The neurotoxic legacy of CAR-T cells: where do we stand?

Abstract

Chimeric antigen receptor T (CAR-T) cell therapy has transformed outcomes for relapsed/refractory B-cell malignancies and is increasingly reshaping the therapeutic landscape of autoimmune disorders and solid tumors, offering curative potential where options were previously limited. Its broader deployment is, however, constrained by immune-mediated toxicities, chiefly cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS). ICANS spans a heterogeneous spectrum from mild aphasia and tremor to seizures, cerebral edema, coma, and death, and remains difficult to predict prospectively. As CAR-T platforms expand beyond CD19 malignancies, neurotoxicity phenotypes are also broadening beyond classical ICANS. In plasma cell dyscrasias, BCMA-directed CAR-T has been associated with delayed non-ICANS neurotoxicities, including movement and neurocognitive/behavioral symptoms, cranial nerve palsies, and peripheral neuropathic presentations. In parallel, early experiences with CAR-T and related immune effector therapies in autoimmune and neuroimmunologic diseases suggest distinct inflammatory contexts and potentially different neurotoxicity patterns, underscoring the need for indication-specific monitoring and attribution frameworks. Converging data implicate a multilayered pathophysiology involving systemic cytokine surges, disruption of the blood–brain barrier, endothelial dysfunction, and context-dependent trafficking of activated CAR-T cells and other immune effectors into the CNS with baseline neurological vulnerability and the peri-infusion inflammatory milieu likely modulating individual risk. Given the frequency of these complications, an active research effort is underway to identify clinical, functional, and biological signals that could predict and improve their management. However, most biomarkers remain investigational, lacking prospective validation and straightforward clinical utility. This review synthesizes current evidence on the epidemiology, mechanisms, and monitoring of ICANS and emerging non-ICANS syndromes, and offers a fresh perspective on integrated, multimodal risk models to enable more precise stratification and timely intervention across indications.

Plain language summary

Understanding brain side effects of CAR-T cell therapy: why they happen and how we can prevent them

CAR-T cell therapy is a breakthrough treatment that can cure some blood cancers when all other options have failed. It works by reprogramming a patient’s own immune cells to recognize and destroy cancer cells. Despite these successes, many patients experience neurological side effects after receiving CAR-T cells. These effects, grouped under the term ICANS (immune effector cell-associated neurotoxicity syndrome), can range from mild confusion, tremors, or trouble speaking to more severe complications such as seizures, brain swelling, or even coma. Beyond this syndrome, these therapies can also cause other neurotoxic complications, including peripheral nervous system involvement and longer term cognitive impairment.

Central neurotoxicity develops through several overlapping processes. The treatment triggers a strong release of inflammatory molecules that disrupts the blood–brain barrier, allowing inflammatory signals and immune cells to enter the brain. Once inside, brain support cells like astrocytes and microglia release substances that can damage or overstimulate nerve cells, leading to swelling and impaired brain function. Some blood vessel support cells in the brain (pericytes and vascular smooth muscle cells) may also display the CD19 marker, creating the possibility of direct, unintended damage that further weakens the blood–brain barrier. The biology behind other types of neurological dysfunction remains incompletely understood.

Assessment of ICANS and other neurological complications after CAR-T cells combines bedside neurological testing, brain scans, electrophysiology testing, and sometimes spinal fluid analysis. Current treatments rely on rapid detection and anti-inflammatory medicines such as corticosteroids and biological therapies, including anti-interleukin treatments. Most cases improve with timely care, but severe ICANS can be life-threatening. Understanding these mechanisms may enable better prediction, safer therapy design, and more effective pr.

Introduction

Chimeric antigen receptor T (CAR-T) cell therapies have ushered in a new era of cancer treatment, transforming once-fatal hematologic malignancies into curable diseases with unprecedented remission rates.^{1 –8} Yet, these remarkable successes have been tempered by a constellation of novel, immune-mediated toxicities, most notably cytokine release syndrome (CRS), immune effector cell-associated neurotoxicity syndrome (ICANS), and the emerging immune effector cell-associated hematotoxicity (ICAHT).^{7,9 –11} Each of these iatrogenic syndromes has deepened our understanding of the complex biology of both engineered and endogenous T cells, revealing previously unrecognized roles of adaptive immunity in driving inflammation and tissue injury.^9,12 In doing so, they have shed light on fundamental mechanisms of T cell trafficking, effector function, and cross-talk with the endothelium and innate immune cells, offering new insights into how T cells can contribute to autoimmune and inflammatory processes beyond their intended therapeutic targets. This paradigm of “discovery by complication” echoes the formative years of allogeneic hematopoietic stem cell transplantation, when graft-versus-host disease first illuminated fundamental principles of T-cell biology and host–tissue interactions. ¹³ Just as those early transplant-related toxicities exposed the double-edged nature of immune activation, today’s CAR-T cell-associated adverse events offer a unique lens into how adoptively transferred lymphocytes—and the immunological cascades they trigger, traffic to, and sometimes damage, non-malignant tissues.

