Source-linked AI summary
U-Time: A Fully Convolutional Network for Time Series Segmentation Applied to Sleep Staging
Mathias Perslev, Michael Hejselbak Jensen, Sune Darkner, Poul Jørgen Jennum, Christian Igel
TL;DR
Manual sleep staging is difficult and time-consuming, while recurrent deep-learning models can be difficult to tune across tasks. U-Time uses a fully convolutional encoder-decoder that densely segments signals and aggregates point-wise scores at a chosen temporal scale, reaching or surpassing state-of-the-art models across seven PSG datasets with fixed settings.
Problem
Manual sleep staging requires time-consuming expert inspection of long multichannel recordings, while recurrent neural networks are difficult to tune and optimize.
Method
U-Time is a fully convolutional U-Net-inspired encoder-decoder that maps whole physiological sequences to dense point-wise segmentations and aggregates them into predictions at a chosen temporal resolution.
Results
Across seven PSG datasets, U-Time reached or surpassed state-of-the-art models and a CNN-LSTM baseline using the same architecture and hyperparameter settings.
Takeaways & Limitations
U-Time is presented as a robust, flexible feed-forward alternative to recurrent architectures for physiological time-series segmentation.
Takeaways & Limitations
The study used a single EEG channel, although multiple channels may benefit sleep staging, particularly where EOG helps distinguish wake from REM sleep.
Abstract
from arXiv · showhide
Neural networks are becoming more and more popular for the analysis of physiological time-series. The most successful deep learning systems in this domain combine convolutional and recurrent layers to extract useful features to model temporal relations. Unfortunately, these recurrent models are difficult to tune and optimize. In our experience, they often require task-specific modifications, which makes them challenging to use for non-experts. We propose U-Time, a fully feed-forward deep learning approach to physiological time series segmentation developed for the analysis of sleep data. U-Time is a temporal fully convolutional network based on the U-Net architecture that was originally proposed for image segmentation. U-Time maps sequential inputs of arbitrary length to sequences of class labels on a freely chosen temporal scale. This is done by implicitly classifying every individual time-point of the input signal and aggregating these classifications over fixed intervals to form the final predictions. We evaluated U-Time for sleep stage classification on a large collection of sleep electroencephalography (EEG) datasets. In all cases, we found that U-Time reaches or outperforms current state-of-the-art deep learning models while being much more robust in the training process and without requiring architecture or hyperparameter adaptation across tasks.
1 Introduction
Sleep staging maps overnight physiological signals to sleep-stage labels, but manual scoring is difficult and time-consuming. U-Time addresses this challenge with a feed-forward approach evaluated across multiple PSG datasets.
- Motivation: Sleep staging supports diagnosis of many sleep-related disorders by mapping transitions between stages from physiological recordings.Relevant signals include EEG, EOG, and EMG measured in polysomnography studies.
- Motivation: Manual scoring requires clinicians to inspect and segment typically 8–24-hour multichannel recordings into individually classified 30-second intervals.The process is described as difficult and time-consuming.
- Existing approaches: Deep-learning sleep-staging systems commonly combine convolutional and recurrent layers to model physiological time-series data.The introduction identifies these combined architectures as the current best systems in the cited work.
- Existing approaches: Recurrent neural networks are often difficult to tune and optimize, motivating feed-forward alternatives for time-series analysis.The cited literature reports that feed-forward systems can replace recurrent models on many tasks without sacrificing accuracy.
- Study approach: The study introduces U-Time and evaluates one fixed architecture and hyperparameter set across seven PSG datasets.Additional recurrent-model experiments assess robustness to architecture and hyperparameter changes.
2 Method
U-Time is a fully convolutional encoder-decoder that performs dense point-wise segmentation and aggregates scores to produce predictions at a chosen temporal resolution. Its convolutional design supports variable-length inputs and flexible segmentation frequencies.
- Architecture: U-Time adapts the U-Net idea to one-dimensional time series by mapping a whole sequence to dense segmentation in one forward pass.It is a fully convolutional encoder-decoder inspired by U-Net and temporal convolutional networks.
- Input and output: The model processes T connected segments of length i and predicts T class labels for K classes from multichannel physiological signals.The target segmentation frequency e determines the number of samples i per segment and labels floor(τ · e).
- Flexible temporal resolution: U-Time assigns class-confidence scores to every input time point, then projects these dense scores onto longer intervals at the selected segmentation frequency.A model trained at e = 1/30 Hz can produce predictions at higher frequencies during inference, including the technically possible e = S case.
- Flexible temporal resolution: Figure 1 depicts point-wise signal segmentation followed by aggregation into final predictions at the chosen temporal scale.The illustration uses T = 4 and e = 1/30 Hz for visual purposes.
- Architecture: The architecture contains an encoder, decoder, and segment classifier that respectively build feature maps, recover dense segmentation, and aggregate scores into sleep-stage predictions.The encoder downsamples through convolution blocks, while the decoder upsamples and combines corresponding encoder feature maps.
- Input and output: Convolutional operations allow the input length to vary between training and testing, provided it is large enough for the encoder’s pooling operations.The implementation requires a minimum input length of tmin = 1920, equivalent to 19.2 seconds at 100 Hz.
- Architecture: The decoder produces t × K point-wise confidence scores, while the segment classifier mean-pools non-overlapping intervals of width i and stride i.During training, this maps the dense segmentation to a T × K output matching the available segment labels.
- Flexible temporal resolution: Learning mean activation over segments encourages confidence scores to be distributed across each segment, enabling shorter-scale predictions at inference.These scores can highlight uncertainty or rapid transitions on timescales shorter than 30 seconds.
