Validation and optimization of automatized sleep spindle detectors in elderly healthy subjects and patients with Parkinson's disease

L. Rose, H. Leonthin, M. Nikolic, P. Jennum, J.A.E. Christensen

Research output: Contribution to journalConference abstract in journalResearchpeer-review

Abstract

Objectives/Introduction: Implementation of an Automatic Sleep Spindle Detector (ASSD) could help standardize scorings of sleep spindles (SS). In this study three ASSDs were implemented to investigate their performance in detecting SS in respectively healthy subjects and Parkinson's Disease (PD) patients, before and after optimization. The implemented ASSDs were inspired from Latreille et al. (ASSD‐L;[1]), Schimicek et al. (ASSD‐S;[2]) and Ferrarelli et al. (ASSD‐F;[3]).

Methods: Polysomnographic electroencephalography data from 15 PD patients and 15 healthy elderly were included. SS were manually identified in 20 minutes of noise‐free N2 sleep by five experts, and a gold standard was defined by a group consensus (GS) [4]. The optimization was based on parameter studies including the parameters: amplitude‐threshold, root‐mean‐square (rms)‐EMG and rms‐ratio between alpha and SS (ASSD‐S), lower and upper threshold (ASSD‐F) and percentile (ASSD‐L). A leave‐one‐subject‐out training approach was applied for 12/15 controls. The performances of the optimized ASSDs were evaluated on a test set of 3/15 controls and 15 patients based on a by‐event F1‐score, where detected events were considered true positives if they overlapped with a SS from GS with more than 20%.

Results: The optimal parameters were estimated to 13, 2 and 1.9 (ASSD‐S); 3·[most‐common‐envelope‐peak‐amplitude] and 6·mean (|signal|) (ASSD‐S); and 96.2th‐percentile (ASSD‐L). The F1‐score (µ±σ) for the test controls changed from 0.20 ± 0.14 to 0.41 ± 0.10 (ASSD‐F), 0.49 ± 0.15 to 0.59 ± 0.02 (ASSD‐S), and 0.36 ± 0.09 to 0.35 ± 0.12 (ASSD‐L). Furthermore, the F1‐score for the PD patients changed from 0.20 ± 0.20 to 0.29 ± 0.30 (ASSD‐F), 0.30 ± 0.30 to 0.35 ± 0.32 (ASSD‐S), and 0.28 ± 0.30 to 0.27 ± 0.28 (ASSD‐L). All ASSDs detected SS in the control group, however, none of the few SS from GS were detected in 2/15 PD patients (ASSD‐F and ASSD‐S) and 3/15 PD patients (ASSD‐L). The inter‐group comparison (Wilcoxon tests) showed that none of the ASSDs performed significantly different (p = 0.94 (ASSD‐F), p = 0.94 (ASSD‐S) and p = 0.48 (ASSD‐L)) between groups.

Conclusions: The parameter studies improved the ASSDs for both groups. Although the mean F1‐scores were consistently lower for the PD group compared to the test controls, it did not differ significantly between groups. Yet, a dataset with more subjects and scorings outside N2 sleep is required to fully investigate the robustness of the ASSDs.

Disclosure: [1] Latreille et al. ‘Sleep spindles in Parkinson´s disease may predict the development of dementia’. Neurobiology of Aging — 2015, Volume 36, Issue 2 [2] Schimicek et. al. ‘Automatic Sleep‐Spindle Detection Procedure: Aspects of Reliability and Validity’., Clinical EEG and Neuroscience 1994, vol. 25, issue 1 [3] Ferarelli, F et al. ‘Reduced sleep spindle activity in schizophrenia patients’. Am J Psychiatry 164:3. March 2007. [4] Christensen, Julie A E et al. “Sleep spindle alterations in patients with Parkinson's disease.” Frontiers in Human Neuroscience vol. 9 233. May 2015.
Original languageEnglish
Article numberP252
JournalJournal of Sleep Research
Volume29
Issue numberSuppl. 1, Sp. Iss. SI
Pages (from-to)193-194
ISSN1365-2869
Publication statusPublished - 2020

Fingerprint Dive into the research topics of 'Validation and optimization of automatized sleep spindle detectors in elderly healthy subjects and patients with Parkinson's disease'. Together they form a unique fingerprint.

Cite this