Amid this evolving landscape, neurotoxicity has emerged as one of the most unforeseen and clinically consequential complications of CAR-T cell therapy. ¹⁴ From the earliest trials, patients experienced a range of neurologic disturbances (including headache, confusion, aphasia, tremor, and, in severe cases, seizures and brain swelling) often in close temporal association with CRS, but at times occurring independently.^14,15 These manifestations were eventually unified under the term ICANS, defining a clinical spectrum that spans from mild, self-limiting encephalopathy to fulminant cerebral edema. Although the pathophysiology of ICANS remains incompletely understood, key mechanisms include disruption of the blood–brain barrier (BBB), endothelial activation, and a surge of pro-inflammatory cytokines.^{14 –16}

In parallel with ICANS, a growing array of non-classical neurologic complications has further revealed the neurotoxic reach of CAR-T cell therapies. These include movement and neurocognitive toxicities (MNT), ¹⁶ particularly observed with BCMA-targeting CAR-T cells in multiple myeloma, as well as rarer but clinically significant events such as cranial nerve palsies, cerebellar dysfunction, myelitis, and ischemic strokes. Tumor inflammation-associated neurotoxicity (TIAN) has been described in patients with primary CNS lymphoma, where localized inflammation at sites of tumor involvement can lead to mass effect, hydrocephalus, or seizures. ¹⁷ Peripheral nervous system complications have also been reported, including demyelinating syndromes and Guillain–Barré-like presentations. In addition, neurotoxicity may rarely arise from the lymphodepleting/conditioning regimen itself, including direct fludarabine-associated neurotoxicity, a clinical entity described well before the CAR-T era. While these syndromes fall outside the canonical definition of ICANS, they highlight that neurologic complications in CAR-T recipients may reflect both immune-mediated toxicity and non-immune effector, regimen-related neurotoxicity, affecting neural structures from cortex to cauda equina.

In this review, we will outline the clinical hallmarks of ICANS and other CAR-T cell-related neurotoxicity, explore emerging mechanistic insights, and examine current and future strategies for prediction, prevention, and management.

Principles of epidemiology and pathophysiology of CAR-T cell neurotoxicity

The incidence of neurotoxicity following CAR-T cell therapy varies depending on multiple factors, including the target antigen, CAR construct, disease type, patient and disease-specific characteristics, and co-occurrence of CRS. ICANS has been most frequently observed in agents targeting CD19-positive cells, used for B-cell malignancies, with reported incidence rates ranging from 20% to over 60% in pivotal trials.^1,2,18 Severe (grade ⩾3) neurotoxicity occurs in approximately 10%–30% of cases. However, this varies markedly between products, being more common with second-generation CARs incorporating CD28 costimulatory domains (e.g., axicabtagene ciloleucel) compared to those using 4-1BB (e.g., tisagenlecleucel), which are generally associated with delayed but less intense immune activation.^19,20 In BCMA-targeting CAR-T cells for multiple myeloma, the overall incidence of neurotoxicity appears lower, with MNT syndromes increasingly recognized.^16,21 Risk factors for ICANS include high disease burden, early and severe CRS, elevated serum cytokines (e.g., IL-6, IL-15), and pre-existing neurologic comorbidities.^15,17,20 While the majority of ICANS episodes are reversible with prompt management, the risk of long-term sequelae or fatal outcomes (though rare) remains a significant concern in high-grade cases. ²²

The pathophysiology of CAR-T cell neurotoxicity is still far from being completely understood. A complex interplay of factors, including systemic inflammation, endothelial and microglial activation, and localized immune responses, converges to produce neurodysfunction. The chronological connection with CRS during early CAR-T cell expansion suggests a key role for cytokines such as IFN-γ, IL-6, IL-15, IL-1β, GM-CSF, TNF-α, CXCL8, and CCL2, likely driven by both CAR-T and myeloid cells.^15,23 Monocyte-derived IL-1 and IL-6 have been directly implicated in both CRS and ICANS pathogenesis.^14,22,23 This cytokine wave activates the endothelium, as evidenced by elevated levels of intercellular adhesion molecule-1 (ICAM-1), vascular cell adhesion molecule-1 (VCAM-1), von Willebrand factor, angiopoietin-2, matrix metalloproteinases, and VEGF, which collectively destabilize the BBB and instigate capillary leakage. ²⁴ Once the BBB integrity is compromised, cytokines and, occasionally, CAR-T cells infiltrate the CNS. CSF analyses during ICANS episodes frequently detect elevated pro-inflammatory mediators, astrocytic injury markers such as GFAP and S100B, and neurofilament light chain (NfL), which may function as brain and neuroaxonal damage indicators. A rhesus macaque model demonstrated perivascular T-cell invasion with panencephalitis, ²⁵ whereas human post-mortem analyses revealed BBB disruption without extensive CNS T-cell infiltration, suggesting that edema and endothelial dysfunction, rather than direct cytotoxicity, might play central roles. ²⁵ Astrocytes and microglia, once activated, may sustain neuroinflammation through glutamate release and excitotoxicity, as evidenced by increased CSF glutamate and quinolinic acid early during ICANS. ¹⁴ A landmark study identified CD19 molecules expressed in human brain mural cells, including pericytes and vascular smooth muscle cells, as a potential source of neurotoxicity. These mural cells are essential for maintaining BBB integrity, and importantly, the CD19 isoform expressed in the brain contains the FMC63 epitope recognized by clinical CAR-T cells and bispecific antibodies, supporting an on-target, off-tumor mechanism of neurotoxicity. ²⁶ In contrast to humans, CD19 expression in murine mural cells is sparse, underscoring the limitations of preclinical mouse models. ²⁶ This framework has practical implications for target selection and construct development (e.g., antigens such as BCMA or CD22 may reduce the likelihood of CD19-mural–cell engagement, while still remaining susceptible to cytokine/endothelial pathways), and it may help explain why neurologic toxicity patterns can differ across indications: autoimmune applications may feature lower baseline antigenic burden and different inflammatory kinetics, potentially attenuating systemic endothelial injury despite shared B-cell targets.^27,28

