0323 Design of a Deep Learning Based Algorithm forAutomatic Detection of Leg Movements During Sleep
Abstract Introduction Leg Movements (LM) and Periodic Leg Movements (PLM) during sleep are a key feature of nocturnal polysomnographic (PSGs) sleep studies. The current practice is manual annotation by technicians, which is time-consuming and prone to human error. To automate scoring, other approach...
Gespeichert in:
| Veröffentlicht in: | Sleep (New York, N.Y.) Jg. 41; H. suppl_1; S. A124 |
|---|---|
| Hauptverfasser: | , , , , , , , , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
US
Oxford University Press
27.04.2018
|
| Schlagworte: | |
| ISSN: | 0161-8105, 1550-9109 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | Abstract
Introduction
Leg Movements (LM) and Periodic Leg Movements (PLM) during sleep are a key feature of nocturnal polysomnographic (PSGs) sleep studies. The current practice is manual annotation by technicians, which is time-consuming and prone to human error. To automate scoring, other approaches use rule-based detection of increased motor activity to identify PLMs. We devised and validated a machine learning-based algorithm for detection of LMs.
Methods
A representative sample of 359 (192 males, 167 females; Age 63.4 ± 7.7(mean±SD); BMI 32.0 ± 7.2) PSGs was drawn from the Wisconsin Sleep Cohort (WSC) and LMs were scored by experts using latest AASM criteria. The average number of LMs per subject was 185.5 ± 176.4 with average duration of 2.9s±1.7s. After adaptive filtering to remove ECG artifact, the combined left/right anterior tibialis channel was used to extract 16 time-domain features. Additional model features included manually scored sleep stages and apneic events, as these are required to eliminate wakefulness and apnea-associated LMs. A long short-term memory (LSTM) machine-learning algorithm was trained on 316 subjects. Model hyperparameters were tuned using an independent development set of 13 PSGs. We used a test set of 30 WSC PSGs to provide preliminary results on model performance reported below. We are now having 5 independent technicians score LMs on the same 30 WSC PSGs. In this context, the algorithm’s performance will be compared to each individual scorer using consensus score as ground truth. We will extend this study to include PSGs from the Stanford Sleep Cohort and will report on both LM and PLM scoring accuracy.
Results
Currently, analysis of the algorithm performance showed precision (positive predictive value)=0.796 ± 0.018, recall (sensitivity)=0.850 ± 0.016 and F1=0.822 ± 0.005 on the test set.
Conclusion
Most LMs were correctly and consistently detected with this automated algorithm. Detection difficulties were in cases with noisy recordings and/or inaccurate annotations. Using an algorithm trained to understand LM features can provide better parameters for defining LMs and PLMs while unburdening technicians/clinicians from scoring these physiologically relevant features.
Support (If Any)
Klarman Family Foundation, H. Lundbeck A/S and Foundation, Technical University of Denmark. |
|---|---|
| Bibliographie: | ObjectType-Article-1 SourceType-Scholarly Journals-1 ObjectType-Feature-2 content type line 14 |
| ISSN: | 0161-8105 1550-9109 |
| DOI: | 10.1093/sleep/zsy061.322 |