Pediatric SleepNet: A Deep Learning Model for Reliable Pediatric Sleep Staging Across Developmental Stages — Trained Models and Code

Abstract

Study Objectives. Manual sleep staging in pediatric populations is challenging due to developmental variability and limited scoring consistency, especially in infants and toddlers. We developed a multimodal deep learning model for pediatric sleep staging and evaluated its performance across a broad age range and diverse clinical subgroups.

Methods. We trained a U-Net-inspired encoder-decoder model (Pediatric SleepNet) using 9-channel input signals — electroencephalography (EEG), electrooculography (EOG), and chin electromyography (EMG) — from 35-epoch segments of clinical pediatric polysomnograms (PSGs). Models were trained separately for three age groups (<6 months, 6–12 months, >1 year) using 9,150 PSGs, with 2,455 PSGs reserved for validation. Evaluation was conducted on 3,804 held-out test recordings. Performance was compared with U-Sleep and the Complete Artificial Intelligence Sleep Report (CAISR), and stratified analyses were performed across ages, sexes, and seven ICD-10-based disease categories. External validation used two independent datasets, CHAT and PATS.

Results. Pediatric SleepNet achieved robust performance across all age groups, with mean Cohen’s Kappa increasing from 0.49 (0–6 months) to 0.72 (>12 years), significantly outperforming U-Sleep and CAISR across early developmental stages. Three-class staging yielded mean Cohen’s Kappa increasing from 0.66 (0–6 months) to 0.79 (>12 years). Sex-based differences were negligible; performance was reduced in children with epilepsy, Down syndrome, and hydrocephalus.

Conclusions. Pediatric SleepNet demonstrates reliable sleep staging across pediatric development. Its robust performance across age, disease, and external datasets supports its potential for clinical and research use in pediatric sleep medicine.

Background

Sleep staging is central to the diagnosis and management of pediatric sleep disorders, but manual scoring is labor-intensive and subject to substantial inter-rater variability. Pediatric sleep staging is especially challenging because neurophysiology and sleep architecture change dramatically with age: EEG waveforms, event morphology, spectral content, and stage transitions differ markedly between neonates, infants, children, and adolescents. As a result, automated models trained on adult data perform poorly when applied to young children, and a single model rarely generalizes across the full pediatric age span.

Deep learning has achieved expert-level automated sleep staging in adults, but reliable, developmentally-aware pediatric models have been lacking. Pediatric SleepNet addresses this gap with a multimodal (EEG+EOG+EMG), age-specific modeling approach trained and validated on a large clinical pediatric PSG corpus, and benchmarked against existing automated methods (U-Sleep, CAISR) and expert scoring.

Model Description

This release provides the trained Pediatric SleepNet models and the training/inference code. It is a model-and-code companion to the underlying data, the Boston Children’s Hospital Sleep Corpus (published separately on BDSP).

Data layout on S3

Files are hosted in the bdsp-opendata-credentialed bucket under prefix pediatric-sleepnet/ (credentialed access).

s3://bdsp-opendata-credentialed/pediatric-sleepnet/
├── models/     # six trained Keras (.h5) models
│   ├── trained_model_usleep_3models_below6mo_EEGEOGEMG.h5     # 0-6 months
│   ├── trained_model_usleep_3models_6mo1y_EEGEOGEMG.h5       # 6-12 months
│   ├── trained_model_usleep_3models_above1y_EEGEOGEMG.h5     # > 1 year
│   └── ..._normalized_...                                     # same 3 age groups, normalized-input variant
├── code/       # training + inference scripts (paths de-hardcoded to relative ./data, ./models, ./output)
├── README.md · REPRODUCE.md · DATA_SOURCE.md · requirements.txt

Sleep stages are 5-class (Wake, N1, N2, N3, REM); the model also supports a 3-class scheme. The exact per-recording train/validation/test split lists were not recoverable from the authors’ archives at release time and are not included; the split methodology is visible in the training scripts.

Technical Implementation

Model

Pediatric SleepNet is a U-Net-inspired encoder-decoder network that ingests 9 channels (EEG, EOG, chin EMG) and classifies sleep stages over 35-epoch segments. To accommodate developmental variability, three separate models are trained for age groups <6 months, 6–12 months, and >1 year; at inference each recording is routed to the age-appropriate model. Both standard and per-recording signal-normalized variants are provided.

Data and preprocessing

Training and evaluation used the Boston Children’s Hospital (BCH) Sleep Corpus (see Data description). Recordings were stored as standardized EDF and processed uniformly with MNE-Python (v1.10.2): 60 Hz notch filtering, band-pass filtering, resampling, montage selection, normalization, and epoching. Training used 9,150 PSGs; 2,455 were held out for validation; 3,804 independent recordings formed the test set.

