analysis of the included validation models

A total of three prediction models—the modified Endothelial Activation and Stress Index (m-EASIX), EASIX, and the CAR-HEMATOTOX score—met the criterion of having at least three external validation cohorts for their respective outcomes and were therefore included in the quantitative synthesis. The pooled AUCs for these models are shown in Table 2, with the corresponding forest plots, subgroup analyses, and sensitivity analyses presented in Supplementary Figures S1–S7.

Table 2.

Results of meta-analyses of AUCs for prediction models

Outcome	Model	Number of datasets	Summary AUCs (95% CI)	Prediction intervals	I2 (%)	τ2
CRS	m-EASIX	6	0.69 (0.57–0.78)	0.41–0.87	44.7	0.1055
	EASIX	5	0.65 (0.57–0.72)	0.42–0.82	98.2	0.0700
ICAHT	CAR-HEMATOTOX	4	0.79 (0.61–0.90)	0.31–0.97	67.2	0.1433
	EASIX	4	0.72 (0.67–0.76)	0.65–0.77	0	0
ICANS	m-EASIX	3	0.61 (0.34–0.83)	0.0065–1.00	75.3	0.1632
	EASIX	3	0.60 (0.47–0.72)	0.23–0.88	4.8	0.0001

95% CI, 95% confidence interval; AUC, Area Under the Receiver Operating Characteristic Curve; CRS, Cytokine release syndrome; EASIX, Endothelial Activation and Stress Index; ICAHT, Immune effector cell–associated hematotoxicity; ICANS, Immune effector cell–associated neurotoxicity syndrome; m-EASIX, modified Endothelial Activation and Stress Index

For the outcome of CRS, two models, m-EASIX and EASIX, were compared. The m-EASIX showed a pooled AUC of 0.69 (95% CI: 0.57–0.78) with a prediction interval of 0.41–0.87 and moderate between-study heterogeneity (I² = 44.7%). The EASIX yielded a pooled AUC of 0.65 (95% CI: 0.57–0.72) with a prediction interval of 0.42–0.82, despite high heterogeneity (I²=98.2%). Further subgroup analysis (Supplementary Figures S1–S3) revealed that the discrimination of EASIX was superior in B-cell malignancies (pooled AUC: 0.70; 95% CI: 0.68–0.72) compared to plasma cell malignancies. When stratified by the timing of prediction, the pooled AUC was 0.61 (95% CI: 0.54–0.68) for pre-infusion and 0.66 (95% CI: 0.58–0.73) for post-infusion, with no statistically significant difference between them (χ²=0.75, p = 0.388). A leave-one-out sensitivity analysis demonstrated that the pooled effect size remained stable after sequentially removing each study, supporting the robustness of this finding.

For the ICAHT outcome, two models, the CAR-HEMATOTOX score and EASIX, were included in the meta-analysis. The CAR-HEMATOTOX score demonstrated a pooled AUC of 0.79 (95% CI: 0.61–0.90) with a prediction interval of 0.31–0.97, showing moderate heterogeneity (I²=67.2%). The EASIX exhibited a pooled AUC of 0.72 (95% CI: 0.67–0.76) with a prediction interval of 0.65–0.77 and low heterogeneity (I² = 0%). To explore the source of heterogeneity for the CAR-HEMATOTOX score, an exploratory subgroup analysis based on event rate was performed (Supplementary Figures S4–S6). The AUCs were 0.80 (95% CI: 0.73–0.86) in cohorts with high event rates and 0.81 (95% CI: 0.51–0.95) in those with low event rates, with no statistically significant difference between the subgroups (χ²=0.00, p = 0.996). The robustness of the pooled results was further confirmed by sensitivity analysis.

For the outcome of ICANS, both the m-EASIX and EASIX demonstrated limited discriminative ability. The m-EASIX had a pooled AUC of 0.61 (95% CI: 0.34–0.83) with a prediction interval of 0.0065–1.00, indicating significant heterogeneity (I²=75.3%). The EASIX showed a pooled AUC of 0.60 (95% CI: 0.47–0.72) with a prediction interval of 0.23–0.88 and minimal heterogeneity (I²=4.8%).

