IntroductionVirtual reality (VR) provides an immersive environment for inducing emotional experiences, offering a naturalistic framework for investigating brain network dynamics. However, traditional emotional neuroscience has largely focused on regional activations, leaving the topological robustness and adaptive capacity of integrated brain networks underexplored. This study addresses this gap by applying a graph-theoretical framework to quantify how different emotional states modulate the resilience of functional architectures against systematic disruptions.MethodsIn this study, we examined the resilience of EEG-based functional brain networks during negative, neutral, and positive emotional states induced by VR stimuli. Functional connectivity was computed using coherence across six frequency bands (delta to high gamma), and graph-theoretical measures were applied to characterize network topology. To assess resilience, we simulated network disruptions using two complementary approaches: targeted attacks (removing high-centrality nodes) and random failures (removing nodes randomly). Changes in global efficiency and the largest connected component were tracked as nodes were progressively removed.ResultsOur findings revealed emotion-specific resilience profiles. In the alpha band, both negative and positive emotions demonstrated enhanced resilience to targeted attacks compared to neutral states, maintaining higher efficiency and greater network integrity, with positive emotions showing particularly strong preservation of large-scale connectivity. In the high gamma band, networks during negative emotional states exhibited greater robustness than those during positive emotions, indicating enhanced capacity to withstand targeted disruptions.DiscussionThese findings suggest that emotional experiences are associated with differences in functional brain architecture that affect network robustness and adaptability, providing insights into neural mechanisms of emotion regulation and potential applications for emotion-aware VR systems.

Decoding language from brain dynamics is an important open direction in the realm of brain-computer interface (BCI), especially considering the rapid growth of large language models. Compared to invasive-based signals which require electrode implantation surgery, non-invasive neural signals (e.g. EEG, MEG) have attracted increasing attention considering their safety and generality. However, the exploration is not adequate in three aspects: 1) previous methods mainly focus on EEG but none of the previous works address this problem on MEG with better signal quality; 2) prior works have predominantly used $``teacher-forcing"$ during generative decoding, which is impractical; 3) prior works are mostly $``BART-based"$ not fully auto-regressive, which performs better in other sequence tasks. In this paper, we explore the brain-to-text translation of MEG signals in a speech-decoding formation. Here we are the first to investigate a cross-attention-based ``whisper" model for generating text directly from MEG signals without teacher forcing. Our model achieves impressive BLEU-1 scores of 60.30 and 52.89 without pretraining $\&$ teacher-forcing on two major datasets ($\textit{GWilliams}$ and $\textit{Schoffelen}$). This paper conducts a comprehensive review to understand how speech decoding formation performs on the neural decoding tasks, including pretraining initialization, training $\&$ evaluation set splitting, augmentation, and scaling law. Code is available at https://github.com/NeuSpeech/NeuSpeech1$.

Intracranial hemorrhage (ICH) is among the most lethal forms of stroke, where timely and accurate diagnosis is critical for patient survival. CNN-based computer-aided diagnosis systems have been applied for ICH subtype detection and classification, but their performances remain insufficient for reliable clinical deployment yet. In this study, we propose a novel integration of Real-Time Detection Transformer (RT-DETR) into a reinforcement learning (RL) framework based on Proximal Policy Optimization (PPO). To enhance the diagnostic capability for ICH subtype detection and classification, an adaptive reward mechanism is also proposed that jointly optimizes detection and classification objectives through an actor-critic learning paradigm. Our experimental results on a public database demonstrate that our proposed RL enhanced RT-DETR outperforms the baseline RT-DETR and YOLO variants. In particular, RT-DETR-PPO achieved 6.7% and 6.0% gains in precision under the detection-oriented criterion (IoU>0.7) for the challenging ICH subtypes subarachnoid hemorrhage (SAH) and subdural hemorrhage (SDH), respectively. Under the classification-oriented criterion (IoU>0.1), RT-DETR-PPO achieved $5.0 \%$ and $4.2 \%$ gains in precision for SAH and SDH, respectively.

