Fujian Star-Net Communication Co., Ltd. has recently achieved a significant advancement in EEG-based artificial intelligence and multimodal intelligent analysis through its subsidiary, Fujian Xingqilingzhi Intelligent Technology Co., Ltd., in collaboration with Professor Chen Jintai's ML4H (Machine Learning for Healthcare) team at the Hong Kong University of Science and Technology (Guangzhou).

The collaborative research has resulted in EEGBind, a novel multimodal AI method for EEG analysis, which has been accepted for presentation at ACM Multimedia 2026. This prestigious conference, hosted by the Association for Computing Machinery (ACM), stands as a leading international academic forum in the computer multimedia field, with a long-standing focus on frontier research areas including multimedia understanding, multimodal learning, and artificial intelligence. ACM MM is also recognized as a Class A international academic conference in computer graphics and multimedia by the China Computer Federation (CCF).

The current research centers on analyzing regional characteristics of EEG signals. By developing a multimodal learning framework with EEG as its core, the team is investigating how AI can extract more distinctive spatial and temporal features from complex brainwave data, while enhancing the model's analytical stability in complex information environments.

For ordinary people, EEG signals may appear as nothing more than intricate "wavy lines," yet for artificial intelligence, these continuously varying waveforms contain rich temporal, spatial, and signal morphology information. Traditional AI methods can help determine whether abnormal activities exist within a given EEG segment. This study, however, aims to go further by enabling AI to identify which source region characteristics these signals most closely resemble from complex waveform variations. The research classifies EEG signals according to five distinct source regions: generalized whole-brain patterns, frontal, temporal, occipital, and central-parietal areas. This means the model must not only "see" the EEG signals but also extract key features from short-term, complex, and individually variable EEG data that reflect spatial distribution and signal morphology differences.

In traditional acquisition processes, EEG signals are typically recorded alongside synchronized video. This video provides contextual background—such as patient movements, postures, and states—that aids in determining the sources of abnormal EEG activities. The research team therefore sought to enable AI to jointly analyze both EEG and video information. They developed an EEG-centric multimodal binding framework that treats EEG as the primary information source while incorporating synchronized video as auxiliary context. The model first extracts core representations from the EEG data, then integrates synchronized video information as corroborative evidence through multimodal binding, ultimately performing classification across the five IED source regions. The entire system operates in three key stages: fusion of EEG with auxiliary information, repair mechanisms when information is disrupted, and final judgment through multiple collaborative models.

When auxiliary information is incomplete, AI systems that rely too heavily on video data may produce compromised results. To address this challenge, EEGBind incorporates a View-Consistent Repair mechanism. During the research process, the team deliberately discards or masks portions of the video information, training the model to maintain stable EEG representations and classification outcomes even when auxiliary information fluctuates, thereby ensuring the AI remains reliable under conditions of incomplete data.

To continuously improve the accuracy of the AI, the research team employed Weighted-F1 as the primary evaluation metric for comparing different methods. Starting from a baseline of 0.6444 using EEG alone, the introduction of EEG-centric multimodal binding improved results to 0.8019, which further advanced to 0.8282 after incorporating view-consistency repair mechanisms. Ultimately, the five-fold model achieved a Weighted-F1 score of 0.8395 on the official evaluation of the NeuroMM 2026 Grand Challenge Track 3 (NMM-Source-IED), ranking first among all participating methods in this track. It is worth noting that this task does not simply determine whether abnormal EEG activity exists, but rather identifies, from candidate segments already confirmed to contain IEDs, which of the five source regions—whole-brain, frontal, temporal, occipital, or central-parietal—each segment belongs to.

In recent years, artificial intelligence has been transitioning from specialized models trained for single tasks toward foundation models and multimodal large models capable of processing increasingly complex information. For EEG analysis, the true value lies not merely in enabling models to "read a waveform," but in allowing AI to simultaneously understand temporal dynamics, spatial distributions, physiological signals, and cross-modal relationships. The exploration of EEGBind represents a meaningful step in this direction: using EEG as core information, introducing synchronized video as supplementary data through multimodal binding, and enabling models to establish associations between different information types—determining which information should be core, which is suitable as auxiliary, and whether AI can maintain stability when information is incomplete. This is precisely the core challenge that EEGBind aims to address.

The paper is available at: https://arxiv.org/pdf/2609.09728v1

This research marks a significant technical accumulation for Xingqilingzhi in the field of EEG-based artificial intelligence and multimodal intelligent analysis. As an AI company dedicated to emotional mental health and sleep health, Xingqilingzhi consistently focuses on the integration of EEG, physiological signals, and artificial intelligence technologies, exploring how complex physiological signals can be better collected, analyzed, and understood. Looking ahead, Fujian Star-Net Communication Co., Ltd. will continue to deepen independent innovation in core areas while strengthening collaborative innovation with universities and research teams, promoting the application of AI technology across more fields, extending AI capabilities to a broader range of industry scenarios, and leveraging the power of AI perception and understanding to support the development of the digital economy.