Discussion

Main findings and interpretation

Although CAR-T cell therapy has achieved substantial efficacy in hematologic malignancies, its broader clinical use remains limited by the high incidence and potential lethality of treatment-related adverse events [44]. Therefore, reliable prediction models are urgently required to enable early risk stratification and support clinical decision-making. This systematic review identified 26 prediction models reported across 29 studies. In the meta-analysis, m-EASIX and EASIX exhibited moderate discrimination for CRS but limited performance for ICANS, while the CAR-HEMATOTOX score demonstrated relatively better discrimination for ICAHT. However, PROBAST assessment indicated that all included studies had a high risk of bias, which substantially limits the reliability and clinical applicability of the available models.

The high risk of bias across the included models was mainly driven by shortcomings in the PROBAST “Participants” and “Analysis” domains, highlighting methodological limitations in existing studies. While retrospective design is common in this field, the risk of bias is primarily driven by sampling strategies and inclusion mechanisms rather than the study design alone. Furthermore, owing to the relatively recent clinical adoption of CAR-T cell therapy and the limited number of eligible patients, many included studies had modest sample sizes. In this review, the median sample size was 155, and the smallest study included 24 participants. EPV is defined as the ratio of the number of outcome events to the total number of candidate predictors considered at any stage of model development [45]. Among the 26 identified prediction models, only three met the conventional threshold of EPV ≥ 20, which may increase the risk of overfitting and compromise the reliability of model performance estimation [25]. Moreover, methodological studies have suggested that the applicability of traditional EPV criteria may be limited for prediction models developed using machine learning, as such models typically require a larger number of events during development to achieve stable and robust performance [15, 46]. Missing data handling posed another major methodological concern, as more than half of the studies did not report how missing values were addressed. Inappropriately excluding missing data may compromise model accuracy [47], and inadequate handling of missing values for key predictors may further weaken their contribution to the model [24, 43]. The reporting of model performance was incomplete. Only 23% of studies presented calibration, while the majority relied solely on AUCs. It is crucial to note that these models would require recalibration or updating before real-world clinical implementation to ensure reliable performance. Predictor selection methods were frequently suboptimal; more than half of the studies used univariable screening, which may enhance interpretability in small samples but overlooks potential interactions and confounding effects. Shrinkage techniques, which can effectively reduce overfitting and improve generalizability, were applied in only a minority of models [48]. Therefore, future studies should more rigorously adhere to the Critical Appraisal and Reporting of a Multivariable Prediction Model Study (CHARMS) recommendations, adopt prospective designs, ensure adequate sample sizes, appropriately address missing data, and implement more robust strategies for predictor selection and overfitting control [49–51].

With regard to the clinical usability of prediction models, we observed that an increasing number of studies have begun to emphasize how these models are presented in clinical practice. For example, several models have been accompanied by user-friendly online calculation platforms, such as the early ICAHT prediction model (eIPM) web-based tool developed by Liang et al. [42] and the prediction model of severe CRS (PrCRS) platform proposed by Wei et al. [25]. These tools allow clinicians to obtain individualized risk estimates by entering routinely collected clinical data. At the level of predictor selection, frequently reported variables—including PLT, C-reactive protein (CRP), and various cytokines reflecting inflammatory and hematologic activity—provide important references for model development and clinical risk stratification [39, 52]. However, it should be noted that despite their strong predictive potential, some biomarkers, particularly cytokines, may have limited applicability in routine clinical practice due to high testing costs and the fact that many assays are not yet fully incorporated into standard diagnostic workflows [30]. This limitation is especially pronounced in resource-limited healthcare settings. In contrast, prediction models based on patient-reported outcomes (PROs), such as those proposed by Zhao et al. [38], offer advantages in terms of accessibility. Therefore, future development and optimization of prediction models should carefully consider both the availability of predictors and the usability of model formats, in order to facilitate the practical implementation of high-quality prediction tools in clinical practice.