Dual-arm robots hold significant potential for performing medical assistive tasks in healthcare environments. However, executing such diverse and complex tasks requires advanced dual-arm robot intelligence, which faces substantial challenges due to multi-agent interactions in sequential long- horizon (LH) actions. This study introduces a novel multi-agent reinforcement learning approach, termed Counterfactual Multi-Agent Demo Augmented Policy Gradient (COMA-DAPG), to learn and perform LH medical assistive tasks for dual-arm robots. The proposed COMA-DAPG integrates a counterfactual critic network and demonstration-augmented policy gradient (DAPG) with three designed reward functions. Our experimental results demonstrate that COMA-DAPG outperforms each COMA and DAPG with over 25% improvement in average success rate across three LH tasks.Clinical Relevance—COMA-DAPG addresses key challenges in dual-arm robotics, such as credit assignment, gradient variance, and collision avoidance, to enable precise, cooperative execution of complex medical tasks, enhancing reliability and efficiency in clinical care settings.

Background Patients with schizophrenia experience accelerated aging, accompanied by abnormalities in biomarkers such as shorter telomere length. Brain age prediction using neuroimaging data has gained attention in schizophrenia research, with consistently reported increases in brain-predicted age difference (brain-PAD). However, its associations with clinical symptoms and illness duration remain unclear. Methods We developed brain age prediction models using structural magnetic resonance imaging (MRI) data from 10,938 healthy individuals. The models were validated on an independent test dataset comprising 79 healthy controls, 57 patients with recent-onset schizophrenia, and 71 patients with chronic schizophrenia. Group comparisons and the clinical associations of brain-PAD were analyzed using multiple linear regression. SHapley Additive exPlanations (SHAP) values estimated feature contributions to the model, and between-group differences in SHAP values and group-by-SHAP value interactions were also examined. Results Patients with recent-onset schizophrenia and chronic schizophrenia exhibited increased brain-PAD values of 1.2 and 0.9 years, respectively. Between-group differences in SHAP values were identified in the right lateral prefrontal area (false discovery rate [FDR] p = 0.022), with group-by-SHAP value interactions observed in the left prefrontal area (FDR p = 0.049). A negative association between brain-PAD and Full-scale Intelligence Quotient scores in chronic schizophrenia was noted, which did not remain significant after correction for multiple comparisons. Conclusions Brain-PAD increases were pronounced in the early phase of schizophrenia. Regional brain abnormalities contributing to brain-PAD likely vary with illness duration. Future longitudinal studies are required to overcome limitations related to sample size, heterogeneity, and the cross-sectional design of this study.

Virtual reality (VR) technologies can induce realistic emotions in controlled experimental settings, offering unprecedented opportunities to study how the human brain processes emotions under real-world conditions. The integration of VR experiences with electroencephalography (EEG) provides a promising potential for gaining novel insights into individual emotional states. However, the complex network dynamics underlying human emotions during VR experiences remain largely unexplored. To address this gap, we leveraged graph-theoretical approaches to investigate functional brain networks derived from EEG signals recorded during immersive VR experiments. We assessed key topological properties of functional brain networks across multiple frequency bands (delta, theta, alpha, beta, gamma, and high gamma) and compared network characteristics between different emotional states (negative, neutral, and positive). Furthermore, we evaluated whether these graph-based features could accurately distinguish between positive and negative emotions using machine learning approaches. Our findings revealed distinct network patterns associated with different emotional states. During negative emotional experiences, we observed two key neural signatures: increased high gamma band activity in the left central region and decreased theta band activity in the occipital region. Conversely, positive emotions were characterized by reduced activity across most frequency bands in the left frontal region. Our machine learning model achieved an average classification accuracy of 79% in differentiating positive and negative emotions using network features that combined graph-theoretical measures and connectivity weights across all frequency bands, with the high gamma band demonstrating particular importance for emotion processing. This study advances our understanding of how brain networks dynamically reorganize during VR-induced emotional experiences and establishes the potential of graph-based EEG features for robust emotion recognition, paving the way for personalized VR applications.

