1. Introduction
Circadian rhythm plays a significant role in our lives as it regulates various physiological processes, biological features, and behaviors. Circadian rhythm has evolved to oscillate with a period of about 24 h to keep in synchronization with the light-dark cycle on earth. Plasma melatonin suppression [
1], circadian gene expression [
2], and cognitive performance [
3] demonstrate spectral sensitivity to light, i.e., they are more sensitive to short-wavelength light [
4]. Numerous circadian rhythm models have been formulated and analyzed based on the dynamics of circadian-related physiological features [
5]. Light has been widely used as an external input in circadian rhythm regulation in previous works [
6]: Diekman constructed a one-dimensional map to demonstrate the process that the light-dark cycle entrains circadian rhythm dynamics [
7]; Booth formulated a simplified dynamical system to investigate the dynamics in circadian regulation of sleep-wake cycle [
8]; Abel and Doyle applied the model predictive control method on the dynamics of the mammalian circadian clock and showed theoretical results of circadian phase resetting via light [
9]; Julius and Forger applied variational calculus on humans’ core body temperature dynamic model, and showed the optimized timing of light exposure dramatically reduced the circadian rhythm entrainment time in various time shifts [
10,
11]. Cognitive impairments, because of circadian disturbance, sleep disorder, and extended wakefulness, are common in modern society and bring negative impacts on our health. Lower concentration and alertness also cause harmful consequences, such as a decline in efficiency, traffic accident, and other failures in jobs and lives. In our common lives, we may give higher priority to the enhancement of our cognitive performance, alertness, and vigilance instead of reducing the time costs of entraining the misaligned circadian rhythm to a normal one. For example, transmeridian travelers who want to catch an important conference or have important personal matters at the destination, night-shift workers who need to relieve fatigue for a whole night, or soldiers with an important mission want to keep fresh. The light exposure should be appropriately set in terms of the dynamics of subjects’ cognitive functions to relieve cognitive impairments.
Previous research has shown that the dynamics of humans’ neurobehavior, cognitive performance, fatigue, and alertness are jointly determined or mutually affected by several processes, including circadian rhythm and sleep-wake cycle [
12]. Empirical data showed that humans’ sleep-wake cycle is closely linked to the circadian rhythm and sleep homeostasis, called
Process C and
Process S, respectively [
13]. Two-process models have been formulated to simulate the sleep-wake cycle and sleep propensity [
14]. However, some experimental results indicated that the alertness level starts at a very low level at the wake-onset time and then increases rapidly. This phenomenon is attributed to sleep inertia, which is ignored in two-process models. To incorporate its impacts, the dynamics of the sleep inertia (usually called
Process W) were formulated [
15,
16] and the three-process models, combining the two-process model with the sleep inertia, were used in several papers [
17,
18] to predict cognitive functions and alertness.
Several existing two-process and three-process models took the circadian process as predetermined skewed sinusoidal waveforms without consideration of light exposure. However, experimental studies indicated light intensity and spectrum during night-shift has a great influence on the alertness level of shift workers: the work in [
19] showed exposure to bright light during nighttime slowed down the decline in alertness during the night-shift. Another study in [
4] showed that subjective alertness at 8 am (the end of night-shift) was less impaired by staying in darkness or wearing circadian light-blocking goggles filtering light spectrum less than 480 mm than staying in unfiltered light exposure during the night-shift between 8 pm and 8 am. Achermann formulated a three-process model with Process C represented by a simpler core body temperature (CBT) model, which contains lighting impacts on its dynamics [
16]. However, the CBT model had been developed several times to incorporate higher-order nonlinearities to capture the latest empirical data and experimental observation that amplitude recovery of circadian oscillator is slower around the singular regions than near the limit cycle [
10], which has been ignored in the three-process model formulation. Previous literature optimized the sleep schedule to regulate cognitive performance [
20]. However, this study ignored the constraints of sleep propensity [
16] and deliberately changed the sleep and wake times. In addition, the existing study of light-based circadian entrainment [
10,
11] used light exposure to entrain Process C but ignored constraints of Process S and the sleep-wake cycle. Results of these studies also contain long wakefulness duration (or long duration under light exposure) beyond normal sleep times and excessive sleep propensity, which could be impractical for human subjects.
In this paper, we develop the three-process model and study the alertness optimization problem based on this model. The main contributions of this paper are as follows:
This paper simulates the subjective alertness and sleepiness with various light inputs and sleep schedules by a three-process hybrid dynamic model. Different from previous work [
14,
16], the latest dynamic model of CBT [
10,
11] is used to represent the Process C in the three-process model to incorporate the lighting effects and higher-order nonlinear terms of circadian rhythm on cognitive performance;
This paper proposes a tunable sleep schedule with appropriate sleep/wake time constraints in the three-process model to avoid excessive sleepiness and guarantee the optimal sleep schedule and light exposure are more practical for human subjects to follow;
The numerical solution method of optimal light exposure and sleep schedule in alertness optimization problems is proposed based on the calculus of variations, and the results demonstrate that, compared with empirical data, the optimized light and sleep schedule relieve humans’ cognitive impairment caused by shift works and other issues.
In
Section 2, the three-process hybrid dynamic model with light input is formulated and validated, and the problem formulations of several alertness optimization cases with light and sleep constraints are proposed. In
Section 3, the numerical algorithm for solving the alertness optimization problems is proposed. In
Section 4, the solutions of alertness optimization problems are demonstrated and compared with previous experimental data to discuss the regulation capability of light and sleep on human cognitive performance. Finally, the conclusion is summarized in
Section 5.
2. Mathematic Model and Problem Formulation
We represent the Process C of the three-process model based on the CBT circadian rhythm dynamic model in [
10,
11]. The mathematical formulation of the CBT model is listed as follows:
where parameter values are given as
h
−1,
h
−1,
lux,
,
,
h
−1,
,
h,
h
−1, and
h [
10,
11]. The term
I represents the light input with a unit of
.
represents the sleep state, i.e.,
means the subject is sleeping at time
t and
means the subject is awake. The state
n in Equation (
1) is the receptor state used to simulate the time-varying relation between the light input
I and circadian drive
u. The state
x is the normalized state of CBT and
is a complementary state.
The sleep homeostatic oscillator, i.e., Process S, represents the accumulation of substance that generates the sleep drive [
16]. The dynamics of sleep homeostasis
H are given as
where the time constants during sleep and wake periods are
h and
h, respectively [
16]. Sleep propensity
is jointly affected by both Process S and Process C. We define its value as
where
. If a subject falls asleep and wakes up spontaneously, the change in his sleep state
is fully determined by
(and thus, essentially, by
and
). The dynamic of the spontaneous sleep schedule is expressed as
where
is the time just before
t, i.e.,
means the left limit of
. The upper and lower bounds of spontaneous sleep schedule are
and
, respectively [
16]. The dynamic of sleep inertia
W with sleep state is given as
where the time constant
h [
16,
20]. At the wake (onset) time, the value of sleep inertia
W is reset to 0.32. The three-process model used in this paper consists Process C in Equations (
1)–(
4), Process S in Equation (
5), and Process W in Equation (
8). We define the value of alertness and sleepiness level, denoted as
and
, as
The alertness level during sleep is set (by definition) as 0. Under a periodic light-dark cycle, the state of the three-process model, which is denoted as
may run periodically in the same period, i.e., the three-process model keeps synchronization with the light-dark cycle. We define a 24 h periodic light-dark cycle in the following form
which simply simulates the reference daily light-dark cycle. We set that the sunrise/light-on time
corresponds to 6 am and the sunset/light-off time
corresponds to 10 pm. Under the periodic light-dark cycle in Equation (
11) and the spontaneous sleep schedule in Equation (
7), the state of the three-process model runs periodically in the same period. The stable periodic solution of the three-process model is the reference state
, denoted as
which, along with the corresponding sleep state
and alertness level, is plotted in
Figure 1. We can observe the daily alertness starts from a low level at the wake time, as a result of sleep inertia, then increases rapidly and reaches the daily maximum value at about 9 am. Then alertness declines steadily in the afternoon and evening until sleep begins, which agrees with daily profiles of cognitive performance and subjective sleepiness [
21,
22]. Note the post-lunch dip in alertness is widely experienced during the early afternoon in human beings, especially habitual nappers. Experimental observations of university students showed the alertness of these subjects show a progressive increase and peaks before noon, and slightly declines around noon to 4 pm. Then alertness increases again and reaches another peak, finally demonstrating a sharp drop to sleep [
23]. One way to simulate the biphasic sleep pattern and post-lunch dips is by tuning the upper and lower thresholds in Equation (
7) to generate one short and one long daily sleep bouts [
14]. Another way for post-lunch dips’ simulation is adding another process (usually called Process U) with a period of 12 h [
24]. Sleep and nap habits, as well as the extent of post-lunch dip, vary largely with persons and regions, here we study the alertness improvement problem with external light and sleep regulation, in which we assume the sleep is consolidated into one continuous nocturnal bout by long-term, daily 16-h photoperiod [
22] and we only tune this single bout in the tunable sleep schedule. Experimental data of night-shift we compared in this paper comes from constant routine protocol and photoperiod experiments, which prohibited daytime naps and no obvious post-lunch dip was found in these experiments. Therefore, this model is sufficient for our simulation and study.
Sleep propensity and recovery sleep duration (after sleep deprivation) both do not monotonically increase with the time length of sleep deprivation (duration of keeping awake after spontaneous sleep time) [
13], which can be simulated and explained by the three-process model. The simulation results of the three-process model and experimental data in [
25] in
Figure 2 both indicate the recovery sleep duration firstly shows a decline as the sleep deprivation increases from 0 to about 14 h, and then shows a discontinuous increase to a duration longer than normal night sleep when sleep deprivation is around 18 h.
Figure 3 shows the simulated sleep propensity during sleep deprivation with various time lengths and the following recovery sleep duration. After spontaneous sleep time, the sleep propensity firstly grows with the duration of sleep deprivation but then shows a decline when the sleep deprivation duration increases to more than 8 h, as a result of the increase in
x from Process C reducing circadian sleep drive. This simulation is consistent with the fact that subjects may feel less fatigue and cannot fall asleep immediately once the night-shift work ends or having kept awake for several hours beyond normal sleep time. The blue curve of
in
Figure 3 indicates the sleep homeostasis at wake time after sleep deprivation and determines the recovery sleep duration based on Equation (
5). As the sleep deprivation duration increase to 14 h, the corresponding homeostasis at wake time increases gradually and shortens the recovery sleep duration after sleep deprivation. When the sleep deprivation duration increases from 14 to about 18 h, the homeostasis at the following wake time shows a sharp drop, resulting in a sudden growth in recovery sleep duration. Then,
increases again and reduces the recovery sleep duration as the sleep deprivation continues to increase to more than 20 h.
To further validate the accuracy of the three-process model in the prediction of subjective alertness and sleepiness, we introduce and simulate two experimental protocols in the previous reference. The first experiment asked subjects to remain awake for an extended period of 36–60 h under constant light (150 lux), which is called the
constant routine experiment protocol [
26]. Subjective alertness was assessed three times every hour using a linear 100-mm bipolar visual-analogue scale (VAS) during the experiment, which is a questionnaire that asks subjects to indicate their alertness on a visual scale of 100 mm length. We use the three-process model to simulate this experimental protocol and predict the subjective alertness, as shown in
Figure 4. Note that the alertness value
in the three–process model has no unit. To fit the empirical units (0–100 mm), we scale the predicted subjective alertness as
where 56.18 mm and 0.85 are the differences between maximum and minimum in empirical data and simulation results, 45.07 mm and 0.26 are the average values in empirical data and simulation results. The average mean square error between predicted alertness and the average empirical data is 7.91 mm, smaller than the standard deviation of the empirical data (9.80 mm). The coefficient of determination of the predicted alertness to the empirical data is 0.811. As shown in
Figure 4, the predicted alertness from the three-process model shows good correspondence to the subjective alertness in empirical data.
The second experiment, called
photoperiod experiment in [
22], asked subjects to stay in two different photoperiods: the long photoperiod has 16 h daily light in one week and the short has 10 h daily light in four weeks. The 24 h profiles of sleepiness were measured during a constant routine protocol with dim light less than 1 lux after each photoperiod by Stanford Sleepiness Scale ratings. To fit the Stanford Sleepiness Scale (SSS), the predicted sleepiness from the three-process model is scaled as
The corresponding values of alertness in VAS and sleepiness in SSS with respect to
and
in the three-process model are listed in
Table 1. The comparison of predicted sleepiness from the three-process model and empirical data is shown as
Figure 5. The average mean square deviations of the predicted sleepiness from the empirical data are 0.68 and 0.51 in the long and short photoperiods, smaller than the standard deviations (1.08 and 0.79) in empirical data. The coefficients of determination of the predicted sleepiness to the empirical data in long and short photoperiods are 0.777 and 0.714, respectively. The predicted sleepiness values are closely consistent with the empirical data in both long and short photoperiod schedules, indicating that the three-process model is accurate in subjective alertness and sleepiness rating prediction.