Among the models included in this review, more than half were constructed using logistic regression, whereas only 27% employed machine learning. Logistic regression remains a commonly used method in clinical prediction modeling because of its interpretability; however, it has well-recognized limitations in capturing nonlinear relationships [53, 54] and complex interactions among multiple predictors [55]. In addition, traditional statistical models typically rely on data collected at a single fixed time point (e.g., before lymphodepletion), making it difficult to account for the dynamic changes in biomarkers that occur throughout the course of CAR-T cell therapy. In contrast, machine learning have shown considerable potential for handling complex and dynamic data patterns. For example, gradient boosting machines (GBM) and random forests (RF) have demonstrated superior performance in specific datasets and tasks [56, 57]. GBM is an ensemble algorithm that iteratively minimizes prediction error to improve model accuracy. Su et al. [27] applied GBM to three clinically relevant time windows—pre-treatment, fever onset, and CRS peak—to identify stage-specific key predictors and enhance model performance. The advantages of machine learning are further highlighted in several practical challenges. For large datasets or high-dimensional features, Jung et al. [36] utilized an Extreme Gradient Boosting (XGBoost)-based GBM model to efficiently process 3,280 reports. When the number of potential predictors is extensive, the BorutaShap algorithm—a feature-selection method built on XGBoost—can jointly evaluate all variables and substantially reduce the risk of false-positive findings associated with traditional multiple-comparison procedures. Collectively, these studies suggest that, in specific contexts, machine learning has the potential to complement and even improve upon the limitations of traditional modeling approaches.

Among the available prediction models, the CAR-HEMATOTOX score and EASIX are the most extensively validated. The CAR-HEMATOTOX score, developed by Rejeski et al., is one of the most widely used tools for assessing hematologic toxicity in the CAR-T cell therapy setting. This model integrates six routinely available baseline parameters—PLT, ANC (absolute neutrophil count), hemoglobin, lactate dehydrogenase (LDH), CRP, and ferritin [20]. The CAR-HEMATOTOX score has been externally validated across multiple cohorts of relapsed/refractory hematologic malignancies, including large B-cell lymphoma (LBCL), mantle cell lymphoma (MCL), and MM treated with BCMA-directed CAR-T cell therapy. It demonstrated relatively consistent performance across these diverse disease settings [21, 22]. Consistent with previous studies, our meta-analysis showed that the CAR-HEMATOTOX score exhibited moderate discrimination for ICAHT. In addition, the wide prediction interval suggests variability in discrimination across external validation cohorts, whereas subgroup analyses stratified by event rate revealed no significant differences between high- and low-event-rate cohorts, suggesting a certain degree of consistency across baseline risk levels. Nevertheless, given the limited number of external validation cohorts, these findings require further confirmation in larger, independent cohorts.

In contrast, the EASIX was originally developed to assess endothelial dysfunction following allogeneic hematopoietic stem cell transplantation and was later adopted in the CAR-T cell therapy setting to predict CRS, ICANS, and ICAHT [39]. To enhance its performance for CAR-T cell therapy-related adverse events, several modified versions have been proposed, including m-EASIX, simplified EASIX (s-EASIX), and EASIX-ferritin/CRP (EASIX-F/C) [30, 58, 59]. All models within the EASIX family rely on routinely measured laboratory parameters-LDH, creatinine, platelet count, and CRP-offering advantages of simplicity and broad clinical applicability. Our findings indicate that the discrimination of EASIX and its derivatives varies substantially across different CAR-T cell therapy-related adverse events. For CRS, m-EASIX outperformed EASIX, which presented substantial heterogeneity and a broad prediction interval indicative of between-cohort variability. While subgroup analyses revealed no significant difference between pre- and post-infusion time points, EASIX performed better in B-cell malignancies than in plasma cell disorders, suggesting limited cross-disease generalizability. For ICAHT, EASIX showed moderate discrimination with low heterogeneity, and its narrow prediction interval confirmed consistent performance across external validation cohorts. In contrast, both EASIX and m-EASIX demonstrated limited discrimination for ICANS, with wide prediction intervals observed for both models. This result may be related to statistical uncertainty as well as the complex pathophysiology of ICANS, which involves alterations in blood–brain barrier integrity, neuroimmune dysregulation, cytokine dynamics, and baseline neurological status-critical biological domains not captured by current EASIX-based models [60]. Therefore, predicting ICANS remains a substantial challenge. Future studies may benefit from developing higher-quality models based on more comprehensive data and validating them in independent cohorts to clarify their optimal application scenarios and enhance model stability.

Strengths and limitations

The main strengths of this review lie in its comprehensive literature search, rigorous study selection, and standardized extraction of key characteristics from prediction models related to CAR-T cell therapy-related adverse events. In addition, the use of the PROBAST tool allowed for a structured assessment of risk of bias and facilitated a critical appraisal of methodological quality. By synthesizing the AUC from external validation cohorts for selected models, we further obtained pooled estimates that reflect their overall discrimination.

