Seminar Papers

[Article] Beyond Brain Decoding: Visual-Semantic Reconstructions to Mental Creation Extension Based on fMRI

Summary: In this paper, by integrating LLM with fMRI through its Brain Expert Adaption and Prompt Variant Alignment modules, enables robust cross-subject decoding and sophisticated multimodal ‘mental creation’ capabilities beyond mere visual reconstruction.

Jing, Haodong, et al. “Beyond brain decoding: Visual-semantic reconstructions to mental creation extension based on fmri.” Proceedings of the IEEE/CVF International Conference on Computer Vision. 2025.

[Article] Neuro-Vision to Language: Enhancing Brain Recording-based Visual Reconstruction and Language Interaction

Summary: This paper introduces a framework that uses a Vision Transformer 3D (ViT3D) to preserve the 3D structure of brain data for better visual decoding. It eliminates the need for subject-specific models, allowing high-quality reconstruction from a single experimental trial across different people. By integrating with Large Language Models (LLMs), the system can also perform tasks like brain captioning and complex reasoning using natural language.

Shen, Guobin, et al. “Neuro-vision to language: Enhancing brain recording-based visual reconstruction and language interaction.” Advances in Neural Information Processing Systems 37 (2024): 98083-98110.

[Article] Modulatory dynamics of periodic and aperiodic activity in respiration-brain coupling

Summary: This study utilized resting-state MEG and EEG data from 78 participants to investigate the previously unstudied link between respiration and aperiodic neural activity, specifically the 1/f slope which serves as a marker for the excitation-inhibition (E:I) balance. The authors discovered that fluctuations in the aperiodic slope are significantly phase-locked to the respiratory cycle, with flatter slopes (indicating higher excitation) occurring during inspiration and steeper slopes (indicating higher inhibition) during expiration, particularly over parieto-occipital regions. These findings demonstrate that respiration acts as a physiological modulator of spontaneous brain state shifts and that this coupling to aperiodic activity follows distinct temporal dynamics compared to periodic oscillatory activity.

Kluger, Daniel S., et al. “Modulatory dynamics of periodic and aperiodic activity in respiration-brain coupling.” Nature communications 14.1 (2023): 4699.

[Article] Unlocking the potential of wearable technology: Fitbit-derived measures for predicting ADHD in adolescents & Machine Learning–Based Prediction of Attention-Deficit/Hyperactivity Disorder and Sleep Problems With Wearable Data in Children & Using explainable machine learning and fitbit data to investigate predictors of adolescent obesity

Unlocking the potential of wearable technology: Fitbit-derived measures for predicting ADHD in adolescents

Summary: This study examined whether Fitbit-derived physical activity data could enhance the identification of ADHD in adolescents, addressing the limitations of traditional, subjective diagnostic methods. Using ABCD Study data (release 5.0), researchers analyzed Fitbit daily and weekly activity summaries—such as sedentary time, resting heart rate, and energy expenditure—and tested their predictive value through correlation analyses, logistic regression models, and multiple machine learning classifiers. Several Fitbit features were significantly associated with ADHD diagnosis, and the Random Forest model achieved the highest predictive performance (AUC = 0.95, accuracy = 0.89).

Rahman, Muhammad Mahbubur. “Unlocking the potential of wearable technology: Fitbit-derived measures for predicting ADHD in adolescents.” Frontiers in Child and Adolescent Psychiatry 4 (2025): 1504323.

Machine Learning–Based Prediction of Attention-Deficit/Hyperactivity Disorder and Sleep Problems With Wearable Data in Children

Summary: This study used wearable-device data and K-SADS diagnostic results from the ABCD cohort to develop machine-learning models that detect ADHD and sleep problems in children using daily-life digital phenotypes. Researchers extracted 64 circadian-rhythm features using cosinor analysis and physical activity metrics from 21 days of wearable data. The ML models achieved reasonable performance for ADHD (AUC = 0.798) and sleep problems (AUC = 0.737), with especially high negative predictive values.

Kim, Won-Pyo, et al. “Machine learning–based prediction of attention-deficit/hyperactivity disorder and sleep problems with wearable data in children.” JAMA network open 6.3 (2023): e233502-e233502.

Using explainable machine learning and fitbit data to investigate predictors of adolescent obesity

Summary: This study investigated whether Fitbit-derived sleep, activity, and cardiovascular fitness measures (combined with sociodemographic factors) can predict obesity risk in early adolescence, a period when long-term health trajectories are shaped. Using data from 2,971 adolescents in the ABCD Study, researchers extracted mean and variability-based features from two weeks of Fitbit recordings and applied glass-box machine learning models to identify key predictors of obesity. The models highlighted several strong risk indicators, including non-White race, low household income, later bedtimes, shorter and more irregular sleep, low step counts, and elevated heart rates (AUC ≈ 0.73).

Kiss, Orsolya, et al. “Using explainable machine learning and fitbit data to investigate predictors of adolescent obesity.” Scientific Reports 14.1 (2024): 12563.