3 Experiments and Evaluation
The experiments evaluate U-Time on seven heterogeneous sleep EEG datasets using a fixed five-stage setup, preprocessing pipeline, architecture, and hyperparameter set. Subject-level cross-validation and literature-compatible F1 evaluation were used to assess performance across varied cohorts and recording conditions.
- Experimental setup: The experiments used five AASM sleep stages, with 30-second segments and a fixed input of 35 connected segments.For 100 Hz signals, each input spanned 17.5 minutes and contained 105,000 samples.
- Preprocessing: EEG signals were resampled to 100 Hz, standardized per record using median 0 and IQR 1, and harmonized to the label set {W, N1, N2, N3, R}.Earlier S3 and S4 annotations were merged into N3, while rare movement and non-scored stages were discarded.
- Optimization: U-Time optimization used Adam with learning rate η = 5 · 10^-6 and a generalized Dice cost with uniform class weights.Training batches were sampled by first selecting a sleep-stage class and then a corresponding sleep period from a random PSG record.
- Experimental setup: A fixed architecture and hyperparameter set was applied across datasets without task-specific architectural modifications.The study was designed to test performance across significantly different cohorts under the same model configuration.
- Evaluation: Evaluation used subject-level cross-validation and class-wise F1/Dice scores computed from confusion matrices aggregated across records and splits.The comparison included literature studies with subject-level rather than segment-level cross-validation and clearly described evaluation strategies.
- Datasets: U-Time was evaluated on seven public and non-public PSG datasets spanning diverse sleep-staging cohorts and real-life recording scenarios.The datasets covered multiple decades, sites, instruments, protocols, and healthy or diseased individuals.
4 Results
Across seven datasets, U-Time achieved high performance without task-specific modifications and performed similarly to or better than the CNN-LSTM baseline. Experiments also showed that modest DeepSleepNet changes could substantially affect performance, especially for N1 and REM.
- U-Time results: U-Time achieved high performance across seven datasets spanning dataset sizes, population health status, EEG channels, sample rates, protocols, and sites.The same architecture and hyperparameters were used across these varied settings.
- U-Time results: U-Time performed at least as well as any fairly comparable automated literature method, including methods tailored to individual datasets.The comparison was restricted to studies using subject-level cross-validation and sufficiently clear evaluation reporting.
- Baseline comparison: In all evaluated cases, U-Time performed similarly to or better than the CNN-LSTM baseline.Both models processed single-channel EEG data in the reported comparison.
- CNN-LSTM robustness: Across 13 DeepSleepNet architectural changes, no improvement over the published Sleep-EDF-39 baseline was observed.On DCSM, a smaller DeepSleepNet with smaller CNN filters improved performance slightly over its baseline.
- CNN-LSTM robustness: Modest DeepSleepNet changes could produce large performance changes, particularly for the N1 and REM sleep stages.The authors interpret the lack of Sleep-EDF-39 improvement as indicating that the published architecture was already highly optimized for that cohort.
5 Discussion and Conclusions
U-Time combines fully convolutional time-series segmentation with flexible temporal outputs, achieving strong sleep-staging performance while remaining robust and practical across varied settings.
- U-Time properties: U-Time implicitly segments every input time point and aggregates class scores to produce outputs at the desired temporal resolution.This fully convolutional encoder-decoder approach maps whole sequences to dense segmentations in one forward pass.
- Robustness: The study evaluated U-Time on seven sleep PSG datasets using identical network architecture and hyperparameter settings across tasks.The fixed configuration was intended to demonstrate robustness without task-specific tuning.
- Robustness: U-Time reached or surpassed state-of-the-art models and a CNN-LSTM baseline across the evaluated datasets without task-specific tuning.The CNN-LSTM baseline showed transfer-performance drops and substantial F1 variation under small architecture changes.
- Evaluation context: Human annotators may disagree on sleep staging, with typical expert mean F1 overlap at or slightly above 0.8.N1 is especially difficult because of its transitional nature and weak separation from wake and deep sleep.
- Limitations: Using only a single EEG channel likely limited performance relative to multi-channel assessment, particularly for distinguishing REM from wake.EOG information improved REM classification on ISRUC and other datasets.
- Practical implications: U-Time can score an entire PSG in one forward pass and provide higher-temporal-resolution classifications than the training labels.These properties may support rapid clinical scoring, prediction explanation, and analysis of shorter sleep-stage dynamics.
- Conclusion: The authors conclude that fully convolutional feed-forward architectures are promising alternatives to recurrent models for time-series segmentation.They expect the basic design to apply readily beyond sleep staging.
Supplementary Material
The supplementary material documents sleep-stage definitions, PSG inputs, U-Time architecture and hyperparameters, detailed performance summaries, confusion matrices, and multi-channel or hyperparameter experiments.
- Sleep-stage reference: Table S.1 characterizes the typical features of the five sleep stages defined by the AASM manual.The study uses the five-class system W, N1, N2, N3, and R.
- Data reference: Figure S.1 illustrates a 30-second PSG segment containing EOG, EEG, EMG, and ECG channels used for expert sleep-stage scoring.Most experiments use only one EEG channel, such as C3.
- Architecture: Table S.2 specifies U-Time topology dimensions for i = 3000, C = 1, T = 35, and K = 5, with dilated encoder kernels.The architecture description covers encoder, decoder, pooling, convolution, and skip-connection components.
- Additional experiments: Additional tables report multi-channel results and hyperparameter experiments, with subset-based comparisons requiring interpretation against their designated baselines.The hyperparameter experiments use reduced record subsets and should not be compared directly with full-dataset results.