However, this review has several limitations. First, we excluded prediction models incorporating imaging variables or electrophysiological measurements, including one PET-based model for CRS prediction [61] and two ICANS prediction models based on Positron Emission Tomography (PET) [62] or electroencephalography (EEG) [63]. Because imaging- and EEG-based models differ fundamentally from models built on routine clinical and laboratory parameters in terms of data modalities and modeling approaches, pooling these models could introduce substantial methodological heterogeneity. From an implementation perspective, imaging and EEG examinations are not routinely performed in most CAR-T cell therapy monitoring settings and often require additional resources and specialized expertise. In contrast, models based on routinely available laboratory tests and clinical characteristics are generally more cost-efficient, reproducible, and amenable to broader clinical implementation. Therefore, this exclusion strategy may, to some extent, limit the applicability of our findings to prediction strategies that rely heavily on imaging or electrophysiological information. Second, calibration was generally underreported across the included studies, which precluded a meta-analysis of model calibration. Furthermore, for several prediction models, the scarcity of external validation cohorts limited the quantitative pooling of discrimination and the use of meta-regression to systematically explore sources of heterogeneity, and precluded a reliable assessment of publication bias. Previous studies have shown that CAR-T cell therapy-related adverse events are influenced by a wide range of factors, including demographic characteristics (age, sex, ethnicity), functional status, disease type, CAR-T product used, and prior treatment burden [4, 9, 28, 41, 64, 65]. Therefore, future prediction models should undergo extensive external validation across cohorts with diverse clinical and demographic characteristics to improve generalizability and robustness.

Conclusion

This systematic review and meta-analysis identified 26 prediction models developed to estimate the risk of adverse events in patients with hematologic malignancies undergoing CAR-T cell therapy. Pooled AUC estimates indicate moderate discrimination of the included models. However, all included models were associated with a high risk of bias due to methodological limitations, indicating that these estimates reflect discriminative ability under uncertainty rather than direct clinical readiness. Future studies should strictly adhere to established methodological standards when developing prediction models, improve study design, and conduct multi-center external validation to enhance the validity and reliability of adverse event prediction models in CAR-T cell therapy.

Abbreviations

ALL

Acute lymphoblastic leukemia

ANC

Absolute neutrophil count

AUC

Area under the receiver operating characteristic curve

BCMA

B-cell maturation antigen

CAR-T

Chimeric antigen receptor T-cell

CD19

Cluster of differentiation 19

CD20

Cluster of differentiation 20

CD22

Cluster of differentiation 22

CHARMS

Critical appraisal and reporting of a multivariable prediction model study

CI

Confidence interval

CNKI

China National Knowledge Infrastructure

CBM

China biology medicine disc 

CRP

C-reactive protein

CRS

Cytokine release syndrome

CRS

Cytokine release syndrome

CT

Computed tomography

DIC

Disseminated intravascular coagulation

EASIX-F/C

EASIX-ferritin/CRP

EEG

Electroencephalography

EPV

Events per variable

eIPM

Early immune effector cell–associated hematotoxicity prediction model

ICAHT

Immune effector cell–associated hematotoxicity

ICANS

Immune effector cell–associated neurotoxicity syndrome

IEC-HS

Immune effector cell-associated hemophagocytic lymphohistiocytosis-like syndrome

LBCL

Large B-cell lymphoma

LDH

Lactate dehydrogenase

MM

Multiple myeloma

NHL

non-Hodgkin lymphoma

OS

Overall survival

PCT

Procalcitonin

PET

Positron emission tomography

PFS

Progression-free survival

PLT

Platelet count

PROBAST

Prediction model risk of bias assessment tool

RR

Response rate

TRIPOD-SRMA

Transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analyses

XGBoost

Extreme gradient boosting

Authors’ contributions

YLR, LLQ, and XXH conceived and designed the study. LLQ and YLR performed the literature search and data extraction. ZXJ and WLL assessed the quality of the included studies. ZJT conducted the statistical analyses. YLR and LLQ drafted the manuscript, with critical revisions and supervision provided by XXH. All authors contributed to data interpretation, manuscript preparation, and review, and approved the final version for submission.

Funding

None.

Data availability

The data supporting the main findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Protocol of this study was registered at PROSPERO (CRD420251062713) since 28/5/2025. Amendment was submitted and has been approved on 9/8/2025. None reported.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.