Santos-Mayo, A. | bioRxiv Science

bioRxiv2025

Concept2Brain: An AI model for predicting subject-level neurophysiological responses to text and pictures

The current growth of artificial intelligence (AI) tools provides an unprecedented opportunity to extract deeper insights from neurophysiological data while also enabling the reproduction and prediction of brain responses to a wide range of events and situations. Here, we introduce the Concept2Brain model, a deep network architecture designed to generate synthetic electrophysiological responses to semantic/emotional information conveyed through pictures or text. Leveraging AI solutions like CLIP from OpenAI, the model generates a representation of pictorial or language input and maps it into an electrophysiological latent space. We demonstrate that this openly available resource generates synthetic neural responses that closely resemble those observed in studies of naturalistic scene perception. The Concept2Brain model is provided as a web service tool for creating open and reproducible EEG datasets, allowing users to predict brain responses to any semantic concept or picture. Beyond its applied functionality, it also paves the way for AI-driven modeling of brain activity, offering new possibilities for studying how the brain represents the world.

neuroscience↗

bioRxiv2024

Decoding in the fourth dimension: Classification of temporal patterns and their generalization across locations

Neuroscience research has increasingly used decoding techniques, in which multivariate statistical methods identify patterns in neural data that allow the classification of experimental conditions or participant groups. Typically, the features used for decoding are spatial in nature, including voxel patterns and electrode locations. However, the strength of many neurophysiological recording techniques such as electroencephalography or magnetoencephalography is in their rich temporal, rather than spatial, content. The present report proposes a new decoding method that relies on the time information contained in neural time series. This information is then used in a subsequent step, generalization across location (GAL), which characterizes the relationship between sensor locations based on their ability to cross-decode. Two datasets are used to demonstrate usage of this method, referred to as time-GAL, involving (1) event-related potentials in response to affective pictures and (2) steady-state visual evoked potentials in response to aversively conditioned grating stimuli. In both cases, experimental conditions were successfully decoded based on the temporal features contained in the neural time series. Cross-decoding occurred in regions known to be involved in visual and affective processing. We conclude that the time-GAL approach holds promise for analyzing neural time series from a wide range of paradigms and measurement domains providing an assumption-free method to quantifying differences in temporal patterns of neural information processing and whether these patterns are shared across sensor locations. Author summaryDecoding and classification approaches are widely used in computational biology. In the field of neuroscience, pattern classification approaches typically use spatial information. In many instances, neural time series however are best defined by the their temporal, rather than spatial, features. Here, we propose a novel decoding approach taking advantage of the temporal information. Specifically, we utilize the waveform of neural time series as features for decoding experimental conditions and quantify each decoders generalization across locations (GAL) in multi-channel recordings. We illustrate the usage of our open source toolbox using two datasets, showing the sensitivity of the method to systematic condition differences in temporal dynamics along with its ability to capture and quantify spatial dependencies between recording locations.

neuroscience↗