Performance of an automated sleep scoring approach for actigraphy data in children and adolescents

Authors: CHEN, PIN - The Children'S Hospital Of Philadelphia; JANSEN, ERICA - University Of Michigan; CIELO, CHRISTOPHER - University Of Pennsylvania; WILLIAMSON, ARIEL - University Of Oregon; BANKER, MARGARET - Northwestern University; KAYE, MICHAEL - University Of Michigan; SONG, PETER - University Of Michigan; PETERSON, KAREN - University Of Michigan; CANTORAL, ALEJANDRA - Ibero-American University; TÉLLEZ-ROJO, MARTHA - Ibero-American University; GOLDSTEIN, CATHY - University Of Michigan; ZANNA, KHADIJA - Rice University; SANO, AKANE - Rice University; MORENO, JENNETTE - Children'S Nutrition Research Center (CNRC); KALKWARF, HEIDI - University Of Cincinnati College Of Medicine; ZEMEL, BABETTE - University Of Pennsylvania; MITCHELL, JOHNATHAN - University Of Pennsylvania

Submitted to: Sleep
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/8/2025
Publication Date: 9/19/2025
Citation: Chen, P.W., Jansen, E.C., Cielo, C.M., Williamson, A.A., Banker, M., Kaye, M., Song, P.X., Peterson, K.E., Cantoral, A., Téllez-Rojo, M.M., Goldstein, C., Zanna, K., Sano, A., Moreno, J.P., Kalkwarf, H., Zemel, B.S., Mitchell, J.A. 2025. Performance of an automated sleep scoring approach for actigraphy data in children and adolescents. Sleep. https://doi.org/10.1093/sleep/zsaf282.
DOI: https://doi.org/10.1093/sleep/zsaf282

Interpretive Summary: GGIR is a free software tool that processes raw movement data from wearable devices to estimate sleep patterns. Researchers wanted to see how well GGIR's three sleep algorithms (Cole-Kripke, Sadeh, and van Hees) detect sleep and wake compared to the gold standard test called polysomnography (PSG). They also wanted to check if GGIR gives reasonable sleep estimates in real-world studies with kids. To examine the accuracy of the GGIR algorithms, 30 children (ages 8–16) wore an activity tracker during an overnight PSG test in a lab. To examine the real-world implementation of these algorithms, data from four large studies of typically developing kids (ages 3–18) were analyzed using GGIR. The in the lab tests revealed that GGIR's algorithms were pretty good at detecting sleep vs. wake. Cole-Kripke and van Hees worked best (about 80% accuracy) and Sadeh was less accurate (about 67% accuracy). All algorithms tended to detect sleep easily (high sensitivity) but were less precise at spotting wake periods (moderate specificity). In real-world data, the results looked similar: Cole-Kripke and van Hees gave similar sleep duration estimates, while Sadeh estimated about an hour more sleep. GGIR is a useful, open-source option for researchers to automatically process movement data and estimate sleep in children, including those with sleep disorders like obstructive sleep apnea.

Technical Abstract: GGIR is an R package for processing raw acceleration data to estimate sleep health parameters. We aimed to 1) assess the performance of three sleep algorithms within GGIR against PSG for detecting sleep/wake in clinically referred, typically-developing children (criterion validity); and 2) describe GGIR-derived sleep estimates from typically developing children enrolled in multiple cohort studies (face validity). For criterion evaluation, children (8-16y, N=30) wore an actigraphy device for one night during in-lab polysomnography with performance assessed using epoch-by-epoch analyses. For face validity evaluation, four community/free living datasets were used: 1) BMAYC (3-5y, N=310), 2) SSS (5-8y, N=118), 3) S-Grow2 (12-13y; N=291) and 4) ELEMENT (9-18y; N=543). All raw acceleration data were processed using GGIR (v.3.0-0) with the Cole-Kripke (CK), Sadeh (S), and van Hees (vH) algorithm settings. Following the in-lab test, 60% of children were diagnosed with mild to severe obstructive sleep apnea (OSA). For criterion evaluation, the 30-s epoch-by-epoch analyses revealed that average balanced accuracies were 0.80 (Sensitivity=0.80; Specificity=0.79), 0.76 (Sensitivity=0.86; Specificity=0.65), and 0.67 (Sensitivity=0.95, Specificity=0.39) for GGIR-CK, GGIR-vH, and GGIR-S, respectively. For face validity evaluation, sleep estimates mirrored the in-lab performance metrics (e.g., sleep duration estimates were similar when using GGIR-CK and GGIR-VH but approximately one hour longer when using GGIR-S). The in-lab performance metrics, from typically-developing children with and without OSA, and cohort-based descriptive statistics from samples of typically-developing children, provide benchmark data to guide investigators on the suitability of GGIR for automated processing of raw acceleration data for pediatric sleep estimation.