Pan Liao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Pan Liao has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Jia-Hong Gao
Center for MRI Research, Academy for Advanced Interdisciplinary Studies, Peking University
3 shared publications - Wei Xu
Center for MRI Research, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China
2 shared publications Bingjiang Lyu
Changping Laboratory
1 shared publication- Bingjiang Lyu
Changping Laboratory, Beijing, China
1 shared publication Congcong Li
Changping Laboratory
1 shared publication- Cunxin Lin
Capital Medical University, Headache Center, Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China, Beijing, 100071, CHINA
1 shared publication
Latest Publications
- Top Cited PapersOpen Access
- 13 January 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
A unified time-frequency foundation model for sleep decoding
Sleep decoding is key to revealing sleep architecture and its links to health, yet prevailing deep-learning models rely on supervised, task-specific designs and dual encoders that isolate time-domain and frequency-domain information, limiting generalizability and scalability. We introduce SleepGPT for sleep decoding, a time-frequency foundation model based on generative pretrained transformer, developed using multi-pretext pretraining strategy on 86,335 hours of polysomnography (PSG) from 8,377 subjects. SleepGPT includes a channel-adaptive mechanism for variable channel configurations and a unified time-frequency fusion module that enables deep cross-domain interaction. Evaluations across diverse PSG datasets demonstrate that SleepGPT sets a new benchmark for sleep decoding tasks, achieving superior performance in sleep staging, sleep-related pathology classification, sleep data generation, and sleep spindle detection. Moreover, it reveals channel- and stage-specific physiological patterns underlying sleep decoding. In sum, SleepGPT is an all-in-one method with exceptional generalizability and scalability, offering transformative potential in addressing sleep decoding challenges.
- 13 January 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Top Cited PapersOpen AccessA unified time-frequency foundation model for sleep decoding
Sleep decoding is key to revealing sleep architecture and its links to health, yet prevailing deep-learning models rely on supervised, task-specific designs and dual encoders that isolate time-domain and frequency-domain information, limiting generalizability and scalability. We introduce SleepGPT for sleep decoding, a time-frequency foundation model based on generative pretrained transformer, developed using multi-pretext pretraining strategy on 86,335 hours of polysomnography (PSG) from 8,377 subjects. SleepGPT includes a channel-adaptive mechanism for variable channel configurations and a unified time-frequency fusion module that enables deep cross-domain interaction. Evaluations across diverse PSG datasets demonstrate that SleepGPT sets a new benchmark for sleep decoding tasks, achieving superior performance in sleep staging, sleep-related pathology classification, sleep data generation, and sleep spindle detection. Moreover, it reveals channel- and stage-specific physiological patterns underlying sleep decoding. In sum, SleepGPT is an all-in-one method with exceptional generalizability and scalability, offering transformative potential in addressing sleep decoding challenges.
- 8 January 2026
- Published by IOP Publishing in Journal of Neural Engineering
- Vol. 23 (1) , 016005
A pretrained foundation model for headache disorders based on magnetoencephalography
Objective. Foundation models have demonstrated transformative potential in medical artificial intelligence but remain underexplored in functional neuroimaging, particularly magnetoencephalography (MEG). This study aims to develop a domain-specific, self-supervised MEG clinical foundation model tailored for headache disorders to address the challenges of high-dimensional data and limited labeled datasets in clinical research. Approach . We developed a transformer-based model pretrained on a large-scale dataset comprising multi-state MEG recordings (resting-state, auditory, and somatosensory stimulation) from 416 participants (362 headache patients and 54 healthy controls). The model utilized a self-supervised masked-signal reconstruction strategy to learn latent spatiotemporal representations of neural activity. We evaluated the model’s performance through signal reconstruction, visualization of attention weights, and downstream classification tasks comparing model-derived features against original MEG signals for migraine diagnosis. Main results . The pretrained model successfully reconstructed both continuous MEG signals and stimulus-specific evoked responses, effectively capturing intrinsic spatiotemporal brain dynamics. Visualization of the model’s attention weights demonstrated spatial alignment with corresponding sensory brain regions, confirming its neurophysiological interpretability. Furthermore, classifiers trained on features extracted from the pretrained model significantly outperformed those using original MEG signals in identifying migraine patients, revealing distinct neural response patterns. Significance . This study introduces a scalable, data-efficient framework for clinical MEG analysis that significantly reduces reliance on manual feature extraction and labeled data. It demonstrates the efficacy of foundation models in decoding complex neural dynamics, offering promising implications for understanding neuropathology and facilitating precision diagnostics in neurology.
