We are evaluating how well individuals perform during repeated cycles of chronic sleep restriction in the Human Exploration Research Analog (HERA). Crews of four complete the psychomotor vigilance task (PVT) up to five times every third day during the 45-day mission. They also collect information on their sleepiness and mood and baseline information about their personality and general fatigue levels. Our goal is to determine how well bio-mathematical models designed to predict performance impairment correlate to the objective measures of performance that we are collecting in each mission. In addition, we have added graphics of model predictions to the crew scheduling tool in order to determine how useful crewmembers find the model information for making planning decisions in real time. We are also evaluating how well pre-mission personality characteristics predict in-mission mood.
To date, we have collected data from three complete missions (n = 12), and have a partial dataset from a mission that was evacuated due to an environmental hazard surrounding the habitat (n = 4). We plan to collect data from one final crew in this campaign, during the spring of 2018, for a total of n =16 participants. We have not analyzed any of the data as the study is ongoing. Specific information about our data analysis approach is listed below for each of our aims.
Specific Aim 1 is focused on determining how well the proposed model(s) estimate sleep deprivation and circadian outcomes traditionally assessed by the clinical gold standard, PVT. To address this aim, our study biostatistician will compare the gold standard PVT data to the model predictions. The PVT includes several outcome metrics; however, our primary focus will be two of the PVT’s most common performance indicators--lapses > 500 ms and the inverse mean reaction time (1/RT). Each of the candidate models also estimates these to outcomes, making it possible to compare the gold standard measures to each candidate surrogate measure. More specifically, we will compute root mean square error (RMSE) values and calculate Spearman correlations to determine how well the model predictions fit the observed data. We will repeat our analysis using the outcomes from a robotic trainer task in order to evaluate whether the model predictions are linked with decrements in operational performance.
A sub-aim of Specific Aim 1 is to evaluate the potential impact of individual characteristics generated during screening to PVT performance. In order to accomplish this aim, we will regress incorporate the individual subject characteristics collected at screening (sex, age, morningness-eveningness score, fatigue severity scale score, Epworth Sleepiness Score, Tiredness Symptom Scale (TSS) domain, personality traits, and anxiety traits) on our primary PVT performance outcomes using mixed-effects regression methods that incorporate within-subjects variance components. In addition, we will calculate the squared semi-partial correlation coefficients in order to evaluate the independent variance contributions (after removing shared covariance among other predictors) of each of the subject characteristics.
Specific Aim 2 is aimed at qualitatively evaluating the usability, acceptability, and feasibility of using the models within an operational environment. In order to accomplish this aim, we will test the effectiveness of providing modeling software to the crewmembers by conducting user interface sessions while crewmembers describe their understanding of the model and interface using a think-aloud approach. We will evaluate the usefulness of providing modeling information to crewmembers by collecting usability ratings following each model usability session.
Specific Aim 3 is aimed at incorporating the models into Playbook, the crew’s self-scheduling tool. In order to qualitatively evaluate this aim, we will use a Wizard-of-Oz user testing approach to integrate the model predictions into the scheduling software. In addition, we will compare the usability, feasibility, and acceptability of using Playbook to using existing model software by comparing the usability metrics that we obtain during the implementation of the existing model software to the usability metrics that we obtain during the Wizard-of-Oz implementation of the Playbook software.