Lecanemab use in Chinese patients with Alzheimer's disease: a 12-month multicenter real-world study.

Abnormal manifestations in human conflict monitoring and processing may indicate pathological deficits, such as those observed in individuals with ADHD, autism spectrum disorder, or obsessive-compulsive disorder, which can be assessed using the Stroop task. In the Stroop task, incongruent stimuli evoke additional cognitive processes, leading to longer reaction delays or higher error rates compared to congruent or neutral stimuli. Similarly, the underlying cognitive states associated with these processes are reflected in neural activity, such as changes in hemoglobin concentration in the cortex. Employing functional near-infrared spectroscopy (fNIRS) to monitor hemodynamic responses provides valuable insights into these specific cognitive processes, while also supporting the development of computational models and artificial intelligence algorithms that emulate brain function. A key insight is that, drawing on the pretraining paradigm from natural language processing, researchers can perform self-supervised pretraining on large-scale unlabeled brain imaging data to learn generalizable functional and structural brain representations. The key to the success of large language models lies in pre-training on large-scale text data, which is completely unlabeled. Deep learning models (e.g., Transformers) can learn high-dimensional associations between texts—that is, word vectors—by learning from existing data. This idea applies equally well to neuroimaging data. As long as there is sufficient neural data of the same modality (or even different modalities, just as large language models handle different languages), we have reason to believe that large models can autonomously discover the intrinsic associations among neural activities, and may even construct generative digital brain models. Also, these representations can then be fine-tuned for specific downstream tasks—such as disease diagnosis or cognitive state decoding—thereby addressing the core challenges in neuroscience of small sample sizes, high dimensionality, and low signal-to-noise ratios. Moreover, within-subject decoding typically demands a large number of samples, while the slow-varying nature of hemodynamic responses necessitates slow, block-designed experiments. This trade-off prolongs scanning duration and increases data acquisition costs. In addition, repeated exposure to the same task may introduce practice effects, potentially confounding the interpretation of cognitive decline due to pathology or aging. Consequently, behavioral tests alone are insufficient for reliably distinguishing cognitive states. These challenges highlight the critical need for a multi-day, slow block-designed fNIRS dataset tailored for cognitive state decoding. This dataset contains cortex hemodynamic responses recorded via fNIRS from 55 young adults while performing the Stroop task, across three sessions spaced about one week apart. Technical validation confirms its effectiveness in capturing individual cognitive patterns during conflict processing. Traditional univariate methods, such as the general linear model (GLM), may be suboptimal for this dataset, as they do not account for spatial activation patterns in hemodynamic responses. In contrast, multivariate pattern analysis (MVPA) is well-suited for multi-channel fNIRS signals with high spatial resolution. The dataset holds potential for various applications, including: