The iSLEEPS dataset presents 100 overnight polysomnography recordings from ischemic stroke patients collected at NIMHANS, India, described as 'the first Asian and one of the largest stroke-specific sleep databases,' demonstrating high prevalence of sleep-disordered breathing and enabling automated sleep stage classification with deep learning methods.
Key Findings
Background
The iSLEEPS dataset is the first Asian and one of the largest stroke-specific sleep databases, comprising 100 overnight PSG recordings from ischemic stroke patients.
Data collection was carried out between September 2018 and December 2021 at NIMHANS, India.
Each recording includes sleep stages manually scored at 30-second epochs.
Annotations include detailed respiratory events, periodic limb movements, oxygen desaturation episodes, and clinical metrics.
Scoring followed AASM (2017) guidelines.
Results
The ischemic stroke cohort in iSLEEPS demonstrates a high prevalence of sleep-disordered breathing (SDB).
SDB is described as a common comorbidity in ischemic stroke survivors.
The high prevalence of SDB in the cohort enables investigation of stroke-sleep pathophysiology interactions.
Specific prevalence rates are characterized through comprehensive expert annotations of respiratory events and oxygen desaturation episodes.
Results
A Long Short-Term Memory (LSTM) deep learning model achieved the highest automated sleep stage classification accuracy of 74.70% on the iSLEEPS dataset.
LSTM achieved 74.70% accuracy for automated sleep stage classification.
A Transformer model achieved the second-highest accuracy at 67.44%.
A Convolutional Neural Network (CNN) achieved 61.65% accuracy.
These models were implemented to illustrate dataset utility for automated sleep stage classification.
Background
The iSLEEPS dataset addresses a recognized gap in stroke-specific sleep research by providing comprehensive polysomnography data for post-stroke sleep disturbance analysis.
The dataset was created specifically to address 'the lack of stroke-specific sleep data.'
Prior to this dataset, stroke-specific sleep databases from Asian populations were absent.
The dataset supports analysis of sleep architecture and integrity relevant to neural recovery and cognitive restoration in stroke survivors.
The dataset includes 100 recordings with both physiological signals and comprehensive clinical metrics.
What This Means
This research introduces iSLEEPS, a specialized sleep recording database collected from 100 ischemic stroke patients at a major hospital in India between 2018 and 2021. Each recording captured a full night of sleep using polysomnography — a comprehensive sleep monitoring technique that tracks brain activity, breathing, blood oxygen levels, and limb movements. Experts manually labeled each 30-second segment of sleep according to established international standards. This is described as the first such database from Asia and one of the largest focused specifically on stroke patients anywhere in the world.
The study found that breathing problems during sleep (called sleep-disordered breathing) were very common among the stroke patients in this dataset, which is consistent with known links between stroke and disrupted sleep. To demonstrate the dataset's practical value, the researchers tested three types of artificial intelligence models to see how well computers could automatically identify sleep stages. The best-performing model — a type called Long Short-Term Memory (LSTM) — correctly classified sleep stages about 75% of the time, outperforming a Transformer model (67%) and a Convolutional Neural Network (62%).
This research suggests that having a dedicated, well-annotated stroke sleep database could be an important resource for scientists studying how sleep affects recovery after stroke. Poor sleep is thought to interfere with brain healing, and stroke patients frequently experience sleep disturbances, but there has been very little data available to study this specifically in stroke populations, particularly in Asia. By making this dataset available, the researchers aim to enable further investigation into the relationship between sleep problems and stroke recovery, and to support the development of better automated tools for monitoring sleep in clinical settings.