[Article] Autonomic physiological coupling of the global fMRI signal

Summary: This study characterizes a widespread pattern of co-fluctuations between global fMRI brain signals and sympathetic-mediated physiological fluctuations in humans. This pattern of cofluctuations is replicated across multiple independent datasets of multimodal fMRI, EEG, and peripheral physiology recordings. It is widespread across the body and entire nervous system, involving the brain, heart, lungs, exocrine and smooth muscle systems. It is also linked to changes in arousal state induced via deep breathing and intermittent sensory stimulation, as well as spontaneous fluctuations in arousal observed during sleep. Here, they show that the spatial structure of global fMRI signals is maintained under experimental suppression of end-tidal carbon dioxide variations, suggesting that respiratory-driven fluctuations in arterial CO2 accompanying arousal cannot fully explain the origin of these signals in the brain. These findings suggest that the global fMRI signal is a substantial component of the arousal response governed by the autonomic nervous system. Bolt, Taylor, et al. “Autonomic physiological coupling of the global fMRI signal.” Nature Neuroscience (2025): 1-9.

[Article] Estimation of respiratory rate and effort from a chest-worn accelerometer using constrained and recursive principal component analysis & Accelerometer-based estimation of respiratory rate using principal component analysis and autocorrelation

Estimation of respiratory rate and effort from a chest-worn accelerometer using constrained and recursive principal component analysis

Summary: This study proposes a constrained and recursive PCA method to robustly estimate respiratory effort and respiratory rate from a chest-worn 3-axis accelerometer under realistic sleeping conditions with varying sensor positions and body postures. The accelerometer signal is projected onto a gravity-based horizontal plane, and recursive PCA combined with STFT and a quality index is applied to extract respiratory effort and estimate respiratory rate, significantly reducing estimation error compared to conventional block-wise, unconstrained PCA. The method demonstrated a clear trade-off between coverage and accuracy, achieving agreement intervals below 1.5 breaths/min for high coverage settings and below 0.2 breaths/min for low coverage settings, indicating high robustness and flexibility for clinical and wearable applications.

Schipper, Fons, et al. “Estimation of respiratory rate and effort from a chest-worn accelerometer using constrained and recursive principal component analysis.” Physiological Measurement 42.4 (2021): 045004.

Accelerometer-based estimation of respiratory rate using principal component analysis and autocorrelation

Summary: This study proposed a novel PCA–autocorrelation method for estimating respiratory rate (RR) using a tri-axial accelerometer placed on the abdomen and validated it against a reference flow meter. Results from 25 healthy participants showed a very strong correlation (r = 0.99) and narrow limits of agreement (±1.9 bpm), outperforming single-axis approaches. The method demonstrates a low-cost, non-intrusive solution for continuous RR monitoring, with potential for future clinical validation.

Hostrup, Mads CF, et al. “Accelerometer-based estimation of respiratory rate using principal component analysis and autocorrelation.” Physiological Measurement 46.3 (2025): 035005.

[Article] An open resource for transdiagnostic research in pediatric mental health and learning disorders

Summary: Introduces the Healthy Brain Network (HBN), a large open biobank (target n≈10k) with deep phenotyping and multimodal data (EEG, MRI, eye-tracking, voice/video, genetics), enabling transdiagnostic pediatric mental-health research.

Alexander, Lindsay M., et al. “An open resource for transdiagnostic research in pediatric mental health and learning disorders.” Scientific data 4.1 (2017): 1-26.

[Article] MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-Text Decoding & Generative language reconstruction from brain recordings & CorText-AMA: brain-language fusion as a new tool for probing visually evoked brain responses

MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-Text Decoding

Summary: In this seminar, we will explore a novel fMRI-to-text decoding framework named MindLLM. It combines a neuroscience-informed, subject-agnostic fMRI encoder with an off-the-shelf large language model to translate brain activity into coherent text. It introduces Brain Instruction Tuning (BIT), which enriches the model’s capacity to extract and represent diverse semantic information from fMRI signals, enabling versatile decoding across different tasks and subjects. We will discuss how to implement these techniques in our study.

Qiu, Weikang, et al. “MindLLM: A Subject-Agnostic and Versatile Model for fMRI-to-Text Decoding.” arXiv preprint arXiv:2502.15786 (2025).

Generative language reconstruction from brain recordings

Summary: Using non-invasive fMRI, the authors map brain-decoded semantic representations into a large language model so it can autoregressively generate text aligned with the perceived stimulus—achieving direct brain-to-language generation without pre-constructed candidates and better alignment than selection-based baselines.

Ye, Ziyi, et al. “Generative language reconstruction from brain recordings.” Communications Biology 8.1 (2025): 346.

CorText-AMA: brain-language fusion as a new tool for probing visually evoked brain responses

Summary: CorText-AMA introduces an end-to-end brain–language framework that fuses fMRI signals with a large language model to caption and answer questions about natural scenes via an interactive chat interface, enabling targeted probing of what visual-cortex activity encodes and outperforming control models using functional alignment.

Bosch, Victoria, et al. “CorText-AMA: brain-language fusion as a new tool for probing visually evoked brain responses.”

[Article] TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction

Summary: In this paper, they made a step towards an integrative model of the brain during naturalistic perception by training an encoding model on an unprecedently-large fMRI dataset of participants watching videos. Importantly, their model is the first encoding pipeline which is simultaneously nonlinear, multisubject and multimodal d’Ascoli, S., Rapin, J., Benchetrit, Y., Banville, H., & King, J. R. (2025). TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction. arXiv preprint arXiv:2507.22229.

[Article] TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction

Summary: In this paper, they made a step towards an integrative model of the brain during naturalistic perception by training an encoding model on an unprecedently-large fMRI dataset of participants watching videos. Importantly, their model is the first encoding pipeline which is simultaneously nonlinear, multisubject and multimodal TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction