Available online at www.sciencedirect.com
Nurs Outlook 62 (2014) 313e321
www.nursingoutlook.org
A systems science approach to fatigue management in research and health care Kathryn A. Lee, PhD, RN, FAAN, CBSMa,*, Oliwier Dziadkowiec, PhDb, Paula Meek, PhD, RN, FAANc b
a School of Nursing, University of California, San Francisco, San Francisco CA Center for Nursing Research, College of Nursing, University of Colorado at Denver, Aurora, CO c College of Nursing, University of Colorado at Denver, Aurora, CO
article info
abstract
Article history: Received 25 June 2014 Revised 2 July 2014 Accepted 3 July 2014 Available online 9 July 2014
The purpose of this article was to highlight innovative analytic approaches in symptom science. Complex systems modeling is discussed using fatigue as an example. Fatigue is a common symptom among individuals of any age. It can be acute or chronic, and it can vary across the day and on weekends compared with weekdays. Fatigue can overlap with other symptoms, has many dimensions, and impacts daily function as well as society at large. With the complexities surrounding symptom science, innovative models are needed to advance our understanding of factors within the person, contextual and situational factors, and workplace or health care system factors that impact the symptom experience. Advances in methodologies, such as complex systems modeling, allow for more innovative methods to study the complexities of the symptom experience, design better ways to intervene and manage symptoms, and ultimately improve outcomes related to symptom management, quality of life, and health care utilization.
Keywords: Agent-based modeling Big data Complex systems modeling Markov model Precision medicine System dynamics
Cite this article: Lee, K. A., Dziadkowiec, O., & Meek, P. (2014, OCTOBER). A systems science approach to fatigue management in research and health care. Nursing Outlook, 62(5), 313-321. http://dx.doi.org/ 10.1016/j.outlook.2014.07.002.
Introduction Fatigue is a common symptom that is experienced across the life span among different populations (Barroso & Voss, 2013; Hallowell, 2010; Gay, Lee, & Lee, 2004; Miaskowski et al., 2008). Fatigue can be related to either mental or physical exertion or both. It can be an acute cyclic phenomenon in healthy workers, and it can be a more chronic distressful experience for people living with a health problem. For most healthy children and adults, exertion accumulates during the day, creating a pattern of higher fatigue severity in the evening and relief from fatigue by resting or getting a
good night of restorative sleep (Lerdal, Gay, Aouizerat, Portillo, & Lee, 2011). Without an opportunity to rest or get adequate sleep, which may occur with long work shifts, prolonged physical labor, or mentally challenging tasks, fatigue accumulates, and quality of life suffers (Hallowell, 2010). In individuals with a health problem, fatigue can also be related to their disease condition, medical therapy, or side effects of medications (Lerdal, Lee, Bakken, Finset, & Kim, 2012; Meek & Lareau, 2003; Voss, Dodd, Portillo, & Holzemer, 2006). Complicating the experience of fatigue for both healthy and ill populations is its overlapping features with other symptoms such as anxiety or depression, excessive daytime sleepiness, lack of motivation,
* Corresponding author: Kathryn A. Lee, School of Nursing, UCSF, Box 0606, San Francisco CA 94143-0606. E-mail address: [email protected] (K.A. Lee). 0029-6554/$ - see front matter Ó 2014 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.outlook.2014.07.002
314
Nurs Outlook 62 (2014) 313e321
cognitive impairment, or feeling overwhelmed by accumulating daily stressors (Aouizerat, Gay, Lerdal, Portillo, & Lee, 2013). In symptom science, it is fully acceptable to use a person’s self-report to assess the symptom of interest. In other words, the person is the “gold standard” for valid and reliable measurement. To some extent, using validated measures of fatigue only serves to complicate an already complex phenomenon. These fatigue measures are often multidimensional in nature and include impact on physical and mental function over time rather than current experience of fatigue’s frequency or severity. Many measures of fatigue conceptually overlap with measuring other symptoms (Aouizerat et al., 2013) including items on depression, low motivation, or sleepiness (Shen et al., 2006). This makes intervention trials even more complex to design and test because of the varying time frames and the various dimensions of symptoms assessed. A highly effective intervention may change the severity of fatigue but not its frequency, or it may alter the distress associated with the feeling but not the severity. Using the wrong measure or wrong time frame for measurement will only yield false-negative results. Fatigue measures and fatigue items within the Patient Reported Outcomes Measurement Information System (PROMIS) measures (Bevans, Ross, & Cella, 2014) should be evaluated by researchers for dimensions and time frames of relevance to the study design and population of interest. Fatigue is only one example of just how complex symptom research can be. What is discussed in this article can be applied to a multitude of other symptoms. It can also be applied to more complex phenomena such as symptom clusters (Dodd, Miaskowski, & Lee, 2004). Complex systems science has been applied to other areas of health research such as nutritional security and obesity (Hammond, 2009; Hammond & Dube´, 2012). In this article, current knowledge about the symptom of fatigue is used to describe how complex systems modeling approaches can facilitate our understanding of symptom science and gaps in knowledge.
Types of Fatigue and Potential Mechanisms Occupational Fatigue In occupational health and safety, fatigue is primarily focused on physical capacity and work productivity (Hallowell, 2010). Fatigued workers are less productive, but they are also at particularly high risk for accidents and errors in the work environment and in the home setting (Caruso, 2014; Lee & Lipscomb, 2003). Policies are usually in place to allow for adequate rest periods to minimize fatigue and keep everyone safe, but there are no assessments of how fatigued a worker may be going to and from the workplace. This type of fatigue in a healthy employee can endanger the worker’s health
and well-being and impact social and family interactions. There are societal implications as well. Fatigue of workers puts society at risk, particularly for patients getting 24-hour hospital care and for passengers aboard planes, buses, trains, or ships. Occupational fatigue is of great concern to nurses and hospital administrators as well as labor unions and insurance companies (Han, Trinkoff, Storr, & Geiger-Brown, 2011). Adding to the complexity of occupational fatigue is its overlapping features with stress, symptoms of anxiety or depression, or excessive sleepiness from short sleep duration. Most individuals, and their clinicians or workplace supervisors, cannot make distinctions between these confounding features (Akerstedt, Fredlund, Gillberg, & Jansson, 2002; Shen et al., 2006). Thus, the potential mechanisms for occupational fatigue are complex in nature and have implications for the health of individuals and society (Geiger-Brown et al., 2012). Figure 1 is an example of the complexity of fatigue in occupational health (Hallowell, 2010). Better analytic models for this type of fatigue will improve our understanding of mechanisms to target for more precise interventions to reduce the morbidity and mortality associated with occupational fatigue.
Disease-related Fatigue Fatigue is epidemic among children and adults with health problems. Disease-related fatigue is primarily focused on physical and mental function and limitations that either affect adherence to medical therapy or reduce quality of life. Clinicians often see fatigue as a subjective indicator of poor health or inadequate nutrition. This fatigue complaint is often nonspecific and requires the clinician to explore many possible culprits that can be remedied with medical treatment, such as obesity, anemia, thyroid dysfunction, depression, pain, sleep disorders, or physical deconditioning and lack of exercise. Genetic studies are also finding associations between fatigue and cytokine genetic markers (Lee, Gay, Lerdal, Pullinger, & Aouizerat, 2014; Miaskowski et al., 2010; Voss et al., 2013). Fatigue can be a precursor of infection or a warning sign of inflammatory processes during impending exacerbations of immune disorders like rheumatoid arthritis or multiple sclerosis. In most studies of patient populations that include laboratory values for anemia, like hemoglobin or hematocrit, or include clinical indicators of disease progression like viral load with HIV infection, there are usually weak associations with fatigue. Other factors such as socioeconomic status or age and sex are more strongly correlated with fatigue than clinical indicators (Aouizerat et al., 2013). If fatigue is not part of the disease process, it is often a side effect of medical therapy used to treat a health problem. An example of disease-related fatigue is cancer, with interventions that include chemotherapy and radiation (Lui et al., 2013; Miaskowski et al., 2011). Fatigue may be present as part of the initial cancer
Nurs Outlook 62 (2014) 313e321
315
Figure 1 e Occupational Fatigue Management Model (Hallowell, 2010). Reprinted with permission from M. Hallowell.
diagnosis, but treatment either worsens the fatigue severity or frequency or changes how fatigue is perceived as bothersome or highly distressing. Further complicating the fatigue symptom experience is the use of over-the-counter or prescription medications for sleep or the use of stimulants to manage daytime fatigue. These substances may have side effects of daytime sleepiness, fatigue, or inertia, and stimulants can linger and cause insomnia at bedtime. Fatigue can also impede recovery after surgery or acute illness onset such as in stroke (Lerdal et al., 2012). Figure 2 is one example of the complexity of fatigue in HIV disease (Voss et al., 2006). Better analyses and modeling of this type of fatigue and its mechanisms can improve our understanding of the fatigue experience and provide more precise clinical interventions to improve quality of life.
Environmental and Sociocultural Influences on Fatigue With both occupational fatigue and disease-related fatigue, there can be external environmental and social factors that the individual, the researcher, and the clinician may fail to consider. These external factors can exist alone but often interact with what is going on inside the person, physiologically or psychologically.
Work shift schedules and airline flights to different time zones are some of the more obvious factors (Geiger-Brown et al., 2012; Gander et al., 2013). However, for someone with diabetes or irritable bowel disease, the fatigue can be worsened by travel or a change in schedule. Other factors may be less obvious, such as appropriately timed light exposure. Seasonal variations in day lengths can precipitate mood disorders or “winter depression” because of a lack of sunlight, but poor or dim lighting at work or home can result in fatigue, insomnia, or daytime sleepiness (Baars, Gans, & Ellis, 2008; Rastad, Ulfberg, & Lindberg, 2011). A night shift worker who tries to sleep during the day gets little exposure to daylight. A person living with a chronic illness may also get inadequate light during the day because of physical or social limitations (Lui et al., 2013). Environmental noise pollution (Blomkvist, Eriksen, Theorell, Ulrich, & Rasmanis, 2005; Konkani, Oakley, & Penprase, 2014) and overcrowding conditions can also be fatiguing (Hallowell, 2010). These types of environmental stressors contribute to the fatigue experience on many levels, regardless of the underlying physiological mechanism or occupational influences. In addition, exposures to these environmental stressors on a chronic basis may be the sole source of fatigue. Confounding both occupational fatigue and disease-related fatigue is the environmental consideration of timing and distance. The distance to a
316
Nurs Outlook 62 (2014) 313e321
Figure 2 e Multidimensional Model of HIV-related Fatigue (Voss et al., 2006). Reprinted with permission from Dr. J. Barroso.
scheduled treatment (Lerdal et al., 2012), or the stress of commuting to work can influence fatigue (Han et al., 2011), with added complexity for navigating bus connections, day care or school schedules, and time of day contributing even more complexity to the fatigue experience. In addition, the common cultural experience of a daytime nap is disappearing from many societies because of longer commute distances. There is very little consideration given to scheduling medical appointments or work to allow an individual to awaken without alarm clocks and have sufficient time to bathe and dress for the appointment or workplace. There is also very little consideration for how to best schedule a nap as a preventive strategy for fatigue (Hartzler, 2014; Kassardjian et al., 2013). Finally, societal pressures, cultural attitudes, and stigma are part of the external environment that can add to the complexity of understanding and managing fatigue (Han et al., 2011). Today’s global marketplace and 24-hour access to social media place pressure on individuals and corporations to be “on call” at any time, even on days off. The workplace may have a culture that promotes a healthy lifestyle with scheduled meal and rest breaks and with space for social interactions outdoors, but the workplace may also have a culture of hiding fatigue and have expectations that taking stimulants is the solution rather than demanding rest breaks or an adequate night’s sleep before coming to work (Han et al., 2011; Hallowell, 2010).
Innovations in Computational Models for Symptom Science With the multifactorial influences on individuals’ perceptions of their symptom experience, computational models are needed in nursing science to handle the dynamic interactions between person, environment, and health and illness. In this final section, we describe three types of computational models that could be considered relevant with high potential for application to symptom science.
System Dynamics System dynamics modeling is a computer-aided, theory building, and decision-making approach used to understand dynamic problems arising from complex (dynamic) systems (Richardson, 2013). Dynamic systems are characterized by circular casualty, feedback, interdependence, and interaction (Richardson, 2013). The SD approach is especially informative when a symptom’s cause and effect might be removed from each other across time, have multiple feedback loops, cause unforeseen ripple effects, or is resistant to change (Sterman, 2000). For example, fatigue might be caused by sleep disturbance or high stress at work, or excessive fatigue may reduce physical activity and create muscle deconditioning that worsens fatigue or
Nurs Outlook 62 (2014) 313e321
leads to depression. It is also possible that stress at work might counteract the positive aspects of physical exercise, making it more difficult to recover from fatigue. If variables are assessed at a single point in time and subjected to a multiple regression types of models, the investigator can only test what is set up as the dependent variable (effect) and independent variables (predictors). Any type of feedback loop, rates of change, or ripple effects will be missed with more traditional logistic or linear regression models. Building an SD model starts with defining the “dynamics” of the problem with graphs over time (Richardson, 2013) and formulating a dynamic hypothesis (Sterman, 2000) through the construction of a conceptual map or model of the problem or symptom under investigation. The conceptual map represents an endogenous, “arising from within” explanation of the problem. Two particularly important conceptual maps are (a) stock and flow diagrams and (b) causal loop diagrams. Figure 3 is an example of a stock and flow diagram, showing two flows (“becoming fatigued” and “recovering from fatigue”) as well as one stock (“population with fatigue”). The “stock” represents the current state of the system, and the “flows” change the state of the system by increasing or decreasing at a certain rate. In the simplest form, this represents a single differential equation, and as more stocks and flows are added, as well as other variables affecting the system, it quickly becomes a set of differential equations. Figure 4 is a causal loop map or model that depicts an SD model that is more circuitous. With an increase in the severity or frequency of sleep disturbance, fatigue and stress also increase. Stock and flow diagrams are useful for identifying inflows and outflows in a system, which are a building block of an SD model. However, it is difficult to show how other variables influence, or are influenced by, a rate of change for people becoming fatigued or recovering from fatigue. Causal loop diagrams are also important for identifying potential feedback loops in the system, but if disconnected from stocks and flows, it is difficult to determine their influence on the stock of a population with fatigue or how they then influence the rate at which a person becomes fatigued or recovers from fatigue. Hence, a complete SD model, which incorporates a stock and flow model with a causal loop model, is shown in Figure 5. After constructing the conceptual models, a formal mathematical model is constructed using coupled nonlinear
Figure 3 e Stock and Flow Diagram. There are two flows (“becoming fatigued” and “recovering from fatigue”) as well as one stock (“population with fatigue”). Stock represents the current state of the system, and the “flows” change the state of the system by increasing or decreasing at a certain rate.
317
Figure 4 e Causal Loop Diagram. As one state increases, the other states also increase. Hence, with an increase in the severity or frequency of sleep disturbance, fatigue and stress also increase.
differential equations and then simulated over time. For example, a complete SD model could be used to study factors that contribute to developing fatigue (e.g., stress, lack of physical activity, depression, or poor sleep) and factors that might facilitate recovering from fatigue (e.g., stress management, exercise, antidepressants, and cognitive behavioral therapy) to see how these factors affect each other. Researchers can then design an intervention based on manipulating parameters in the SD model to decrease rates of fatigue over time. There are a number of software options for SD modeling. The most popular are Stella (Isee Systems, 2014) Vensim (Vensim, 2010; Ventana Systems, Inc, 2014), and Powersim (Powersim Software AS, 2014). For more advanced users, there are open-source options such as simecol (Simecol, Dresden, Germany) (Jones, 2007; Petzoldt & Rinke, 2007) and StellaR packages in R (Naimi & Voinov, 2012). SD can also be done as a secondary feature in some agent-based modeling
Figure 5 e Complete SD Model. The polarity of the loop depends on the number of positive arrows relative to the number of negative arrows.
318
Nurs Outlook 62 (2014) 313e321
(ABM) software packages (NetLogo, Northwestern Institute on Complex Systems, Evanston, IL) (Thiele, 2014; Wilensky, 1997, 1999).
Agent Based Modeling (ABM) Similar to SD modeling, ABM is a computer-aided mathematical simulation method. The theoretical foundation of ABM is based on discrete mathematical models, including the Turing machine, cellular automata, and Conway’s Game of Life (Epstein, 2007; Epstein, Parker, Cummings, & Hammond, 2008; Gotts, 2009). ABM models the interactions among autonomous agents (i.e., individuals). Each autonomous agent or person can have multiple demographic characteristics and be assigned behavioral preferences that can be modeled over time (Verzi, Brodsky, Brown, Apelberg, & Rostron, 2012). ABM can be used to examine the consequences of behaviors over time. ABM would be particularly useful for symptom research because it also handles the complexities of groups and environments. ABM can include groups of individuals with certain characteristics, such as night shift workers, groups living in certain regions (above the Arctic Circle or in humid coastal regions), or work settings (intensive care, labor and delivery, etc.). ABM models could be used to understand how a health-related decision or behavior of an individual agent (night shift nurse) could lead to the development of chronic fatigue. This may be particularly useful for research on symptoms when there is initially no complaint of the symptom.
Unlike more traditional regression models, which might tell us context-level predictors of fatigue, an ABM can help us understand how individual characteristics and behaviors over time might contribute to the development of chronic fatigue in a specific environment in which individual choices are constrained by the characteristics of that environment. Similar to SD models, there are a number of software choices for creating ABM models. Two popular commercial options include NetLogo (Wilensky, 1997, 1999) and AnyLogic (XJ Technologies, 2009). For more advanced users, there are open-source options, such as simecol (Petzoldt, 2014) and RNetLogo (Thiele, 2014) packages in R (R Core Team, 2013). A screenshot example of an ABM is shown in Figure 6. This model (Wilensky, 1997, 1999) shows how a population of individuals infected with HIV would change over time based on their average coupling tendency, average commitment, average condom use, and average test frequency. Researchers can then manipulate the parameters to observe their impact on potential infection rates (74.5% infected in the example given the four parameters being modeled).
Markov Models Markov models (MMs) are a group of stochastic time series models based on the idea of a Markov property, which states that a future state of a process depends only on the present state and not on any previous states (Murphy, 2012). MMs are especially useful for understanding complex time series and have been widely used in the area of machine
Figure 6 e Screenshots of an ABM Using NetLogo to Model Rates of HIV Infection Over Time. Reprinted with permission from U. Wilensky, 1997. Setting parameters for four key variables (coupling tendency, time in committed relationship, condom use, and HIV testing frequency) using NetLogo software allows the researcher to observe how each of these parameters would impact HIV infection rates.
Nurs Outlook 62 (2014) 313e321
learning and artificial intelligence. MMs could be applied to understand the stage of disease development from data contained in large databases such as electronic health records (EHRs). The symptom experience in many types of chronic illness populations, including HIV/AIDS, cancer, mood disorders, and type 2 diabetes mellitus, could be modeled to examine patterns over time to identify disease development stages or potential unobserved variables influencing disease progression. This can eventually provide clues to potential interventions and data for patient teaching and anticipatory preparation for future health care needs. MMs can estimate stages of disease or symptom progression or can show multiple ways in which a disease or symptom could develop. A two-state MM example is seen in Figure 7; the circles represent states, and the number of arrows represents probabilities. For example, in a healthy state, the probability that a person will stay in a healthy state is 0.6, whereas the probability is 0.4 that the person will develop fatigue at the next state transition. These are called transition probabilities. With wider use of EHRs, nurse researchers can “train” MMs (in this case hidden MMs [HMMs]) to recognize symptom patterns, levels of pain, anxiety, or fatigue severity for example. Training HMMs consists of performing parameter estimation either on a subsample or full data in order to use these parameter estimates to classify new cases. For example, if we discover the process of becoming fatigued and have trained our HMM, we can use our parameter estimates to classify new cases as being at a certain hidden state or estimate the probability of moving from their current state to a different state. As seen in Figure 8, HMMs allow a researcher to group patients in an EHR dataset by these patterns or states and then design more precision-based interventions to test in future symptom research. An HMM can provide additional detail to better understand whether someone is moving from a healthy state to an unhealthy fatigued state. For example, we might know that there is a higher chance that someone starts in a healthy state, and we might also know the probabilities for feeling stressed, sleepy, or normal while in either a healthy state or fatigued state. These are called emission probabilities. We can easily tell which state someone is in, and these two states are
Figure 7 e Example of a Two-state Markov Chain. Circles represent states, and the numbers of arrows represent probabilities. In a healthy state, the probability that one will stay in a healthy state is 0.6, thus the probability that one will develop fatigue at the next state transition is 0.4.
319
Figure 8 e Example of an HMM. At a baseline measurement, there is a higher probability that a healthy individual will start in a healthy state rather than a fatigued state. The probabilities of feeling stressed, sleepy, or normal while someone is in either a healthy state or fatigued state are called emission probabilities. Fatigued state could then be approximated from the additional information in the model. said to be “hidden and can then be approximated from the additional information in the model. This has now transformed from a two-state Markov chain (MM) into an HMM. MMs can be formulated in MATLAB (MathWorks, Inc., 2014) using a modeling toolkit for a desired MM. Octave (GNU Octave) is an open-source version of MATLAB (Eaton, Bateman, & Hauberg, 2008). There are also a number of MM packages available in R (R Core Team, 2013).