In pre-clinical models, GM-CSF knockout in CAR-T cells has been shown to curtail IL-6-driven endothelial activation, reducing ICANS risk in preclinical models without compromising anti-tumor activity.^29,30 IL-1 blockade, but not IL-6 inhibition, prevented meningeal inflammation in macaque models, providing a rationale for clinical trials targeting IL-1 in ICANS prevention. ³¹ Brain capillary obstruction has been identified as a potential trigger of neurotoxicity in mouse models of ICANS, where increased expression of adhesion molecules such as ICAM-1 and VCAM-1 was observed in CAR-T-treated animals developing this complication. ³²

In patients, elevated plasma levels of soluble ICAM-1 and VCAM-1 have similarly been associated with ICANS and may serve as early biomarkers. ³³ Extracellular vesicles from CAR-T and myeloid cells are also emerging as potential mediators and biomarkers of neurotoxicity. They can be measured in the plasma of CAR-T cell-infused patients. ³⁴ Additional modifiers include catecholamines that amplify cytokine release and hypophosphatemia, potentially exacerbating neurologic vulnerability. ²⁵

Altogether, these findings underscore that ICANS does not arise from a single pathway but rather from the convergence of systemic cytokine-driven inflammation, endothelial and glial activation, and BBB disruption. This complexity highlights the need for integrative biomarkers and targeted interventions to prevent or mitigate this potentially life-threatening complication. In addition, a plausible biological basis for the high incidence of neurotoxicity observed in patients treated with CD19-directed immunotherapies has been identified, emphasizing the importance of refining antigen-targeting strategies to improve safety. ²⁶

Clinical manifestations of ICANS

ICANS typically begins 3–6 days after CAR-T cell infusion, peaks around days 7–8, and generally resolves within days if promptly treated. The clinical picture ranges from transient cognitive impairment (variably termed confusion, delirium, or encephalopathy) to a striking language dysfunction (word-finding pauses to mutism) that often accompanies cognitive changes.^20,15,35 Tremor and headache are common but nonspecific prodromes; with increasing severity, patients may develop depressed consciousness to coma. Seizures (clinical and electrographic) occur in ~5%–10% overall, with reported ranges of 0%–30% after CD28-costimulated CD19 CARs and 3%–14% with 4–1BB constructs; focal neurological deficits are less frequent, although electroencephalographic (EEG) abnormalities are common.^{22,35 –38} Rarely, cortical cytotoxic edema in the acute phase evolves to chronic focal injury with gliosis. ²³ The most feared complication is cerebral edema, estimated at ~1%–2% among CD19-CAR recipients and reported across pediatric and adult populations with acute lymphoblastic leukemia, chronic lymphoid leukemia, and non-Hodgkin Lymphoma and with both CD28 and 4-1BB CARs. ¹⁵

ICANS monitoring

Clinical assessment of ICANS

ICANS monitoring currently relies on clinical assessment, with the American Society for Transplantation and Cellular Therapy (ASTCT) Immune Effector Cell-Associated Encephalopathy (ICE) score serving as the standardized bedside tool to objectively identify and grade clinical manifestations of CNS toxicity. ³⁹ The ICE score is a 10-point cognitive test designed for rapid evaluation at the bedside, targeting higher cortical functions typically impaired during neurotoxicity. It assesses orientation, object naming, command following, writing ability, and attention, with a normal score of 10/10. Declines in performance reflect increasing severity of encephalopathy: scores of 7–9 denote mild impairment, 3–6 moderate, and 0–2 severe. In practice, the ICE score is interpreted alongside other neurological features (including changes in consciousness, seizures, motor deficits, and radiologic evidence of cerebral edema) to establish the final ICANS grade according to ASTCT criteria. Although simple and widely adopted, the ICE score captures neurotoxicity only once neurological dysfunction is clinically manifest, limiting its utility for early detection or preemptive intervention.

Biological monitoring of ICANS

Notably, CAR-T cells are often detectable in CSF during the post-infusion window, but contemporaneous CSF controls are scarce. CSF is rarely obtained from patients without ICANS in the 10–14 days when ICANS typically develops, making it difficult to derive predictive indicators from CSF presence alone. In a pediatric CD19 CAR-T cell cohort with protocolized day-21 lumbar punctures, CAR-T cells were found in the CSF of all patients with ICANS and in 90% of those without neurological symptoms, arguing that mere trafficking is insufficient to explain neurotoxicity. ⁴⁰ CSF protein is frequently elevated (median 80–110 mg/dL, occasionally >1000 mg/dL) and tends to increase with neurotoxicity severity, while CSF cytokines are commonly high but usually mirror serum levels, with occasional enrichment of IL-8, CXCL10, or MCP-1.^14,15,22,41 Taken together, these findings support increased permeability of the BBB b and greater CNS exposure to the systemic secretome during ICANS. More broadly, immunomonitoring strategies remain highly heterogeneous across studies (in sampling time points, biospecimens (blood vs CSF), assay platforms, and analytic pipelines), limiting cross-study comparability and slowing clinical translation. ⁴²

Numerous candidate biomarkers have been proposed, but most remain investigational and not validated for routine clinical use.

Early cytokine panels (notably IL-6, IL-15, IFN-γ, IL-10, and GM-CSF) rise in patients who develop ICANS, mirroring CRS biology, and higher peaks/early surges correlate with worse grades. ¹⁴ Endothelial activation markers such as angiopoietin-2, von Willebrand factor, soluble ICAM-1, and VCAM-1 have been associated with BBB disruption and worse neurotoxicity, yet prospective validation is absent.^15,23 Similarly, neural injury markers, including GFAP, S100B, and NfL, tend to increase during ICANS and may reflect astroglial and/or axonal damage, although their clinical utility remains unproven.^14,23,43 Composite indices such as the modified EASIX, together with routine inflammatory markers (CRP, ferritin), have been associated with ICANS risk and represent simple, routinely applicable measures.^14,22,44 Integrating immune-inflammatory indices with neuronal injury markers could ultimately enable earlier and more accurate risk stratification, but their translation into real-world practice remains challenging. Accordingly, the field requires complementary, scalable tools capable of providing early, actionable assessments of neurotoxicity risk.

The role of EEG

EEG has emerged as a noninvasive, bedside-accessible method for monitoring brain function in real-time during ICANS. While underutilized in current guidelines, several recent studies underscore its potential role in both pre-therapy risk stratification and early neurotoxicity detection and follow-up. A multicenter Spanish study showed that abnormalities on preinfusion EEG were significantly associated with the subsequent development of severe ICANS. Notably, 89% of patients with ICANS exhibited EEG abnormalities at onset, including encephalopathy and interictal epileptiform discharges (IEDs), with the latter strongly associated with rapid clinical deterioration. ³⁸ Among patients with grade 1 ICANS and IEDs, 50% progressed to higher-grade neurotoxicity within 24 h, suggesting a narrow but actionable therapeutic window. ³⁸ Another study identified frontal intermittent rhythmic delta activity (FIRDA) as a highly sensitive (88%) and specific EEG marker for ICANS, with a negative predictive value of 100%, providing a robust argument for the routine use of bedside EEG in high-risk settings. ⁴⁵ Further observations confirmed EEG’s practical advantage, reporting that EEG-informed clinical decision-making more frequently than magnetic resonance imaging (MRI) or lumbar puncture in a real-world cohort. ⁴⁶ A recent report focused on quantitative EEG severity, which correlated with ICANS grade, underpinning its role in indexing encephalopathic burden. ⁴⁷

Despite their value, EEG markers have limitations. Abnormal findings, including IEDs and FIRDA, are highly predictive when present, yet they are neither universal nor specific to ICANS. Many patients who display these patterns never develop severe neurotoxicity, whereas others with initially unremarkable EEGs worsen abruptly. Relying on EEG alone, therefore, risks both overtreatment and undertreatment, each with critical clinical consequences: unnecessary corticosteroids or anti-cytokine agents can impede CAR-T cell expansion and heighten infection risk, while delayed intervention may allow fulminant neurotoxicity to progress unchecked.

The role of brain imaging

Neuroimaging in ICANS serves three purposes: excluding alternative diagnoses (e.g., infection and hemorrhage), identifying patterns that support the clinical diagnosis, and detecting fulminant complications such as cerebral edema, especially in clinically severe scenarios. Across reports, MRI abnormalities in ICANS mirror those seen in inflammatory and infectious encephalopathies, and can be absent in mild cases. Thus, a normal study does not exclude ICANS.^22,48 In cohorts imaged around day 7 after CAR-T cell infusion, ≈32.5% of patients show clinically meaningful abnormalities. ⁴⁹ The most frequent signatures are FLAIR hyperintensities (~50%) and diffusion restriction consistent with cytotoxic injury (~22.5%); less commonly, contrast enhancement (~18.8%), vasogenic/interstitial edema (~8.2%), and hemorrhage (~4.1%) are observed. Lesions preferentially involve periventricular white matter (~21.4%), brainstem (pons ~19.4%, medulla ~17.3%), thalami (~19.4%), and hippocampi (~11.2%); leptomeningeal involvement occurs in ~10%, while subarachnoid findings are rare (~2%). ⁴⁹ The likelihood of abnormal imaging increases with ICANS grade. ⁵⁰ Cohort studies further refine this phenotype, describing limbic-predominant patterns in neurotoxic patients, ¹⁴ encephalitis, and posterior reversible encephalopathy that track systemic inflammatory burden (e.g., ferritin and C-reactive protein), ⁵¹ and focal vascular complications including tumor-associated thrombosis and petechial hemorrhages. ⁵² Taken together, the predominantly symmetric distribution and deep-gray predilection suggest that ICANS is driven by a systemic process, notably cytokine-mediated endothelial/BBB dysfunction, engaging final common pathways also active in infectious and inflammatory encephalopathies. Imaging, therefore, complements, rather than replaces, clinical and laboratory assessment. From a practical standpoint, CT may be normal early but is essential if fulminant cerebral edema is suspected; MRI is more sensitive to the patterns above. The field still lacks prospective data defining when imaging alters management in ICANS and clarifying the clinical utility of imaging (e.g., as an adjunct to EEG), standardizing when and how it should be deployed in monitoring algorithms.

