Sleep assessment using accelerometry: Not all algorithms are equal

Abstract Study objectives Accelerometry devices are increasingly used to assess sleep. However, whether different algorithms provide consistent estimates remains uncertain. This study compared sleep parameters derived from two accelerometry‐based algorithms and a self‐reported sleep journal. Methods Data were obtained from the second (2014–2017; n = 2724; 53.3% women; 62.0 ± 10.0 years) and third (2018–2021; n = 2087; 53.5% women; 65.1 ± 9.6 years) follow‐ups of the CoLaus|PsyCoLaus study. Seven‐day accelerometry data were analysed using GGIR (R‐based) and MACRO (Excel‐based) algorithms. A subset of participants also completed an ecological momentary assessment (EMA) sleep diary. Sleep onset (<22:00, 22:00–23:59, ≥24:00), average sleep duration, and average sleep efficiency were compared. Results In both surveys, GGIR estimated longer sleep duration than MACRO (406 ± 103 vs. 378 ± 79 min; 397 ± 60 vs. 366 ± 84 min; p < 0.001). Sleep duration correlations were moderate (Spearman r = 0.592) with Lin's concordance correlation of 0.269 and 0.513, respectively. GGIR estimates were closer to EMA than MACRO. For sleep onset, GGIR classified >80% of participants before 22:00, compared with 38%–64% (MACRO) and 8%–12% (EMA). GGIR also provided higher sleep efficiency (72 ± 17 vs. 70 ± 14%; 70 ± 7 vs. 67 ± 15%; p < 0.001; r = 0.383). Conclusion When assessing sleep from accelerometry, algorithm choice strongly influences estimates, highlighting the need for standardisation.