Brain-computer interfaces (BCIs) have the potential to revolutionize communication for individuals with severe disabilities. EEG-to-text models, which translate brain signals into written language, offer a promising avenue for restoring communication abilities. Recent advancements in machine learning have improved the accuracy and speed of these models, but their true capabilities remain unclear due to limitations in evaluation methodologies. This study critically examines the performance of EEG-to-text models, focusing on their ability to learn from EEG signals rather than simply memorizing patterns. We introduce a novel methodology that compares model performance on EEG data with that on noise inputs. Our findings reveal that many EEG-to-text models perform similarly or even better on noise, suggesting that they may be memorizing patterns rather than truly learning from EEG signals. These results highlight the need for more rigorous benchmarking and evaluation practices in the field of EEG-to-text translation. By addressing the limitations of current methodologies, we can develop more reliable and trustworthy systems that truly harness the potential of brain-computer interfaces for communication.

Introduction:Brain age prediction using neuroimaging and machine learning has emerged as a promising approach to assess brain health and detect deviations associated with neurological and psychiatric disorders. The difference between chronological age and predicted brain age, known as brain-predicted age difference (brainPAD), is considered a potential biomarker for advanced brain aging. However, most studies rely on single-modality imaging, limiting predictive accuracy and generalization. This study aimed to enhance brain age prediction by integrating multimodal neuroimaging-structural MRI (sMRI) and diffusion MRI-derived fractional anisotropy (FA)-and evaluating its effectiveness in both healthy individuals and schizophrenia patients. Methods:We analyzed a large, multi-site dataset of 2,558 healthy individuals (aged 12-88 years) using machine learning approaches to assess the impact of multimodal inputs on brain age prediction. A stacking model combining sMRI and FA features was developed and validated. To evaluate cross-dataset generalization, the model was tested on an independent dataset comprising 56 healthy individuals (aged 20-58 years) and 48 schizophrenia patients (aged 19-65 years). Statistical analyses were conducted to compare brainPAD scores between groups and assess correlations with clinical measures in schizophrenia patients. Results:The multimodal stacking model achieved superior prediction performance compared to single-modality models, with a mean absolute error (MAE) of 2.675 years and Pearson's correlation (r) of 0.970 between predicted and chronological age in the internal test set. External validation on the COBRE dataset demonstrated MAE of 4.556 years (r = 0.877) for healthy controls and 6.189 years (r = 0.873) for patients with schizophrenia. Schizophrenia patients exhibited significantly higher brainPAD scores compared to healthy controls (t = 3.857; p < 0.001; Cohen's d = 0.769), suggesting advanced brain aging. Additionally, brainPAD was significantly correlated with symptom severity scores in schizophrenia (ρ = 0.331-0.337, p < 0.05). Discussion:Our findings demonstrate that integrating sMRI and FA features improves brain age prediction accuracy and generalization. Furthermore, the correlation between brainPAD and clinical symptoms highlights its potential as a biomarker for disease progression and treatment monitoring. These results underscore the value of multimodal neuroimaging and machine learning in advancing psychiatric neuroimaging research and paving the way for clinical applications in schizophrenia and related disorders. Further investigation with larger sample sizes is required to validate and extend these findings.

