Figure 2
From: Generalizability of A Neural Network Model for Circadian Phase Prediction in Real-World Conditions
(A) Schematic of neural network structure reproduced from15. Blue light irradiance and skin temperature data were inputs to a two-layer perceptron with a hidden layer of 5 neurons, plus a bias term. The input variables included a light variable (blue irradiance) with lags of 0, 1, 2, …, 24 h with either six skin temperature variables (shoulders, sternum, wrists, thighs, ankles, feet) or one skin temperature variable (wrist) with lags of 0, 1, 2, …, 5 h, plus a bias term. Networks using light plus 6 skin temperature variables had 115 inputs, plus a bias term, resulting in a total of 586 adjustable weights. Networks using light and 1 skin temperature variable had 60 inputs, plus a bias term, resulting in a total of 311 adjustable weights. Output was either predicted melatonin concentration or aMT6s excretion rate. (B) Schematic of models trained using cross-validation by sleep schedule: fixed sleep trained on melatonin profiles using light with 1 or 11 skin temperature sensors; fixed and habitual sleep trained on aMT6s profiles using light and 1 or 11 skin temperature sensors; diurnal sleep (fixed, habitual, and shift work day schedule datasets) trained on aMT6s profiles using light and 1 skin temperature sensor; night shift schedule trained on aMT6s using light and 1 skin temperature sensor. Additionally, the fixed and habitual sleep network using light and 1 skin temperature sensor was independently tested on shift work datasets (day and night shift schedules); and the diurnal sleep model was independently tested on the night shift work datasets. Colors indicate model performance based on mean absolute error: dark green < 55 minutes; light green < 65 minutes; light orange < 85 minutes; dark orange < 145 minutes; red > 145 and < 440 minutes.