2.1. Mission Alertness Optimization
Inspired by the study in [
20], we propose the mission alertness optimization problem, that is, to improve the subjective alertness during certain mission periods by optimizing the light exposure and sleep schedule in advance. Note that before the mission time, the subject can partially adapt his sleep schedule instead of following the spontaneous sleep schedule in Equation (
7). This sleep schedule, with tunable sleep and wake times, is called the
tunable sleep schedule in this paper. The sleep state in the tunable sleep schedule is expressed as
with sleep and wake times constraint as
where
and
are the
ith sleep time and
jth wake time,
and
are the total sleep and wake time from the beginning time of circadian regulation to the start of mission.
and
indicate the subject cannot fall asleep earlier than the spontaneous sleep time or wake up later than the spontaneous wake time. Meanwhile, the inequality condition
is used to simulate the cases in our daily lives that we may fall asleep later than normal sleep times, as a result of night-shift, personal reasons or other external disturbance.
corresponds to the cases where we may wake up earlier than spontaneous wake times by setting a morning alarm.
and
are introduced, which were ignored in [
20] and led to 40-h continuous wakefulness, as extra constraints to avoid excessive long wake duration and guarantee the daily sleep duration. The upper thresholds
and
could be set based on personal sleep habits. Individuals show various sleep types with daily sleep duration varying from 6 to 9 h [
27]. In this paper, we choose these values
and
, with sleep propensity at sleep and wake times shown as the red region in
Figure 3, as an example to show that appropriately tuning the sleep schedule could improve cognitive performance without dramatically shortening the daily sleep duration.
Mission Alertness Optimization Problem: Given the initial value of state
the state equation of the three-process model expressed as
and the tunable sleep schedule in Equation (
14), the mission alertness optimization problem is formulated as
which subjects to the light intensity constraint
, sleep state during mission
between
and
, sleep time constraints before mission
,
, and wake time constraints before mission
,
.
is the beginning time of the mission and
is the final time of the mission. The initial state in Equation (
16) is defined based on the periodic solution
.
is the maximum light intensity used before and during the mission. In practical use, the light intensity constraint could be set as a time-varying constraint, i.e.,
, with
valued based on personal preference, health care, work demand, specifications of lighting equipment and other limits. For simplicity and to show the potential of using light to improve alertness, we set
as a constant in this and following optimization problems in
Section 2.2 and
Section 2.3.
2.2. Night-Shift Alertness Optimization
As mentioned in
Section 1, the nighttime light has great impacts on the alertness level of night-shift workers. Based on the experimental protocol in [
19], the night-shift takes place between 10 pm and 8 am. We simulate two cases of this night-shift work: (1) A night-shift worker remains under a bright light
lux between 12 am and 4 am and stays under dim light with 150 lux for the rest of the night shift; (2) A night-shift worker stays under 150 lux dim light during the whole night-shift. The simulation results of these two cases are shown as the blue and red curves in
Figure 6. The alertness level of the shift worker in the first case is larger during the interval between 1 am and 6 am. This result is consistent with the experimental results in [
19] and implies that timed exposure to bright light during the night shift partially improves alertness.
A similar experimental study of lighting impacts on night-shift alertness has been shown in [
4], where the experimental subjects were divided into three groups. One stayed in the darkness, one was given bright light of 1000 lux, and one wore circadian light-blocking goggles between 8 pm and 8 am. The circadian light-blocking goggles filter out all light spectrum with the wavelength less than 480 nm. The experimental results indicate that the alertness levels of the group with the goggles are very close to those of the group that stayed in darkness. After blocking the short-wavelength light, the equivalent light intensity for the circadian rhythm of night-shift workers is 0 lux. Based on the experiments, we simulate two cases in shift work: the first worker stays in darkness (or wears goggles) with
lux during the whole night-shift from 8 pm to 8 am, the second case stays under the bright light of
lux during this night-shift. Both simulation results in
Figure 7 and experimental results in [
4] show that, for the case in darkness during the night-shift, alertness decreases more rapidly in the first several hours of the night-shift, but is slightly larger than the case in bright light at the end of the night-shift.
Based on these experimental studies, we formulate a night-shift alertness optimization problem, given as below:
Night-shift Alertness Optimization Problem: Given the initial condition in Equation (
16), system dynamics in (
17), the optimal light input is determined to maximize the cumulative alertness level during the night-shift. This optimization problem is formulated in the following form:
subjects to the light input constraint
, sleep state
between 0 and
, where
corresponds to the beginning of the night-shift,
is the final time of the night-shift.