Schizophrenia is a complex neuropsychiatric disorder characterized by significant heterogeneity, posing a challenge for accurate classification using neuroimaging data. Graph convolutional networks (GCNs) have emerged as a promising approach for leveraging the inherent graph structure of brain connectivity to discriminate between patients with schizophrenia and healthy controls. However, existing GCN-based methods often struggle to capture the subtle neuroimaging differences associated with the disorder. To address these limitations, we propose a novel GCN framework (MSN-GCN) that integrates morphometric similarity networks (MSN) derived from structural MRI scans. Our method involves constructing individual brain graphs based on multiple morphometric features, including cortical thickness, surface area, gray matter volume, mean curvature, and Gaussian curvature. These individual graphs are then combined into a population-level graph that incorporates both topological and phenotypic information. By employing a variational edge learning approach, our model adaptively optimizes the edge weights to capture the complex relationships between brain structure and schizophrenia. We evaluated our proposed method on a large, multi-site dataset comprising 377 schizophrenia patients and 590 healthy controls. Experimental results demonstrate superior classification performance compared to state-of-the-art methods, achieving a mean accuracy of 80.85%. Notably, the superior temporal gyrus emerged as a key region contributing to classification. Significant differences in the clustering coefficient of the superior temporal gyrus, postcentral gyrus, and lateral occipital cortex between patients and healthy controls, and their correlations with negative symptoms were detected in post-hoc analyses. These findings demonstrate the potential of MSN-GCN for accurate schizophrenia detection and provide valuable insights into the neural correlates of the disorder.

The human brain exhibits a complex organization into functional communities, with interconnected regions of interest (ROIs) playing a critical role in emotional processing. However, traditional transformer models for EEG-based emotion recognition often treat all ROIs equally, neglecting the crucial role of these communities. To address this limitation, we propose the brain network community-aware global-local transformer (BN-BrainTF) model. BN-BrainTF employs source localization to identify brain activity origins within functional communities derived from EEG data. The model then extracts local features specific to each community and global features capturing whole-brain context using a spectral-spatial attention module and a dynamical graph convolutional network based on functional connectivity. A global-local transformer with cross-attention integrates these features within each community, while a fusion transformer captures interactions between all communities. We evaluated BN-BrainTF on two benchmark datasets with distinct emotional classification paradigms. On the SEED dataset (three emotional states: positive, negative, and neutral), our model achieved 77.92% average accuracy across all subjects. For the SEED-IV dataset (four emotional states: happy, sad, fear, and neutral), BN-BrainTF achieved 59.41% average accuracy across all subjects. These results demonstrate the effectiveness of incorporating functional brain community structure for EEG-based emotion recognition, with consistent performance across different emotional classification tasks. The comprehensive representation of brain activity informed by functional communities provides superior emotion recognition performance compared to traditional approaches that ignore the underlying brain network organization.

Diffusion-based generative models have emerged as powerful tools for synthesizing high-fidelity images across diverse domains, yet their application to neuroimaging remains underexplored. We systematically evaluated two representative approaches, namely denoising diffusion probabilistic models (DDPM) and latent diffusion models (LDM), for generating synthetic brain MRI data using the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) dataset. Both approaches were validated through brain age prediction tasks, with performance assessed using mean absolute error (MAE) and Pearson's correlation coefficient (R). Models trained on real MRI achieved MAE of 6.26–6.80 years (R = 0.890–0.910), whereas DDPM-generated data degraded performance to MAE of 11.50–13.57 years (R = 0.866–0.871), and LDM-generated data performed worse with MAE of 15.87–18.19 years (R = 0.792–0.827). Both DDPM and LDM exhibited significant limitations in preserving brain morphological structures critical for accurate brain age prediction. These findings highlight fundamental challenges in adapting conventional diffusion models to neuroimaging applications and underscore the need for specialized architectures and training strategies tailored to the unique characteristics of brain MRI data.