Latent and subclinical neurotoxicity

Although ICANS is typically portrayed as an acute, reversible encephalopathy, bedside resolution of symptoms may not capture the full neurotoxicity burden. A subset of patients reports persistent or delayed neurocognitive and neuropsychiatric sequelae, including attention/executive dysfunction, slowed processing speed, memory complaints, mood or sleep disturbance, and reduced quality of life, despite normalization of ICE-based assessments and discharge from acute monitoring.^53,54 These “latent” outcomes are challenging to attribute because they may overlap with effects of prior therapies, steroids, ICU delirium, infection, anemia, and psychosocial stressors, and because structured neuropsychological testing and patient-reported outcomes are not routinely incorporated into CAR-T follow-up.^55,56 This gap argues for standardized longitudinal assessment (brief cognitive screening at baseline and post-infusion milestones, with formal neuropsychological evaluation in symptomatic patients), and for future studies linking acute neurotoxicity biology (endothelial injury and neuroinjury markers, EEG abnormalities) to longer-term functional outcomes.

Therapeutical management

Therapeutic management of ICANS centers on early recognition and standardized grading, using the ICE score to detect early cognitive changes and to trigger prompt supportive care and stepwise escalation by severity. ⁸ Mild (grade 1) events are managed supportively: frequent neuro checks, prompt correction of fever, hypoxia, and electrolytes, minimization of sedatives, and early neurology input; whereas patients with grade ⩾3 warrant ICU-level monitoring. Corticosteroids remain the cornerstone for grade ⩾2 ICANS: dexamethasone 10–20 mg IV every 6 h with a rapid taper as symptoms improve; for fulminant/grade 4 presentations, methylprednisolone 1 g/day for ~3 days before tapering is recommended. ⁸ Seizures are treated with levetiracetam, and EEG monitoring is prioritized when mental status worsens; routine antiseizure prophylaxis is institution-dependent, and evidence is mixed, with some centers adopting early or pre-emptive levetiracetam in higher-risk settings. ⁵⁷ In a recent multicenter, propensity-matched retrospective study, prophylactic levetiracetam did not reduce ICANS and was associated with a higher incidence of ICAHT, arguing against routine antiseizure prophylaxis in unselected patients. ⁵⁸ Anti-IL-6 therapy is reserved for concomitant CRS: tocilizumab does not reverse isolated ICANS and may theoretically aggravate neurologic toxicity, so steroids are prioritized when CRS and ICANS co-exist. ⁵⁹ For steroid-refractory ICANS, including fulminant cases with evolving cerebral edema, IL-1 blockade with anakinra (IV/SC) is increasingly used, supported by multicenter real-world series (doses ranging from 100 mg q6–8 h to ~2–12 mg/kg/day) and early prospective efforts. Cyclophosphamide has also recently shown signals of benefit for severe non-ICANS neurotoxicities in a retrospective series. ⁶⁰

In parallel, prevention/treatment strategies are rapidly evolving: prophylactic tocilizumab reduces CRS but is insufficient to prevent neurotoxicity^59,61; prophylactic corticosteroids may modestly attenuate CRS yet have not prevented ICANS and raise concerns about blunting CAR-T cell expansion and antitumor activity ⁶² ; prophylactic anakinra has shown encouraging reductions in severe ICANS without impairing CAR-T cell kinetics^62,63; endothelial protection with defibrotide produced a modest decrease in ICANS but the phase II prophylaxis study was terminated early for futility ⁶⁴ ; direct IL-6 neutralization with siltuximab is under prospective evaluation, with signals generally stronger for CRS than for ICANS (NCT04975555) with growing real-world and interim trial signals^65,66; upstream JAK-STAT dampening with the JAK1 inhibitor itacitinib shows emerging randomized evidence of CRS mitigation with possible benefit for ICANS. ⁶⁷ As well, upstream myeloid modulation via GM-CSF neutralization (e.g., lenzilumab) has shown encouraging signals as prophylaxis with low rates of severe CRS/ICANS in early-phase experience, ⁶⁸ although the development program was terminated before phase II expansion (NCT04314843). Intrathecal approaches (IT cytarabine/methotrexate/hydrocortisone) are also prospectively tested to prevent high-grade ICANS (NCT06895473), supported by case-series experience in steroid-refractory disease.^69,70

