🤖 machine learning

STDA-Net: Spectrogram-Based Domain Adaptation for cross-dataset Sleep Stage Classification

The paper proposes STDA-Net, a spectrogram-based unsupervised domain adaptation framework that integrates CNNs, BiLSTMs, and adversarial learning to achieve robust and stable cross-dataset sleep stage classification, outperforming existing 1D signal-based methods across multiple datasets.

Original authors: Unaza Tallal, Shruti Kshirsagar, Ankita Shukla

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Unaza Tallal, Shruti Kshirsagar, Ankita Shukla

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to recognize the different stages of human sleep (like light sleep, deep sleep, or dreaming). You have a lot of training data from one group of people (let's call them "Group A"), but you want the robot to work perfectly on a completely different group of people ("Group B").

The problem is that Group A and Group B are different. Maybe Group A was recorded with a different type of microphone, in a different room, or with a different sampling speed. In the world of data, this is called a "domain shift." It's like trying to teach a student to recognize a dog using only photos of Golden Retrievers, and then expecting them to instantly recognize a Chihuahua without any extra help. The robot gets confused because the "look" of the data has changed.

This paper introduces a new solution called STDA-Net. Here is how it works, broken down into simple concepts:

1. Changing the "Language" of the Data (The Spectrogram)

Most previous methods tried to teach the robot by feeding it the raw sound waves of brain activity (EEG) directly. The authors argue this is like trying to read a book by looking at the ink blots on the page without seeing the letters.

Instead, STDA-Net converts these raw brain waves into spectrograms. Think of a spectrogram as a heat map or a fingerprint of the sound. Instead of just a squiggly line, it turns the data into a 2D picture where you can clearly see the "colors" (frequencies) and the "shapes" (time patterns) of the sleep stages. The paper claims that looking at these 2D pictures helps the robot understand the data much better than looking at the raw lines.

2. The Three-Part Team (The Architecture)

To solve the problem of teaching the robot to work on Group B without ever seeing Group B's answers (labels), the authors built a team of three specialists:

The Artist (CNN): This part looks at the 2D "heat map" pictures and learns to spot the unique patterns of each sleep stage. It's like an artist learning to recognize a style of painting.
The Storyteller (BiLSTM): Sleep doesn't happen in isolated snapshots; it's a story. You don't usually jump instantly from "Deep Sleep" to "Dreaming" without passing through "Light Sleep." This part looks at the sequence of events. It remembers the previous few minutes to understand what is happening right now, just like a storyteller uses context to make sense of the current chapter.
The Spy (DANN): This is the clever part. The Spy's job is to trick the Artist. The Spy tries to guess whether a picture came from Group A or Group B. The Artist tries to make the pictures look so similar that the Spy can't tell them apart. By forcing the Artist to "hide" the differences between the two groups, the robot learns to focus only on the sleep patterns that are universal, ignoring the differences in how the data was recorded.

3. The Training Process

The robot is trained using data from Group A (where we know the answers) and Group B (where we don't know the answers).

The Artist learns to identify sleep stages from Group A.
The Spy tries to tell Group A from Group B.
The Storyteller makes sure the sequence of sleep makes sense.
Through this "game," the robot learns to recognize sleep stages in Group B just as well as it does in Group A, even though it never saw the correct answers for Group B during training.

4. The Results: A More Stable Robot

The researchers tested this system on three different real-world sleep databases. They compared their new method against older methods that only used the raw "squiggly lines" (1D data).

Better Balance: Older methods were good at guessing the most common sleep stages (like being awake) but terrible at the rare ones (like light sleep). STDA-Net was much better at getting all stages right, not just the easy ones.
More Reliable: When the researchers ran the test five times with different random settings, the older methods gave very different results each time (like a coin flip). STDA-Net gave almost the same high score every single time. This means the method is stable and trustworthy.
The Numbers: On average, the new method got about 89% of the sleep stages correct, which was a significant improvement over the previous best methods.

Summary

In short, the paper claims that by turning brain waves into 2D pictures, adding a storyteller to understand the sequence of sleep, and using a spy to hide the differences between different recording setups, they created a system that can accurately classify sleep stages across different datasets. They found this approach to be more accurate, more balanced, and much more stable than previous methods that relied on raw data alone.

Technical Summary: STDA-Net

Problem Statement
Accurate sleep stage classification across different datasets is hindered by significant domain shifts caused by variability in EEG channel montages, sampling rates, recording environments, and subject populations. While deep learning has advanced automated sleep staging, most existing cross-dataset methods rely on one-dimensional (1D) raw EEG signal representations. Furthermore, the application of two-dimensional (2D) spectrogram-based inputs within an unsupervised domain adaptation framework for cross-dataset sleep staging remains largely unexplored. Current approaches often suffer from performance degradation when transferring models between datasets due to these distributional differences, and limited attention has been given to leveraging 2D time-frequency representations to mitigate this gap.