- 8 January 2026
- Published by IOP Publishing in Journal of Neural Engineering
- Vol. 23 (1) , 016005
A pretrained foundation model for headache disorders based on magnetoencephalography
Objective. Foundation models have demonstrated transformative potential in medical artificial intelligence but remain underexplored in functional neuroimaging, particularly magnetoencephalography (MEG). This study aims to develop a domain-specific, self-supervised MEG clinical foundation model tailored for headache disorders to address the challenges of high-dimensional data and limited labeled datasets in clinical research. Approach . We developed a transformer-based model pretrained on a large-scale dataset comprising multi-state MEG recordings (resting-state, auditory, and somatosensory stimulation) from 416 participants (362 headache patients and 54 healthy controls). The model utilized a self-supervised masked-signal reconstruction strategy to learn latent spatiotemporal representations of neural activity. We evaluated the model’s performance through signal reconstruction, visualization of attention weights, and downstream classification tasks comparing model-derived features against original MEG signals for migraine diagnosis. Main results . The pretrained model successfully reconstructed both continuous MEG signals and stimulus-specific evoked responses, effectively capturing intrinsic spatiotemporal brain dynamics. Visualization of the model’s attention weights demonstrated spatial alignment with corresponding sensory brain regions, confirming its neurophysiological interpretability. Furthermore, classifiers trained on features extracted from the pretrained model significantly outperformed those using original MEG signals in identifying migraine patients, revealing distinct neural response patterns. Significance . This study introduces a scalable, data-efficient framework for clinical MEG analysis that significantly reduces reliance on manual feature extraction and labeled data. It demonstrates the efficacy of foundation models in decoding complex neural dynamics, offering promising implications for understanding neuropathology and facilitating precision diagnostics in neurology.
- Open Access
- 2 October 2024
- Published by Society for Neuroscience in The Journal of Neuroscience
- Vol. 44 (47)
Decoding the Temporal Structures and Interactions of Multiple Face Dimensions Using Optically Pumped Magnetometer Magnetoencephalography (OPM-MEG)
Humans possess a remarkable ability to rapidly access diverse information from others’ faces with just a brief glance, which is crucial for intricate social interactions. While previous studies using event-related potentials/fields have explored various face dimensions during this process, the interplay between these dimensions remains unclear. Here, by applying multivariate decoding analysis to neural signals recorded with optically pumped magnetometer magnetoencephalography, we systematically investigated the temporal interactions between invariant and variable aspects of face stimuli, including race, gender, age, and expression. First, our analysis revealed unique temporal structures for each face dimension with high test–retest reliability. Notably, expression and race exhibited a dominant and stably maintained temporal structure according to temporal generalization analysis. Further exploration into the mutual interactions among face dimensions uncovered age effects on gender and race, as well as expression effects on race, during the early stage (∼200–300 ms postface presentation). Additionally, we observed a relatively late effect of race on gender representation, peaking ∼350 ms after the stimulus onset. Taken together, our findings provide novel insights into the neural dynamics underlying the multidimensional aspects of face perception and illuminate the promising future of utilizing OPM-MEG for exploring higher-level human cognition.
- 2 October 2024
- Published by Society for Neuroscience in The Journal of Neuroscience
- Vol. 44 (47)
Open AccessDecoding the Temporal Structures and Interactions of Multiple Face Dimensions Using Optically Pumped Magnetometer Magnetoencephalography (OPM-MEG)
Humans possess a remarkable ability to rapidly access diverse information from others’ faces with just a brief glance, which is crucial for intricate social interactions. While previous studies using event-related potentials/fields have explored various face dimensions during this process, the interplay between these dimensions remains unclear. Here, by applying multivariate decoding analysis to neural signals recorded with optically pumped magnetometer magnetoencephalography, we systematically investigated the temporal interactions between invariant and variable aspects of face stimuli, including race, gender, age, and expression. First, our analysis revealed unique temporal structures for each face dimension with high test–retest reliability. Notably, expression and race exhibited a dominant and stably maintained temporal structure according to temporal generalization analysis. Further exploration into the mutual interactions among face dimensions uncovered age effects on gender and race, as well as expression effects on race, during the early stage (∼200–300 ms postface presentation). Additionally, we observed a relatively late effect of race on gender representation, peaking ∼350 ms after the stimulus onset. Taken together, our findings provide novel insights into the neural dynamics underlying the multidimensional aspects of face perception and illuminate the promising future of utilizing OPM-MEG for exploring higher-level human cognition.