Initial Training in Computational Models All three analytic tools presented in this article require a different skill set than taught in traditional graduatelevel statistics courses. There are many resources available for learning about the theoretical underpinnings for these types of models and how to apply these models to symptom science. Some of these learning resources are also available at no cost. Selflearning is encouraged for individuals with a limited range of programming experience (such as writing simple SPSS syntax) or having a background in elementary calculus. Individuals not experienced in running syntax-based mathematical models are encouraged to take a workshop or a class on one of the discussed approaches.
320
Nurs Outlook 62 (2014) 313e321
The System Dynamics Society (SDS) is an international nonprofit organization with a variety of resources for individuals wanting examples of SD models and wanting to learn about SD (http://www.systemdynamics.org). Similar to SDS, a beginning system scientist can find resources on ABM at a network maintained by ABM researchers (http://www.openabm.org/). There is no central online resource for HMMs, but an online class in machine learning would be beneficial for an introduction to the Markov process, with some elementary illustrations. For more advanced users, Murphy’s Matlab toolbox might be a good starting point (http://www.cs.ubc.ca/ wmurphyk/Software/HMM/hmm.html). For R users, Visser’s “depmixS4” package has a “Github” page that includes sample code and models (https://github.com/ cran/depmixS4).
Conclusions The symptom of fatigue was used in this article as an example to show the complexity of symptom science and the difficulties inherent in assessing and testing interventions designed to improve health outcomes. Other symptoms or symptom clusters could also be examined, including pain, nausea, thirst, tinnitus, insomnia, and dyspnea for example. The three examples of statistical models described in this article hold tremendous promise for increasing our understanding of complex phenomena in symptom research. All models like these will require a time component, whether using crosssectional or longitudinal data. As with any analytic model, the model is only as good as the hypothesis it is testing and the skill of the researcher. The model will not yield correct estimates if the underlying hypothesis is flawed. In addition, as in traditional structural equation or multilevel modeling, researchers should perform sensitivity analysis with ABMs, and model fit statistics should be used to assess the fit of HMMs. SD and ABM offer nurse scientists the potential to develop types of models, whether the data set is small or large. HMMs may perform better with big data than other modeling techniques. These types of models can include many more predictors and determinants of the intensity and distress for a given symptom, symptom cluster, or pattern of symptom experiences over time. This would be especially important in addressing the question of how and whether or not cancer-related fatigue (Barsevick et al., 2013) and HIV-related fatigue (Barroso & Voss, 2013) differ from other types of fatigue like shift work fatigue (Carruso, 2014; Geiger-Brown et al., 2012) or pregnancy fatigue (Gay et al., 2004). Although these analytic techniques are commonly used to understand complex systems of care and estimate potential risk or success with intervention strategies, applying these techniques to symptom science will expand our ability to intervene most effectively and with better precision. In addition, MMs provide a much more quantitative approach for determining a
patient’s improvement or potential for declining over time. This is a novel approach to symptom research that nurse scientists have strongly desired for decades. It is clear that symptom research is in need of these types of analytic techniques, and nurse researchers would do well to learn, collaborate, and incorporate these strategies into their science.