Immersive technologies such as virtual reality and haptic systems create complex multisensory experiences that challenge traditional user experience (UX) evaluation methods. Conventional static questionnaires fail to capture individual variability and dynamic nature of user responses. Electroencephalography (EEG) offers an objective means to monitor attention, cognitive load, and emotional engagement; however, the complexity of interpreting EEG signals has hindered its practical adoption. To address this, we propose EUEQ-mLLM, a framework that integrates EEG data with multi-agent large language models (LLMs) to generate adaptive UX questionnaire. The framework employs a four-stage workflow in which specialized LLMs collaborate to analyze EEG patterns, extract experiential significance, and dynamically create personalized assessment instruments tailored to individual brain functional states during interaction. EUEQ-mLLM generates contextually relevant questionnaires responding to actual neurophysiological patterns rather than static assumptions. Evaluation using G-Eval demonstrated EUEQ-mLLM achieved 80.96, representing 45.32% improvement over single-stage baselines. Designed for small-scale language models, our approach demonstrates the potential of EEG-informed multi-agent architectures for precise, adaptive UX evaluation methodologies that evolve with modern interactive experience complexity.

Major depressive disorder (MDD) is a serious psychiatric disorder characterized by persistent feelings of sadness, hopelessness, and lack of interest or pleasure in daily activities. Yet, reliable diagnostic tools for this brain disorder remain lacking. Functional near-infrared spectroscopy (fNIRS), an optical brain imaging technique, offers a promising approach for monitoring cerebral hemodynamic activity associated with MDD. In this study, we propose a novel algorithm based on wavelet coherence and a state-pathology separation network (WCSN) to automatically detect MDD using a dual-channel fNIRS system. The fNIRS signals were first preprocessed and transformed into two-dimensional feature maps using a wavelet coherence method. Following this, a wrapped exhaustive search was applied to select the optimal subset of feature maps, which was then utilized to reconstruct the dataset. Finally, samples were classified using the state-pathology separation network that employed a dual-encoder convolutional autoencoder (DCoAE) module to separate the feature maps into state features and pathology features, while a Transformer module distinguished MDD patients from healthy controls based solely on pathology features. The WCSN algorithm achieved exceptional performance with an accuracy of 0.923 +/- 0.068 and a subject accuracy of 0.918 +/- 0.076. Our result highlights the WCSN algorithm's ability to isolate pure pathology features, enhancing classification robustness and generalizability under dual-channel data conditions. Taken together, the proposed WCSN algorithm is well-suited for home-based MDD screening applications.

Since sudden and recurrent epileptic seizures seriously affect people's lives, computer-aided automatic seizure detection is crucial for precise diagnosis and prompt treatment. A novel seizure detection algorithm named channel selection-based temporal convolutional network (CS-TCN) was proposed in this article. First, electroencephalogram (EEG) recordings were segmented into 2-s intervals and features were extracted from both the time and frequency domains. Then, the expanded fisher score channel selection method was employed to select channels that contribute the most to seizure detection. Finally, the features from selected EEG channels were fed into the TCN to capture inherent temporal dependencies of EEG signals and detect seizure events. Children Hospital Boston and Massachusetts Institute of Technology (CHB-MIT) and Siena datasets were used to verify the detection performance of the CS-TCN algorithm, achieving sensitivities of 98.56% and 98.88%, and specificities of 99.80% and 99.88% in samplewise analysis, respectively. In eventwise analysis, the algorithm achieved sensitivities of 97.57% and 95.00%, with delays of 6.91 and 18.62 s, and FDR/h of 0.11 and 0.39, respectively. These results surpassed state-of-the-art few-channel algorithms for both datasets. CS-TCN algorithm offers excellent performance while simplifying model complexity and computational requirements, thus showcasing its potential for facilitating seizure detection in home environments.