2.3. Consecutive Night-Shift Alertness Optimization
Another case considered in this paper comes from the experimental studies in [
28]: These experimental studies asked subjects to remain awake for several consecutive night-shifts and investigated the cognitive impairments in days following night-shifts. These subjects were separated into two groups: control group and treatment group. In the control group, the subjects were assigned ordinary indoor light with 150 lux during the second through fifth nights, while the subjects in the treatment group received bright light with 7000∼12,000 lux. This experiment showed exposure to bright light during several consecutive night-shifts led to a significant shift (about 9 h) in the patterns of core body temperature and subjective assessment of alertness, but could result in serious cognitive impairments in the following day-shift. In this case, we determine the optimal light input and sleep schedule to maximize the cumulative (or average) alertness during either the consecutive night-shifts or the day after these night-shifts, or both of them.
Consecutive Night-shift Alertness Optimization Problem: Given the boundary condition in Equation (
16), the system dynamics in Equation (
17), and the tunable sleep schedule in Equation (
14), we calculate the optimal light input and sleep schedule by solving the consecutive night-shift alertness optimization problem, formulated as below:
which subjects to the light input constraint
, the sleep state during shift working time
, sleep time constraints outside the shift working time
, and wake time constraints outside the shift working time
.
represents the set of all night-shift working periods and
represents the day-shift period this week. The positive constant
and
represent the weight of alertness during night-shifts and the day-shift in the objective function.
3. Solution Strategies and Algorithms
We apply the variational calculus and functional gradient descent algorithms to solve the alertness optimization problems proposed in the previous section. Equation (
5) indicates that, in the three-process model, the state
follows different dynamics when the subject is sleeping (
) and awake (
). Therefore, the alertness optimization problems should be solved on a hybrid dynamical system with two kinds of modes: sleep mode and wake mode. The dynamics of the hybrid dynamical system in the
ith mode are expressed in the following form:
where
N is the total number of modes,
is the final time and
. Assume that the sleep-wake switching time
follows the constraint given in the following form:
where
represents the feasible region of switching time based on problem formulation. It could be defined based on the constraint of the switching condition in Equation (
15) and the sleep and wake time constraints in the problem formulation. We formulate the objective cost function as
where the integrand
maps
into a scalar. To include the constraints of the dynamical system in Equation (
21) into the cost function, we introduce the Lagrange multipliers
and formulate the augmented cost function as
To determine the extreme value of the cost function, we add perturbation terms of
and
into the switching time
and light
, respectively, resulting in perturbed state
and a perturbed augmented cost function, given as
where
is a small scalar,
represents a higher order term of
.
and
represent first-order (linear) variations of
and
in
and
, given as
We subtract cost function in Equation (
25) by (
24) and take first-order variation of the augmented cost by dividing their difference by
and taking its limit with
, then the first-order variation of the augmented cost is given explicitly as
The selection of Lagrange multipliers
does not affect the result of the augmented cost function. We choose
in the following forms to offset terms of
in
:
where
represents the multiplier of the
jth state in
. As
,
,
,
are all continuous at all switching times, i.e.,
for
j={1,2,3,4}, then we have Equation (
28). The sleep inertia
is reset to 0.32 and discontinuous at wake times, as mentioned in
Section 2. We set the multiplier value
follows (
29) with
at wake times, which will be used in backward simulation of Equation (
27). After some simplifications, the first-order variation of the augmented cost function with respect to
and
is given as
This equation implies the linear variation of the augmented cost function resulting from a small perturbation in light input and switching times. As the cost function
J is equal to
with (
21) satisfied, the gradients of
J with respect to
I and
are given as
Based on the gradient descent method, if we represent the light input and
ith switching time at the
jth iteration as
and
, their values can be updated by:
where
and
are the updating steps for light input and wake/sleep time and determined by a line search in the simulation. It should be guaranteed in the gradient descent process that
satisfies the constraints in the problem formulation. Steps of the proposed gradient descent process in the calculation of the optimal light and sleep schedule for alertness optimization are listed below:
Step 1: Set and choose an initial guess of the light input and the sleep schedule (i.e., );
Step 2: Integrate the state equation in (
21) forward and determine the state
, alertness
, and objective function value
;
Step 3: Integrate (
27) backward with the terminal condition in (
26), using switching conditions in (
28), (
29) at switching times, to get
;
Step 4: Determine the gradient of the objective function to
and the sleep/wake times
based on Equations (
30) and (
31), update
and
based on Equations (
32) and (
33);
Step 5: Increment
j by 1. Repeat from
Step 2 until
or
We define the cost function in (
23) as
, and set the dynamics in (
21) based on the three-process model. The stable solutions of this algorithm, denoted as
and
, correspond to the optimal light input and sleep schedules that maximize the alertness. As discussed in [
10,
11], the optimal light in circadian regulation is bang-off control without singular regions, i.e.,
or
for
. In the following figures, we use red regions to demonstrate the optimal light exposure time.