Evaluation

Models were compared with U-Sleep and CAISR and against expert scoring using Cohen’s Kappa, with stratification by age, sex, and seven ICD-10-based disease categories. External validation used the CHAT and PATS cohorts. Final models were saved after convergence (early stopping on validation accuracy, patience 50) and used for inference without post-processing.

Ethics

Retrospective analysis under IRB protocols BIDMC #2016P000058 and MGH #2013P001024, with a waiver of consent.

Installation and Requirements

No installation is required to browse this project. To run the models, clone the code repository and install dependencies:

git clone https://github.com/bdsp-core/BCH-SleepStaging.git
cd BCH-SleepStaging
pip install -r requirements.txt

The trained model weights (.h5) are downloaded from this project’s S3 area (s3://bdsp-opendata-credentialed/pediatric-sleepnet/models/) after credentialed access is granted. A GPU is recommended for training; inference runs on CPU or GPU.

Usage Notes

Getting started

pip install -r requirements.txt   # TensorFlow/Keras, MNE-Python 1.10.2, NumPy, pandas, h5py, scikit-learn

Inference

Place the age-appropriate model(s) in ./models/ and BCH recordings (arranged .h5) in ./data/BCH_h5_arranged/, then run code/3models_EEGEOGEMG_predictions.py (or code/normalized_make_predictions.py for the normalized variant). Each recording is automatically routed to its age-group model; predictions and reference labels are written to ./output/. code/parallel_test.py performs batched inference.

Training

code/train_3models_EEGEOGEMG_parallel.py and code/train_3models_EEGEOGEMG_normalized_parallel.py rebuild the three age-group models from the BCH recordings.

Data and code

Data: the BCH Sleep Corpus at bdsp.io/content/l8c86mgywuneuy74ae71 (credentialed). Code is also mirrored at github.com/bdsp-core/BCH-SleepStaging. Model weights and data are subject to the BDSP credentialed Data Use Agreement.

Release Notes

Version 1.0.0 — initial release of the Pediatric SleepNet trained models (six Keras .h5 checkpoints: three age groups × standard/normalized variants) and the training/inference code, accompanying Tripathi et al., Sleep 2026 (doi:10.1093/sleep/zsag064).

Ethics

This retrospective data analysis was conducted under IRB protocols BIDMC #2016P000058 and MGH #2013P001024, with the MGH and BIDMC IRBs granting a waiver of consent. The underlying Boston Children’s Hospital Sleep Corpus is de-identified and access-controlled under the BDSP credentialed Data Use Agreement.

Acknowledgements

This work was supported by grants from the National Institutes of Health (R01AG073410, R01HL161253, R01NS126282, R01AG073598, R01NS131347, R01NS130119) and by Amazon Web Services (AWS).

Conflicts of Interest

Dr. Westover is a co-founder of, scientific advisor and consultant to, and holds personal equity in Beacon Biosignals. Dr. Clifford has received research funding from the NSF, NIH, and LifeBell AI, and unrestricted donations from AliveCor Inc, Amazon Research, the Center for Discovery, the Gates Foundation, Google, the Gordon and Betty Moore Foundation, MathWorks, Microsoft Research, NextSense Inc, the One Mind Foundation, and the Rett Research Foundation; he holds advisory roles and financial interests in AliveCor Inc and NextSense Inc, and is Chief Technology Officer of MindChild Medical with significant stock ownership. Dr. Thomas is a co-inventor of intellectual property licensed by Beth Israel Deaconess Medical Center to MyCardio, LLC (cardiopulmonary sleep spectrogram) and has submitted patents related to high-loop-gain sleep apnea; he provides paid consulting to GLG Councils, Guidepoint, Beacon Biosignals, and Jazz Pharmaceuticals. Dr. Stone reports grant funding from Eli Lilly, consults for Axsome Therapeutics, and receives a stipend from the Sleep Research Society as deputy editor for SLEEP. Dr. Maski consults for Alkermes, Avadel, Harmony Biosciences, Jazz Pharmaceuticals, and Takeda; has received grant funding from Harmony Biosciences and Jazz Pharmaceuticals; serves as DSMB chair for Idorsia; and is a collaborator on clinical trials sponsored by Alkermes and Takeda. Dr. Sun reports grant funding via a Strategic Research Award from the American Academy of Sleep Medicine. Dr. Trotti is a member of the Board of Directors of the American Academy of Sleep Medicine. All other authors report no conflicts of interest relevant to this work.

References

1. Tripathi A, Gupta A, Ganglberger W, Waters S, Sun H, Nasiri S, Mitra A, Stone KL, Mignot E, Hwang D, Reyna MA, Trotti LM, Clifford GD, Maski K, Katwa U, Thomas RJ, Westover MB. Pediatric SleepNet: a deep learning network for reliable pediatric sleep staging across developmental stages. Sleep. 2026 Aug 14;49(8):zsag064. doi: 10.1093/sleep/zsag064. PMID: 41804802; PMCID: PMC13472892.