Atypical neurotoxicity syndromes

As CAR-T platforms expand beyond CD19 malignancies, neurotoxicity phenotypes are diversifying beyond “classical” ICANS. In plasma cell dyscrasias, BCMA-directed CAR-T (notably cilta-cel) has been associated with delayed, non-ICANS syndromes encompassing a clinically distinct spectrum that includes movement and neurocognitive treatment-emergent adverse events (MNTs) (characterized by movement disorders, cognitive symptoms, and personality/behavioral changes) as well as cranial nerve palsies and other peripheral neuropathies, including Guillain–Barré-like syndromes.^{71 –73} Notably, MNTs can present in a delayed fashion, may be progressive and corticosteroid-refractory, and have been associated with severe outcomes in a subset of patients. Although the pathophysiology remains incompletely defined, high CAR T-cell expansion/transgene levels have repeatedly emerged as a potential risk factor, supporting the need for dedicated surveillance and tailored mitigation strategies distinct from standard ICANS algorithms.^71,74 This non-ICANS neurotoxicity may occur late and may not track with the usual CRS/ICANS timing or grading frameworks. Early recognition is critical, and emerging management approaches include escalation beyond corticosteroids (case-based strategies such as CAR-T ablation with cyclophosphamide or intrathecal approaches have been reported in steroid-refractory cases).^60,71 In parallel, CAR-T for CNS/solid tumors can trigger a distinct syndrome termed TIAN, conceptually overlapping with pseudoprogression but broader than edema alone, and clinically dominated by focal deficits, seizures, and increased intracranial pressure/hydrocephalus driven by intense on-target intratumoral inflammation.^17,75 TIAN has been highlighted across early-phase neuro-oncology CAR-T experiences (e.g., GD2 CAR-T in H3K27M-mutant DMG/DIPG and IL-13Rα2 CAR-T in recurrent high-grade glioma), where neurocritical monitoring and targeted anti-inflammatory strategies (including IL-1 blockade in some protocols) are central to safely advancing efficacy. ⁷⁵

Neurotoxicity considerations in autoimmune and neuroimmunologic indications

Beyond hematology, CAR-T and related immune effector cell therapies are increasingly being explored in autoimmune and neuroimmunologic diseases, where both the therapeutic targets and the neurotoxicity landscape may differ from classical CD19 malignancy settings. Early clinical experiences with CD19-directed CAR-T in systemic lupus erythematosus and other refractory autoimmune diseases suggest that, despite profound immune resetting, high-grade ICANS appears uncommon compared with hematologic malignancies, potentially reflecting lower baseline systemic inflammatory burden, differences in disease-related endothelial vulnerability, and distinct cytokine kinetics.^{76 –78} Conversely, when immune effector strategies are deployed in patients with primary neurologic autoimmunity (e.g., chronic inflammatory demyelinating polyradiculoneuropathy, myasthenia gravis, or broader antibody-mediated disorders), a key concern is not only classical ICANS but also the possibility of disease-specific neurological destabilization (e.g., fluctuating neuromuscular weakness, peripheral nerve demyelination, or autonomic dysfunction) and the challenge of differentiating treatment-related toxicity from disease activity, infection, metabolic derangements, or concomitant immunosuppression.^76,79,80 These emerging indications therefore emphasize the need for neurologist-tailored surveillance paradigms, including baseline and longitudinal neurofunctional assessments, low thresholds for EEG in encephalopathy/seizures, and context-specific use of CSF and neuroinjury markers, while also offering a unique “human model” to disentangle immune effector-driven neuroinflammation from confounders intrinsic to heavily pretreated cancer populations. As these applications expand, comparative analyses across oncology versus autoimmune cohorts may clarify which components of neurotoxicity are primarily driven by immune effector expansion and cytokine dynamics, versus host susceptibility (pre-existing CNS/PNS pathology, BBB integrity, and endothelial reserve), informing both mitigation strategies and rational trial design. ⁸¹

Pragmatically, we suggest an indication-adapted monitoring bundle for autoimmune/neuroimmunologic CAR-T cell indications: (i) pre-infusion baseline neurological examination with a disease-specific functional measure (e.g., neuromuscular/neuropathy disability scale where relevant) and a brief cognitive/psychiatric screen; (ii) early post-infusion daily focused neurological assessments, using ICE/ASTCT when encephalopathy is present but not as a stand-alone tool, with a low threshold for EEG in any confusion, aphasia, suspected seizures, or fluctuating mental status; and (iii) structured post-discharge follow-up (e.g., 4–6 weeks and ~3 months) to capture delayed or subclinical sequelae (cognition, mood, sleep/fatigue, focal deficits) and to systematically evaluate PNS/autonomic symptoms when applicable, with targeted CSF/neuroinjury markers reserved for atypical, persistent, or severe presentations.