Social hierarchies are a common form of social organization across species. Although hierarchies are largely stable across time, animals may socially ascend or descend within hierarchies depending on environmental and social challenges. Here, we develop a novel paradigm to study social ascent and descent within male CD-1 mouse social hierarchies. We show that mice of all social ranks rapidly establish new stable social hierarchies when placed in novel social groups with animals of equivalent social status. Seventy minutes following social hierarchy formation, males that were socially dominant prior to being placed into new social hierarchies exhibit higher increases in plasma corticosterone and vastly greater transcriptional changes in the medial amygdala (MeA), which is central to the regulation of social behavior, compared to males who were socially subordinate prior to being placed into a new hierarchy. Specifically, the loss of social status in a new hierarchy (social descent) is associated with reductions in MeA expression of myelination and oligodendrocyte differentiation genes. Maintaining high social status is associated with high expression of genes related to cholinergic signaling in the MeA. Conversely, gaining social status in a new hierarchy (social ascent) is related to relatively few unique rapid changes in the MeA. We also identify novel genes associated with social transition that show common changes in expression when animals undergo either social descent or social ascent compared to maintaining their status. Two genes, Myosin binding protein C1 (Mybpc1) and μ-Crystallin (Crym), associated with vasoactive intestinal polypeptide (VIP) and thyroid hormone pathways respectively, are highly upregulated in socially transitioning individuals. Further, increases in genes associated with synaptic plasticity, excitatory glutamatergic signaling and learning and memory pathways were observed in transitioning animals suggesting that these processes may support rapid social status changes.

Machine learning (ML) techniques are increasingly used in brain age prediction to assess brain health and detect neurological and psychiatric disorders. The availability of large, publicly accessible imaging datasets has accelerated the adoption of ML-based methods. However, multi-site MRI datasets present challenges due to site effects, which can introduce biases and affect the accuracy of brain age prediction. In this study, we examined the influence of MRI data harmonization on brain age prediction by comparing models trained with and without harmonization across a large-scale, multi-site dataset of 10,938 healthy individuals aged 5 to 95 years. Using automated ML approaches, we trained various models and computed SHapley Additive exPlanations (SHAP) values to identify the key features driving brain age predictions. Our results showed that while a weighted ensemble method achieved high prediction accuracy (MAE = 7.013; R = 0.860), data harmonization reduced prediction performance, indicating that site-related variability contains valuable information influencing model predictions. SHAP analysis also revealed substantial site-specific biases impacting the predictions. These findings suggest the need to account for site-specific factors in multi-site MRI studies. Understanding the impact of site harmonization is crucial for developing robust and generalizable brain age prediction models that can be applied across diverse populations and imaging settings.

The integration of multisensory, particularly the fusion of visual and tactile inputs, plays a crucial role in human perception and environment interaction. Despite its importance, the neural mechanisms underlying visuo-haptic integration remain poorly understood. This study investigates the brain's functional connectivity during visuo-haptic tasks within virtual environments using electroencephalography (EEG) and graph theoretical analysis. We collected EEG data from participants engaged in tasks requiring simultaneous processing of visual and tactile stimuli. Functional connectivity was constructed from the EEG signals, and we analyzed key graph theoretical metrics to characterize the network's properties. Specifically, in the tactile-only condition, we observed decreased global efficiency and small-worldness in both beta and theta frequency bands, accompanied by an increase in characteristic path length. In addition, strength, local efficiency, and clustering coefficient were diminished in the tactile-only condition, compared to conditions with visual input. However, we found no significant differences in brain network topology across different tactile feedback modalities. These results provide novel insights into how the brain's dynamic coordination of multisensory inputs and establish a solid foundation for future investigations of visuo-haptic processing using graph theoretical approaches.

Because large brains are energetically expensive, they are associated with metabolic traits that facilitate energy availability across vertebrates. However, the biological underpinnings driving these traits are not known. Given its role in regulating host metabolism in disease studies, we hypothesized that the gut microbiome contributes to variation in normal cross-vertebrate species differences in metabolism, including those associated with the brain's energetic requirements. By inoculating germ-free mice with the gut microbiota (GM) of three primate species - two with relatively larger brains and one with a smaller brain - we demonstrated that the GM of larger-brained primates shifts host metabolism towards energy use and production, while that of smaller-brained primates stimulates energy storage in adipose tissues. Our findings establish a causal role of the GM in normal cross-host species differences in metabolism associated with relative brain size and suggest that the GM may have been an important facilitator of metabolic changes during human evolution that supported encephalization.

