TY  - CONF
AU  - Bi, Hanwen
AU  - Bülow, Robin
AU  - Deantoni, Michele
AU  - Elmenhorst, David
AU  - Elmenhorst, Eva-Maria
AU  - Ewert, Ralf
AU  - Ferrarelli, Fabio
AU  - Frenzel, Stefan
AU  - Grabe, Hans J
AU  - Hoepel, Sanne J. W.
AU  - Hoffstaedter, Felix
AU  - Jahanshad, Neda
AU  - Keihani, Ahmadreza
AU  - Küppers, Vincent
AU  - Luik, Annemarie I.
AU  - Mayeli, Ahmad
AU  - Mortazavi, Nasrin
AU  - Nilsonne, Gustav
AU  - Rupp, Julia S
AU  - Saberi, Amin
AU  - Schmidt, Christina
AU  - Spiegelhalder, Kai
AU  - Tamm, Sandra
AU  - Thomopoulos, Sophia I.
AU  - Thompson, Paul M.
AU  - Valk, Sofie
AU  - Vandewalle, Gilles
AU  - Völzke, Henry
AU  - Weihs, Antoine
AU  - Wexler, Joseph
AU  - Wittfeld, Katharina
AU  - Eickhoff, Simon
AU  - Patil, Kaustubh
AU  - Raimondo, Federico
AU  - Tahmasian, Masoud
TI  - Associations between sleep and cognitive performance using the ENIGMA-Sleep data
M1  - FZJ-2023-05233
PY  - 2023
AB  - Background: Sleep disturbance is considered a potential risk factor for cognitive decline and dementia. Previous large-scale studies using the UK-Biobank data have highlighted a significant, albeit small effect size, non-linear relationship between self-reported sleep duration and cognitive performance. In this study, we performed machine learning (ML) analysis based on both self-reported and objective sleep duration and sleep efficiency using the ENIGMA-Sleep data to predict cognitive scores at the individual level.Methods: A total of 1,040 subjects from two ENIGMA-Sleep collaboration sites (SHIP-TREND, Liege) were included. Sleep measurements were sleep duration and sleep efficiency extracted from Polysomnography (PSG) data and Pittsburgh Sleep Quality Index (PSQI). Cognitive performance was measured using Stroop scores (Liege and SHIP-Trend) and N-back Working Memory Accuracy (Liege). In addition, age, sex, BMI, and depression (Beck DepressionInventory (BDI-II)) scores were added to our ML models as input features. Multiple ML models were tested, including polynomial regression, support vector machine (SVM) with linear and rbf kernel, and random forest (RF) using the Julearn python library. ML models’ performance was evaluated using mean absolute error, mean squared error, R2, and holdout correlation.Results: Our preliminary results demonstrated that sleep measurements and demographic data were predictive of the Stroop interference score in the Liege dataset(r2 = 0.283 ± 0.231, by RF) and the Stroop reaction time in the SHIP-Trend dataset (r2 = 0.108 ± 0.077, by SVM-rbf) using RF, SVM-linear, and SVM-rbf, outperforming polynomial regression. Age was a significant feature in both datasets with mean feature importance of 0.240 ± 0.0586 (RF) in SHIP-Trend and 0.616 ± 0.218 (SVR-rbf) in Liege. Sleep measurements showed weaker importance in the prediction model. In SHIP-Trend dataset, PSG Sleep Duration was the most impactful sleep measurement (feature importance = 0.00185 ± 0.00223 by RF), whereas in Liege dataset, it was Self-report Sleep Efficiency(feature importance = 0.0479 ± 0.0477 by SVM-rbf). However, the Liege dataset's sleep measurements and demographic data could not predict N-back working memory accuracy.Conclusions: Based on two independent samples, our findings demonstrate that Stroop interference and reaction time scores can be predicted by subjective and objective sleep measurements and demographic data (mainly age) based on ML models.
T2  - eSleep Europe Virtual Congress 2023
CY  - 4 Oct 2023 - 6 Oct 2023, Virtual (Germany)
Y2  - 4 Oct 2023 - 6 Oct 2023
M2  - Virtual, Germany
LB  - PUB:(DE-HGF)6
DO  - DOI:10.34734/FZJ-2023-05233
UR  - https://juser.fz-juelich.de/record/1019187
ER  -