Current limitations and perspectives

Ongoing work to capture subclinical neurotoxicity after CAR-T is moving beyond bedside encephalopathy scoring toward standardized longitudinal neurocognitive follow-up, EEG-based continuous/quantitative assessment, and fluid biomarkers. Prospective/real-world cohorts integrating patient-reported outcomes and objective neurocognitive testing show that perceived cognition, mood, and specific cognitive domains can fluctuate for weeks to months after infusion (sometimes despite apparent clinical recovery), supporting baseline and serial assessments rather than reliance on ASTCT/ICE scores alone. ⁵⁴ In parallel, EEG-driven tools (including graded EEG abnormalities and continuous EEG in selected patients) have been proposed to detect early network dysfunction, non-convulsive seizures, or preclinical deterioration and to provide a quantitative correlate of ICANS severity (and ongoing dysfunction). ⁸² As seen, “fluid biopsy” approaches are emerging, with neuroaxonal injury markers such as serum/plasma NfL elevated at baseline or early after infusion in patients who later develop more severe neurotoxicity, suggesting a role for identifying vulnerable patients and monitoring subclinical CNS injury; additional candidate pathways (e.g., kynurenine metabolites linked to mood/neurotoxicity) are also under evaluation. ⁸³

Despite their value, EEG, radiological, and immune-inflammatory markers have clear limitations for early detection of neurological complications after CAR-T cell therapy. First, no adequately powered prospective study has established the predictive value of these parameters (alone or in combination) across cohorts with and without neurotoxicity. Second, while abnormal functional or imaging findings can be highly predictive when present, they are neither universal nor specific to ICANS: many patients with EEG abnormalities never develop severe neurotoxicity, whereas others deteriorate abruptly despite initially unremarkable EEGs. Likewise, MRI/CT signs lack sensitivity and specificity for immune-related neurotoxicity, and circulating biomarkers often reflect systemic inflammation more than CNS-specific processes. Relying on these signals in isolation risks both overtreatment and undertreatment: unnecessary corticosteroids or anti-cytokine agents may blunt CAR-T cell expansion and heighten infection risk, while delayed intervention may permit fulminant neurotoxicity to progress unchecked. Consequently, individualized, multimodal assessment, integrating longitudinal clinical, electrophysiological, imaging, and laboratory data, is essential to align each patient with the most appropriate, timely intervention.

Virtual patients and personalized medicine in CAR-T cell-related neurotoxicity

More recently, efforts have shifted from static risk stratification to patient-specific prediction using “digital-twin” approaches. ⁸⁴ In neurology, individualized computational models have already been used to simulate brain dynamics and support clinical decisions, for example, to help identify candidate seizure-generating regions in drug-resistant epilepsy, to infer myelin-related damage beyond conventional MRI sensitivity in multiple sclerosis, or to simulate the impact of different L-DOPA doses on brain activities in Parkinson’s disease.^{85 –88} These methods share a common principle: a mechanistic model of physiological interactions (often expressed as differential equations) whose parameters are personalized using patient data, such as structural connectivity (“connectome”) or electrophysiological features.^{89 –91} Once personalized, the model can generate time-varying outputs (e.g., simulated EEG signals), thereby creating a virtual representation of an individual patient, termed “virtual twin.”^82,83,85 In simple terms, a virtual model is a computer-based representation of a patient (or organ/tissue) that integrates clinical, biological, imaging, and EEG data to simulate disease-relevant processes and predict how risk may evolve under different interventions. Similar virtual-patient strategies are now being explored in oncology, epidemiology, and systems biology,^{92 –95} and hybrid frameworks that combine mechanistic modeling (when biology is understood) with data-driven learning (when cross-system interactions are uncertain) may provide a route from bedside observations to actionable prevention strategies for ICANS.^89,96

Despite the recent advances and remarkable successes in these fields, relatively little has been done to develop models that account for interactions between the nervous system and the immune system. However, the evidence showing that alterations in the pre-infusion EEG carry predictive power with respect to the development of severe ICANS suggests that the brain is not limited to a “passive spectator,” but its state at the beginning of the therapy is relevant to the development of side effects such as ICANS. In practice, we propose a multiscale digital twin, which would accommodate (i) a systemic immune module (projections for CAR-T cell expansion and cytokine/myeloid feedback) to (ii) an endothelial/BBB state model (informed by factors such as inflammation, coagulation, permeability) and to (iii) a whole-brain model (e.g., coupled neural masses) that reproduces EEG signatures (e.g., FIRDA, IEDs). In particular, the “Virtual-patient” models, combining mechanistic physiology, when pathophysiology is known, with data-driven learning, when interactions between body systems are unknown, could become the missing link between bedside observations and actionable prevention of ICANS.^89,96 As observed for other diseases, this kind of personalized twin could be deployed to (i) infer latent states of neurovascular risk before symptoms, (ii) provide individualized forecasts of ICANS onset and grade, and (iii) run in silico trials to rank prophylactic or early interventions (e.g., IL-1 or GM-CSF blockade, endothelial stabilizers, and steroid timing/dosing) under patient-specific constraints. To ensure clinical utility, models should incorporate causal structure (e.g., graphical models and counterfactual simulators) that provide predictions about potential alternative therapeutic strategies, thereby separating correlation from mechanism and hypothesizing personalized treatment outcomes. While this is a major research direction of our group, we emphasize that digital-twin implementations should be considered a long-term research vision rather than a near-term clinical tool.

Take-home message for clinical neurologists

At this stage, the neurotoxic legacy of CAR-T therapy is less a single syndrome than a shift in how neurologists define, detect, and attribute immune effector-related neurological injury. First, neurotoxicity phenotypes now extend beyond classical, early ICANS to include delayed and non-ICANS presentations (e.g., movement, neurocognitive, behavioral, cranial nerve, or peripheral neuroimmune syndromes), whose timing and clinical signatures may differ across targets and disease settings. Second, apparent bedside recovery does not necessarily equal neurological recovery: latent or subclinical cognitive, affective, and sleep sequelae can persist or fluctuate for weeks to months and are not captured by encephalopathy scores alone. Third, as CAR-T expands into autoimmune and neuroimmunologic indications, attribution and risk stratification become indication-specific, requiring explicit separation of treatment-related toxicity from baseline neurological vulnerability, disease activity, infection, and concomitant immunosuppression.

