First published: 28 May 2018
SummarySleep loss, which affects about one‐third of the US population, can severely impair physical and neurobehavioural performance. Although caffeine, the most widely used stimulant in the world, can mitigate these effects, currently there are no tools to guide the timing and amount of caffeine consumption to optimize its benefits. In this work, we provide an optimization algorithm, suited for mobile computing platforms, to determine when and how much caffeine to consume, so as to safely maximize neurobehavioural performance at the desired time of the day, under any sleep‐loss condition. The algorithm is based on our previously validated Unified Model of Performance, which predicts the effect of caffeine consumption on a psychomotor vigilance task. We assessed the algorithm by comparing the caffeine‐dosing strategies (timing and amount) it identified with the dosing strategies used in four experimental studies, involving total and partial sleep loss. Through computer simulations, we showed that the algorithm yielded caffeine‐dosing strategies that enhanced performance of the predicted psychomotor vigilance task by up to 64% while using the same total amount of caffeine as in the original studies. In addition, the algorithm identified strategies that resulted in equivalent performance to that in the experimental studies while reducing caffeine consumption by up to 65%. Our work provides the first quantitative caffeine optimization tool for designing effective strategies to maximize neurobehavioural performance and to avoid excessive caffeine consumption during any arbitrary sleep‐loss condition.
...MUCH MORESleep loss, which is a common stressor for both civilians and military personnel, can severely impair cognitive and physical performance, and thereby diminish productivity and compromise safety. Several studies have demonstrated that, when safely used, caffeine can help to sustain cognitive performance during prolonged periods of restricted sleep (Doty et al., 2017; Kamimori et al., 2015; Killgore et al., 2008; Mclellan, Bell, & Kamimori, 2004; Mclellan et al., 2005; Wesensten, Killgore, & Balkin, 2005). However, these investigations offer caffeine countermeasure guidance that is study‐specific, and which cannot be readily adaptable to any arbitrary sleep‐loss condition. Providing a foundation for addressing this need, our group has previously developed and validated a mathematical model, the unified model of performance (UMP), which can predict the effects of sleep loss and caffeine, as a function of time of day, on objective measures of neurobehavioural performance (i.e. the psychomotor vigilance task, PVT) across a wide range of sleep–wake schedules and caffeine doses (Ramakrishnan et al., 2013, 2014, 2016). More recently, we have built upon the UMP to develop the open‐access Web tool 2B‐Alert (Reifman et al., 2016), a decision aid to help users design sleep studies and work schedules, and the smartphone 2B‐Alert app (Reifman et al., 2017) for real‐time, individualized performance prediction (Liu, Ramakrishnan, Laxminarayan, Balkin, & Reifman, 2017).
In this work, our goal was to develop a computational tool to provide, in real time, effective caffeine‐dosing strategies for any arbitrary sleep‐loss condition. Once incorporated into a mobile computing device, such a tool could provide customized caffeine‐consumption guidance to, for example, sustain the attention of sleep‐deprived military personnel. To this end, using the predictive ability of the UMP, we formulated an optimization problem to determine when and how much caffeine to consume, so as to safely maximize neurobehavioural performance at the desired time of the day for the desired duration. To solve this problem, we developed an efficient optimization algorithm that was able to find near‐optimal solutions in real time. We assessed the optimization algorithm by comparing the effects of its predicted caffeine‐dosing (timing and amount) strategies with those obtained in four experimental studies previously used to validate the UMP (Ramakrishnan et al., 2016). In particular, we obtained caffeine‐dosing strategies that enhanced PVT performance while using the same total amount of caffeine as in the original studies, and strategies that yielded equivalent levels of performance as in the original studies while reducing caffeine consumption.
2.1 The unified model of performanceOur goal was to find a caffeine‐dosing strategy that would minimize neurobehavioural performance impairment based on the PVT for a given sleep–wake schedule. To this end, we sought to minimize the objective function Z (Equation 1, Table 1), which considers both the area under the UMP‐predicted PVT mean RT curve (AUCC) that is above the baseline, and the worst performance (WPC) (i.e. the difference between the peak of the mean RT curve and the baseline) (Figure 1). As the baseline mean RT, we used the highest predicted value of mean RT when an average individual has no sleep debt, wakes up at 07:00 and is awake for 16 hr. We normalized AUCC and WPC by the corresponding values for the predicted mean RT curve without caffeine consumption, AUCNC and WPNC, respectively. In addition, we included a penalty term in Z to limit the accumulation of caffeine in the blood [C(ti, Di)], which could result in unsafe consumption (Killgore et al., 2008). This term penalizes Z when the maximum level of caffeine in the blood is higher than the maximum level (Cmax) achieved by a single 400‐mg dose (Institute of Medicine, 2001). (Note that the value of Cmax can be readily changed in the algorithm.) Hence, without considering the penalty term, Z varies from 0 (for a strategy that consistently maintains the mean RT below the baseline) to 100 (for a strategy that is no better than using no caffeine). Therefore, the smaller the value of Z, the better is the dosing strategy....
The UMP has two components. The first, based on Borbély's two‐process model (Borbely, 1982), describes performance as a function of the circadian cycle and a homeostatic process. The second is a pharmacokinetic and pharmacodynamic model, which estimates the caffeine level in the blood and predicts the duration and magnitude of the effect of caffeine intake on neurobehavioural performance. For a given sleep–wake schedule and caffeine consumption strategy, which constitute the inputs to the model, the UMP predicts the PVT mean response time (RT) for an ‘average’ individual. We refer the reader to Ramakrishnan et al. (2016) for detailed descriptions of the UMP, the parameter estimation process and model validation. We have provided the UMP equations and parameter values in Section I of the Supporting Information.
2.2 Optimization problem
Or 10 page PDF