Generalizability of A Neural Network Model for Circadian Phase Prediction in Real-World Conditions

Julia E. Stone, Andrew J.K. Phillips, Suzanne Ftouni, Michelle Magee, Mark Howard, Steven W. Lockley, Tracey L. Sletten, Clare Anderson, Shantha M.W. Rajaratnam, Svetlana Postnova

Research output: Contribution to journalArticleResearchpeer-review

26 Citations (Scopus)


A neural network model was previously developed to predict melatonin rhythms accurately from blue light and skin temperature recordings in individuals on a fixed sleep schedule. This study aimed to test the generalizability of the model to other sleep schedules, including rotating shift work. Ambulatory wrist blue light irradiance and skin temperature data were collected in 16 healthy individuals on fixed and habitual sleep schedules, and 28 rotating shift workers. Artificial neural network models were trained to predict the circadian rhythm of (i) salivary melatonin on a fixed sleep schedule; (ii) urinary aMT6s on both fixed and habitual sleep schedules, including shift workers on a diurnal schedule; and (iii) urinary aMT6s in rotating shift workers on a night shift schedule. To determine predicted circadian phase, center of gravity of the fitted bimodal skewed baseline cosine curve was used for melatonin, and acrophase of the cosine curve for aMT6s. On a fixed sleep schedule, the model predicted melatonin phase to within ± 1 hour in 67% and ± 1.5 hours in 100% of participants, with mean absolute error of 41 ± 32 minutes. On diurnal schedules, including shift workers, the model predicted aMT6s acrophase to within ± 1 hour in 66% and ± 2 hours in 87% of participants, with mean absolute error of 63 ± 67 minutes. On night shift schedules, the model predicted aMT6s acrophase to within ± 1 hour in 42% and ± 2 hours in 53% of participants, with mean absolute error of 143 ± 155 minutes. Prediction accuracy was similar when using either 1 (wrist) or 11 skin temperature sensor inputs. These findings demonstrate that the model can predict circadian timing to within ± 2 hours for the vast majority of individuals on diurnal schedules, using blue light and a single temperature sensor. However, this approach did not generalize to night shift conditions.

Original languageEnglish
Article number11001
Number of pages17
JournalScientific Reports
Publication statusPublished - 29 Jul 2019


  • circadian rhythms and sleep
  • machine learning

Cite this