In practice, this synthesis supports a pragmatic change in current care pathways: (i) implement baseline neurological assessment and risk profiling (including prior CNS disease, cognitive reserve, and seizure history) before infusion; (ii) pair early bedside grading with indication-triggered electrophysiological and biomarker-informed assessment rather than relying on ICE/ASTCT scoring alone; (iii) organize structured post-acute follow-up focused on cognition, mood, fatigue, sleep, and functional status to detect latent sequelae and guide rehabilitation; and (iv) embed these steps into prospective, harmonized datasets to enable the validation of expanded neurotoxicity scores and monitoring frameworks as the most realistic near-term advance for the field.

Conclusion

CAR-T neurotoxicity has matured from an acute encephalopathy paradigm into a heterogeneous spectrum of immune effector-related neurological syndromes that can be early or delayed, central or peripheral, and increasingly indication-dependent. The practical consequence for clinical neurology is a move from reactive grading toward structured pathways: baseline neurological assessment and risk profiling, multimodal evaluation when symptoms emerge, and longitudinal follow-up that captures cognition, mood, sleep, and function beyond hospital discharge. In the near term, the field will advance most through harmonized monitoring bundles and the prospective development and validation of expanded neurotoxicity scores that are applicable to both malignant and autoimmune/neuroimmunologic CAR-T settings (Figure 1, Box 1).

Figure 1.

Proposed pathophysiology and multimodal assessment of ICANS. (a) Pathophysiology. After CAR-T engagement and tumor killing, a systemic cytokine surge (illustrative mediators shown: IL-1, IL-6, IFN-γ, IL-10, IL-15, GM-CSF, TNF-α) and a myeloid amplification loop (monocyte/macrophage IL-1/IL-6) drive endothelial activation at the neurovascular unit (↑ Ang-2:Ang-1, vWF release, ↑ ICAM-1/VCAM-1), resulting in capillary leak and increased BBB/BCSFB permeability. Within the CNS, astrocyte and microglial activation promotes excitotoxic mediators (glutamate, quinolinic acid), and platelet/vWF microthrombi may form, together producing neuroinflammation characteristic of ICANS. CD19 expression on mural cells (pericytes and vascular smooth muscle cells) provides a potential on-target substrate for CAR-T cell-mediated injury, further contributing to BBB disruption. Trafficking of CAR-T cells into CSF can occur, but is not required for neurotoxicity. (b) Multimodal assessment. Conceptual timeline aligning CAR-T expansion kinetics with the typical temporal windows of CRS and ICANS, and delayed NINT. ICANS is commonly characterized by headache, aphasia, delirium, tremor, ataxia, myoclonus, and seizures, whereas NINT encompasses delayed movement and neurocognitive/behavioral symptoms and peripheral neuropathic presentations (including Guillain–Barré-like syndromes). Icons indicate complementary monitoring modalities deployed along the clinical course, including bedside neurologic assessment/ICE score and immunomonitoring, EEG, brain MRI, and CSF analysis. The time course and associations are schematic and may vary across products and indications.

Box 1.

Practical points.

• CAR-T neurotoxicity is broader than “classical ICANS” and can be delayed/subclinical. The clinical spectrum is expanding (non-ICANS syndromes, delayed presentations, peripheral nervous system phenotypes), and neurologists should actively consider these entities rather than forcing all events into the ICE/ASTCT framework. • Bedside scores grade severity but miss key phenotypes; monitoring should become structured and longitudinal. The practical consequence is to implement baseline neuro risk assessment and standardized longitudinal follow-up beyond hospital discharge. • The most actionable next step is validated, expanded neurotoxicity scores, and tailored monitoring frameworks that incorporate non-ICANS phenotypes, cognition/behavior/sleep outcomes, and scalable biomarkers/EEG features. • CAR-T in other fields than malignant hematology (i.e., autoimmune/neuroimmunologic disorders) needs indication-specific monitoring. Rates of high-grade ICANS may differ, but neurotoxicity can be qualitatively different, and attribution is harder (baseline neuro disease, relapse/flare, infections, concomitant immunosuppression). Monitoring should be tailored rather than copied from lymphoma trials. What should change tomorrow in practice: 1. Document a baseline neurological status (including cognition/psychiatric symptoms when feasible). 2. Use symptom-triggered EEG/neurology assessment early. 3. Plan a post-acute follow-up window (e.g., 4–12 weeks) focused on cognition, mood, fatigue, sleep, and focal deficits. 4. Feed standardized data into registries/prospective studies to enable score validation.

ASTCT, American Society for Transplantation and Cellular Therapy; CAR-T, chimeric antigen receptor T; EEG, electroencephalographic; ICANS, immune effector cell-associated neurotoxicity syndrome; ICE, immune effector cell-associated encephalopathy score.