Open-Source EEG Neural Network Checkpoint Management Dashboard: Sleep Staging Inference and Visual Analysis Tool

A PyTorch Lightning model management dashboard developed with Streamlit, supporting model checkpoint management, batch inference, sleep staging visualization, and multi-model comparative analysis for SHHS sleep EEG data.

• Original Author/Maintainer: dhinostroza
• Source Platform: GitHub
• Original Title: EEG-Neural-Network-Checkpoint-Dashboard
• Original Link: https://github.com/dhinostroza/EEG-Neural-Network-Checkpoint-Dashboard
• Release Date: June 2026

---

Sleep disorders have become a global public health issue, affecting the quality of life of hundreds of millions of people. Accurate sleep staging is a key step in diagnosing sleep disorders. Traditional methods rely on manual interpretation of polysomnography (PSG), which is time-consuming and highly subjective. With the development of deep learning technology, automatic sleep staging systems based on electroencephalography (EEG) have shown great potential. However, researchers face a common challenge in their work: how to efficiently manage large numbers of model checkpoints, quickly compare the performance of different models, and intuitively display inference results.

The open-source project introduced in this article was created to address this pain point—a checkpoint management dashboard specifically designed for EEG neural networks, which integrates model management, inference execution, and result visualization into an intuitive web interface.

---

One of the core components of this dashboard is an automated model registry system. It can intelligently scan specified directories for .ckpt checkpoint files and automatically extract key metadata, including:

• Validation Loss: A core metric for evaluating model generalization ability
• Network Architecture Information: Model structure parameters and configuration
• Training Epochs: Model convergence level
• Parameter Count: Quantification of model complexity

The system uses color coding to intuitively display model performance—green for high-performance models (low validation loss) and red for models that need further tuning. This visual feedback mechanism allows researchers to quickly locate the best model among numerous checkpoints.

The dashboard provides a fully functional inference environment that supports multiple input formats:

Supported File Types:

• EEG spectrogram files in .parquet format
• EEG files in .edf (European Data Format) standard
• Files in .bdf (BioSemi Data Format)

Intelligent Preprocessing Flow: Uploaded EDF/BDF files are automatically converted to spectrogram format. The system has built-in unit conversion logic (BDF units are automatically scaled by 1e6 times to match the microvolt standard) to ensure data consistency.

Instant Result Feedback: The system first queries precomputed results in the predictions.sql database for millisecond-level responses. For newly uploaded files, it triggers a real-time inference process.

Inference results are presented in the form of an intuitive hypnogram, clearly showing the整晚 sleep structure:

• Wake
• N1 (Light Sleep)
• N2 (Medium Sleep)
• N3 (Deep Sleep)
• REM (Rapid Eye Movement Sleep)

The visualization component uses optimized plotting parameters—wider chart area and enhanced line contrast—to ensure that even long sleep records are clearly readable.

---

The project uses a modern Python technology stack:

• Streamlit: As the web application framework, providing concise declarative UI building capabilities
• PyTorch Lightning: Abstracts PyTorch training processes and standardizes checkpoint formats
• Timm (PyTorch Image Models): Provides a rich library of pre-trained visual models
• Altair: A declarative statistical visualization library for generating interactive charts
• Pandas/NumPy: Infrastructure for data processing and numerical computation

To meet the batch analysis needs in research scenarios, the system implements dual-model parallel inference functionality:

1. Baseline Model (2025-09-04): As a performance reference
2. Ensemble Model: Fuses multi-model predictions to improve stability

Batch processing results generate a comparative report, including:

• Agreement Matrices: Show the consistency of model predictions
• Comparative Hypnograms: Intuitively compare staging results of different models
• Category Distribution Statistics: Quantitative comparison of the proportion of each sleep stage