A unified time-frequency foundation model for sleep decoding.

SleepGPT was pretrained on a large-scale polysomnography dataset comprising 86,335 hours of recordings from 8,377 subjects using a multi-pretext pretraining strategy.

• The model is based on a generative pretrained transformer architecture adapted for sleep decoding.
• Pretraining used polysomnography (PSG) data, which typically includes multiple physiological channels recorded during sleep.
• A multi-pretext pretraining strategy was employed to enable the model to learn diverse representations from the data.
• The scale of pretraining data (86,335 hours from 8,377 subjects) represents a large foundation for a sleep-specific model.

SleepGPT incorporates a unified time-frequency fusion module that enables deep cross-domain interaction between time-domain and frequency-domain information.

• Prior prevailing deep-learning models used dual encoders that isolate time-domain and frequency-domain information, limiting generalizability.
• The unified time-frequency fusion module contrasts with the dual-encoder approach by enabling joint processing of both domains.
• This architectural choice was designed to overcome limitations of task-specific designs in existing models.
• The module is described as enabling 'deep cross-domain interaction' rather than parallel but separate processing.

SleepGPT includes a channel-adaptive mechanism that accommodates variable channel configurations across different PSG datasets.

• Different PSG datasets may record different numbers and types of physiological channels (e.g., EEG, EOG, EMG).
• The channel-adaptive mechanism allows the model to handle these varying configurations without requiring dataset-specific retraining.
• This feature contributes to the model's generalizability across diverse PSG datasets.
• The mechanism is presented as a key architectural innovation enabling the 'all-in-one' nature of SleepGPT.

SleepGPT achieved superior performance in sleep staging compared to existing approaches across diverse PSG datasets.

• Evaluations were conducted across diverse PSG datasets to assess generalizability.
• Sleep staging involves classifying sleep into standard stages (e.g., Wake, N1, N2, N3, REM).
• SleepGPT is described as setting 'a new benchmark' for sleep staging performance.
• The model outperformed supervised, task-specific deep-learning models in this task.

SleepGPT demonstrated superior performance in sleep-related pathology classification.

• Sleep-related pathology classification is one of the four evaluated sleep decoding tasks.
• The model's performance on pathology classification was part of its benchmark-setting evaluation across diverse datasets.
• This task tests the model's ability to generalize beyond sleep staging to clinically relevant disorder identification.
• The model's unified architecture enabled strong performance on this task without task-specific design modifications.

SleepGPT achieved superior performance in sleep data generation as evaluated across diverse PSG datasets.

• Sleep data generation is identified as one of the four primary sleep decoding tasks evaluated.
• The generative pretrained transformer architecture is particularly suited to data generation tasks.
• Strong generation performance suggests the model learned meaningful representations of sleep physiology.
• This capability is noted as part of the model's 'all-in-one' functionality.

SleepGPT demonstrated superior performance in sleep spindle detection.

• Sleep spindle detection is a fine-grained event detection task within sleep recordings.
• Sleep spindles are brief bursts of oscillatory neural activity occurring during NREM sleep with relevance to memory and health.
• The model's ability to detect spindles alongside other tasks demonstrates its versatility as a foundation model.
• This is one of the four main sleep decoding tasks on which SleepGPT set a new benchmark.

SleepGPT revealed channel- and stage-specific physiological patterns underlying sleep decoding.

• The model's analysis uncovered patterns specific to individual recording channels (e.g., specific EEG leads).
• Stage-specific patterns were identified, meaning different sleep stages have distinct physiological signatures captured by the model.
• These findings provide interpretability insights into what the model learned during pretraining.
• The discovery of these patterns is presented as an additional scientific contribution beyond benchmark performance.

Prevailing deep-learning models for sleep decoding rely on supervised, task-specific designs and dual encoders that isolate time-domain and frequency-domain information, limiting generalizability and scalability.

• This limitation motivated the development of SleepGPT as a foundation model approach.
• Dual encoders process time-domain and frequency-domain information separately rather than jointly.
• Task-specific designs require separate models or significant modification for different sleep decoding tasks.
• The supervised-only approach limits the ability to leverage large unlabeled PSG datasets.