Methodology
The authors propose STDA-Net (Spectrogram-based Temporal Domain Adaptation Network), a framework designed for unsupervised domain adaptation in sleep stage classification using 2D EEG spectrograms. The architecture integrates four primary components:

Preprocessing and Spectrogram Generation: Raw single-channel EEG signals from source and target domains are standardized (downsampled to 100 Hz, bandpass filtered 0.5–30 Hz) and segmented into 30-second epochs. Each epoch is converted into a 2D time-frequency representation (spectrogram) of size $76 \times 60$ using the Short-Time Fourier Transform (STFT). To model temporal dependencies, sequences of $L=10$ consecutive epochs (5 minutes) are constructed with a stride of 5 epochs.
CNN Epoch Encoder: A shared Convolutional Neural Network (CNN) backbone processes each spectrogram independently to extract discriminative epoch-level features. The encoder consists of four convolutional blocks with residual units, batch normalization, and GELU activation. It outputs a 128-dimensional feature vector. An auxiliary classification head is attached to these features to provide direct per-epoch supervision, encouraging the extraction of discriminative spectral features.
BiLSTM Temporal Module: The sequence of epoch-level features is fed into a two-layer Bidirectional Long Short-Term Memory (BiLSTM) network. This module captures temporal dependencies across consecutive epochs, leveraging biologically constrained sleep-stage transition patterns (e.g., the rarity of direct transitions between certain stages) to refine predictions.
Domain-Adversarial Training (DANN): To align source and target feature distributions without labeled target data, a domain discriminator is trained to distinguish between source and target features. A Gradient Reversal Layer (GRL) is applied to the CNN encoder, training it to generate features that confuse the discriminator, thereby learning domain-invariant representations. This adversarial adaptation is applied at the per-epoch CNN feature level, as spectral discrepancies are the primary source of domain gap, while temporal patterns remain relatively consistent.

The total training loss combines three objectives: the main classification loss ( $L_{main}$ ) on source labels, an auxiliary classification loss ( $L_{aux}$ ) on CNN features, and the adversarial domain adaptation loss ( $L_{adv}$ ). The weight for the adversarial loss follows a progressive schedule to allow the model to first learn discriminative source features before gradually increasing alignment pressure.

Key Contributions

Framework Proposal: The introduction of STDA-Net, a novel framework for cross-dataset sleep stage classification operating on 2D EEG spectrograms. It uniquely integrates a CNN encoder for spectral feature extraction, a BiLSTM for temporal modeling, and DANN for unsupervised domain alignment.
Systematic Ablation: A comprehensive ablation study quantifying the individual contributions of temporal modeling, auxiliary supervision, and adversarial adaptation, demonstrating that the combination of all components yields the optimal performance.
Robust Cross-Dataset Evaluation: Extensive experiments conducted on three public datasets (Sleep-EDF, SHHS-1, SHHS-2) across six distinct cross-dataset transfer settings, validating the framework's robustness against domain shifts.

Results
Experiments were conducted on Sleep-EDF, SHHS-1, and SHHS-2 under six transfer settings. The proposed STDA-Net achieved:

An average accuracy of 89.03%.
An average macro F1-score of 87.64%.
Consistently outperformed existing 1D baseline methods (including Deep CORAL, DANN, ADAST, and ADAADL) in terms of balanced classification performance (macro F1).
Demonstrated substantially lower variance across five independent runs compared to baselines, indicating improved stability and reproducibility.

Notably, the model showed a directional asymmetry where transfer from SHHS datasets to Sleep-EDF outperformed the reverse, suggesting that training on physiologically diverse source populations yields more generalizable spectral representations.

Significance
The paper claims that 2D spectrogram-based representations, when combined with temporal modeling and adversarial domain adaptation, provide a robust and competitive alternative to conventional 1D EEG inputs for cross-dataset sleep staging. The findings suggest that moving away from raw 1D waveforms to 2D time-frequency representations helps mitigate the performance degradation typically caused by domain shifts in sleep data. The framework successfully addresses the challenge of generalizing sleep staging models across different acquisition protocols and subject populations without requiring labeled target data, offering a practical solution for improving automated sleep analysis in clinical and research settings. Future work is proposed to validate the framework on additional databases (ISRUC, MASS) and extend it to multi-channel EEG configurations.