Proper monitoring of anesthesia stages can guarantee the safe performance of clinical surgeries. In this study, different anesthesia stages were classified using near-infrared spectroscopy (NIRS) signals with machine learning. The cerebral hemodynamic variables of right proximal oxyhemoglobin (HbO2) in maintenance (MNT), emergence (EM) and the consciousness (CON) stage were collected and then the differences between the three stages were compared by phase-amplitude coupling (PAC). Then combined with time-domain including linear (mean, standard deviation, max, min and range), nonlinear (sample entropy) and power in frequency-domain signal features, feature selection was performed and finally classification was performed by support vector machine (SVM) classifier. The results show that the PAC of the NIRS signal was gradually enhanced with the deepening of anesthesia level. A good three-classification accuracy of 69.27% was obtained, which exceeded the result of classification of any single category feature. These results indicate the feasibility of NIRS signals in performing three or even more anesthesia stage classifications, providing insight into the development of new anesthesia monitoring modalities.

Abstract Chimeric antigen receptor T-cell (CAR-T) therapy is a significant advancement in treating hematological malignancies, yet it faces challenges due to its variable therapeutic responses and the risk of severe toxicities. This study delves into the effects of tumor burden and CAR-T cell doses on the toxicity profile of the therapy, employing a PBMC-humanized mouse model engrafted with luciferase-labeled Raji B cell lymphoma (Raji-luc). By exploring high vs. low tumor burden and high vs. low CAR-T dose scenarios, we aim to gain insights into the dynamic relationship between these factors and CAR-T-induced toxicity and efficacy. In the first experiment, we established PBMC humanized mouse models with a high tumor burden and a low tumor burden. 10 days (high burden) or 7 days (low burden) after the Raji-luc inoculation, mice were engrafted with 10M human PBMCs and dosed with CD19 autologous CAR-Ts. In mice with low tumor burden, CAR-T treatment demonstrated significant efficacy, as evidenced by IVIS imaging, while the high tumor burden model exhibited reduced efficacy. CAR-T treatment in Raji-bearing mice did not induce weight loss in either high or low burden models. CD19 CAR-T cells effectively reduced the target cell population in peripheral blood, with greater expansion observed in the higher tumor burden model. We also evaluated human cytokine levels post-CAR-T treatment, revealing higher cytokine induction in the high tumor burden model, peaking at 2 days post-treatment. Secondly, in a CAR-T dose-response study, we treated PBMC-humanized Raji-bearing mice with 10M, 15M, and 20M CAR-T cells. Higher CAR-T cell doses (15M and 20M) resulted in observable toxicity measured by body weight loss, while the 10M dose did not. All CAR-T doses effectively slowed tumor growth and induced significant CAR-T cell expansion. Selected human cytokines, such as IL-5 and RANTES, demonstrated a dose-response correlation with CAR-T treatment. Additionally, IL-6 and IL-10 were significantly correlated with tumor burden rather than CAR-T doses. The differential cytokine responses observed in our study provide valuable insights into the utility of the PBMC-humanized mouse model for investigating the biological responses associated with CD19 CAR-T therapy. Our findings underscore the utility of the PBMC-humanized mouse model in assessing variability in toxicity and cytokine responses to CAR-T therapy. This model offers valuable insights into the factors influencing CAR-T treatment outcomes and provides a platform for planning more precise treatment and enhancing the safety and efficacy of CAR-T therapy. Citation Format: Won Lee, Destanie Rose, James G. Keck, Jiwon Yang. Assessing impacts of tumor burden and CAR-T cell dosage on the toxicity and efficacy profile of CD19 CAR-T therapy in a PBMC-humanized mouse model [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4010.