SleepFM-Clinical

🔥 News

• Our paper is out on bioarxiv.

📖 Introduction

Sleep is a fundamental biological process with broad implications for physical and mental health, yet its complex relationship with disease remains poorly understood. Polysomnography (PSG), the gold standard for sleep analysis, captures rich physiological signals but remains underutilized due to challenges in standardization, generalizability, and multimodal signal integration. To address this, we curated over 585,000 hours of PSG data from approximately 65,000 participants across multiple cohorts and developed SleepFM, a multimodal sleep foundation model trained with a novel contrastive learning approach that accommodates any PSG montage. SleepFM produces sleep embeddings that enable accurate prediction of future disease risk. We demonstrate that SleepFM achieves a C-Index and AUROC of at least 0.75 (Bonferroni-corrected p < 0.01) for 130 conditions, including death (C-Index: 0.84), heart failure (0.80), chronic kidney disease (0.79), dementia (0.85), stroke (0.78), atrial fibrillation (0.78), and myocardial infarction (0.81). The model generalizes well to the Sleep Heart Health Study (SHHS), a dataset excluded from pretraining, where it achieves strong transfer learning performance. In addition, SleepFM performs competitively on traditional sleep analysis tasks, including sleep staging (mean F1 scores: 0.70–0.78) and sleep apnea classification (0.69 and 0.87 accuracy for severity and presence, respectively). Further analysis reveals that different sleep stages and physiological signals carry distinct predictive power for specific diseases. This work demonstrates that foundation models can extract clinically meaningful features from raw PSG data, enabling scalable, label-efficient analysis and disease prediction.

📖 Table of Contents

1. Installation
2. Usage
3. Licence

💿 Installation

Follow the steps below to set up an environment for running SleepFM.

We recommend using Python 3.10. All required packages and their specific versions are listed in the requirements.txt file. Installing all dependencies should take only 2–3 minutes.

🖥️ System Requirements

SleepFM was developed and tested on Linux systems with the following configuration:

• GPU: NVIDIA A40, A100, and RTX 2080 Ti
• CUDA: 12.4
• CPU: 8 cores recommended
• RAM: At least 32 GB
• OS: CentOS Linux 7.9.2009 (Core)

Although optimized for A40/A100 GPUs, the model can be run on smaller GPUs (e.g., RTX 2080 Ti) by reducing the batch size. For smooth performance during preprocessing and training, we recommend using at least 8 CPU cores.

🚀 Demo Run

This codebase includes a demo using the MESA dataset. On an NVIDIA A40, pretraining on MESA for one epoch takes approximately 1 hour, and fine-tuning for one epoch takes about 1 minute. Note that MESA is significantly smaller than the full dataset used in our main experiments.

git clone https://github.com/zou-group/sleepfm-clinical.git
cd sleepfm-clinical
conda env create -f env.yml
conda activate sleepfm_env

👩‍💻 Usage

This codebase will serve as a framework that you can adapt to your dataset for pretraining and testing. Below, we outline the steps to pretrain and adapt the model on a publicly available dataset called MESA. Please keep in mind that this dataset is small and will most likely not yield optimal results.

Note: Please make sure to download the dataset with in your local path, with dataset name, mesa. Later on, we will need this path.

Preprocessing Dataset

PSG files may be stored in different formats. Here, we specifically provide scripts to process .EDF file format.

• Step 0: preprocessing/preprocessing.py
  • This script converts .EDF file into .hdf5 files with is the format that the model will expect below.

Pretraining

Note that we provide with dataset split as json file here: configs/dataset_split.json. We also provide with different channel groups within a modality: configs/channel_groups.json.

• Step 1: sleepfm/pipeline/pretrain.py
  • This script has our main pretraining config. Its corresponding config file is inside configs/config_set_transformer_contrastive.yaml, where you will set all the parameters and data path. This step will roughly take about an hour for an epoch on MESA.
• Step 2: sleepfm/pipeline/generate_embeddings.py
  • After pretraining our model, we want to generate the embeddings for train/valid/test so that we can train a model for downstream classification. We do sleep stage classification here. This step will roughly take few minutes on MESA.

Evaluation

Note: These evaluation results will not match the ones that we have in our paper as this is a small dataset. This step does not require GPU support.

You should also have extracted the sleep stage labels, which should look like this:

Start,Stop,StageName,StageNumber
0.0,5190.0,Wake,0
0.0,5190.0,Wake,0,
0.0,5190.0,Wake,0

These labels files are stored inside a folder as such <path>/mesa/mesa-sleep-0001.csv. Note that mesa-sleep-0001 is the filename that should correspond with the original .EDF file and .hdf5 files.

• Step 3: sleepfm/pipelinefinetune_sleep_staging.py

  • This will finetune the pretrained model on sleep stage classification task. Please make sure to check config configs/config_finetune_sleep_events.yaml. This step roughly takes less than a minute on MESA.

• Step 4: sleepfm/pipeline/evaluate_sleep_staging.py

  • This will evaluate the model on test set. This step only takes few seconds on MESA.

For disease prediction task:

• Step 3: sleepfm/pipeline/finetune_diagnosis_coxph.py
  • This will finetune the pretrained model on disease prediction task, using CoxPH loss function. Note that you will need to provide your own data, and set up dataloaders. Please see corresponding config sleepfm/configs/config_finetune_diagnosis_coxph.yaml

BibTeX

@article{thapa2025multimodal,
  title={A Multimodal Sleep Foundation Model Developed with 500K Hours of Sleep Recordings for Disease Predictions},
  author={Thapa, Rahul and Kj{\ae}r, Magnus Ruud and He, Bryan and Covert, Ian and Moore, Hyatt and Hanif, Umaer and Ganjoo, Gauri and Westover, Brandon M and Jennum, Poul and Brink-Kj{\ae}r, Andreas and others},
  journal={medRxiv},
  pages={2025--02},
  year={2025},
  publisher={Cold Spring Harbor Laboratory Press}
}

🪪 License

MIT License