references
Akerstedt, T., Fredlund, P., Gillberg, M., & Jansson, B. (2002). Work load and work hours in relation to disturbed sleep and fatigue in a large representative sample. Journal of Psychosomatic Research, 53(1), 585e588. Aouizerat, B. E., Gay, C. L., Lerdal, A., Portillo, C. J., & Lee, K. A. (2013). Lack of energy: An important and distinct component of HIV-related fatigue and daytime function. Journal of Pain Symptom Management, 45(2), 191e201. Baars, E. W., Gans, S., & Ellis, E. L. (2008). The effect of hepar magnesium on seasonal fatigue symptoms: A pilot study. Journal of Alternative and Complementary Medicine, 14(4), 395e402. Barroso, J., & Voss, J. G. (2013). Fatigue in HIV and AIDS: An analysis of evidence. Journal of the Association of Nurses in AIDS Care, 24(1 Suppl), S5e14. Barsevick, A. M., Irwin, M. R., Hinds, P., Miller, A., Berger, A., Jacobsen, P., . Cella, D. (2013). National Cancer Institute clinical trials planning meeting. Recommendations for highpriority research on cancer-related fatigue in children and adults. Journal of the National Cancer Institute, 105(19), 1432e1440. Bevans, M., Ross, A., & Cella, D. (2014). Patient-reported outcomes measurement information system (PROMISÒ): Efficient, standardized tools to measure self-reported health and quality of life. Nursing Outlook, 62(5), 339e345. Blomkvist, V., Eriksen, C. A., Theorell, T., Ulrich, R., & Rasmanis, G. (2005). Acoustics and psychosocial environment in intensive coronary care. Occupational and Environmental Medicine, 62(3), e1. Caruso, C. C. (2014). Negative impacts of shiftwork and long work hours. Rehabilitation Nursing, 39(1), 16e25. Dodd, M. J., Miaskowski, C., & Lee, K. A. (2004). Symptom clusters: The new frontier in symptom management research. Journal of the National Cancer Institute Monographs, 2004(32), 17e21. Eaton, J. W., Bateman, D., & Hauberg, S. (2008). GNU Octave Manual Version 3: A high-level interactive language for numerical computations. Bristol, UK: Network Theory Ltd. Epstein, J. M. (2007). Generative social science: Studies in agent-based computational modeling. Princeton University Press. Retrieved from http://press.princeton.edu/chapters/s8277.html. Epstein, J. M., Parker, J., Cummings, D., & Hammond, R. A. (2008). Coupled contagion dynamics of fear and disease: Mathematical and computational explorations. PLoS One, 3(12). http://dx.doi.org/10.1371/journal.pone.0003955. Gander, P. H., Signal, T. L., van den Berg, M. J., Mulrine, H. M., Jay, S. M., & Jim Mangie, C. (2013). In-flight sleep, pilot fatigue and psychomotor vigilance task performance on ultra-long range versus long range flights. Journal of Sleep Research, 22(6), 697e706. Gay, C. L., Lee, K. A., & Lee, S. Y. (2004). Sleep patterns and fatigue in new mothers and fathers. Biological Research for Nursing, 5(4), 311e318. Geiger-Brown, J., Rogers, V. E., Trinkoff, A. M., Kane, R. L., Bausell, R. B., & Scharf, S. M. (2012). Sleep, sleepiness, fatigue,
Nurs Outlook 62 (2014) 313e321
and performance of 12-hour-shift nurses. Chronobiology International, 29(2), 211e219. Gotts, N. M. (2009). Ramifying feedback networks, cross-scale interactions, and emergent quasi individuals in Conway’s game of Life. Artificial Life, 15(3), 351e375. Hallowell, M. R. (2010). Worker fatigue: Managing concerns in rapid renewal highway construction projects. Professional Safety (pp. 18e26), Retrieved from http://www.asse.org/professional safety/pastissues/055/12/018_026_F1Hallo_1210Z.pdf. Hammond, R. A. (2009). Complex systems modeling for obesity research. Preventing Chronic Disease, 6(3), A97. Hammond, R. A., & Dube´, L. (2012). A systems science perspective and transdisciplinary models for food and nutrition security. Proceedings of the National Academy of Sciences of the United States of America, 109(31), 12356e12363. Han, K., Trinkoff, A. M., Storr, C. L., & Geiger-Brown, J. (2011). Job stress and work schedules in relation to nurse obesity. Journal of Nursing Administration, 41(11), 488e495. Hartzler, B. M. (2014). Fatigue on the flight deck: The consequences of sleep loss and the benefits of napping. Accident Analysis & Prevention, 62(1), 309e318. Isee Systems, Inc. (2014). STELLA. Isee Systems, Inc. Retrieved from http://www.iseesystems.com/softwares/Education/ StellaSoftware.aspx. Jones, O. (2007). Simco: A Package to Import Structure Files and Deduce Similarity Coefficients from Them. R package version, 1, 2007. Kassardjian, C. D., Murray, B. J., Kokokyi, S., Jewell, D., Barnett, C., Bril, V., . Katzberg, H. D. (2013). Effects of napping on neuromuscular fatigue in myasthenia gravis. Muscle & Nerve, 48(5), 816e818. Konkani, A., Oakley, B., & Penprase, B. (2014). Reducing hospital ICU noise: A behavior-based approach. Journal of Healthcare Engineering, 5(2), 229e246. Lee, K. A., Gay, C. L., Lerdal, A., Pullinger, C. R., & Aouizerat, B. E. (2014). Cytokine polymorphisms are associated with fatigue in adults living with HIV/AIDS. Brain, Behavior, & Immunity, 40, 95e103. Lee, K. A., & Lipscomb, J. (2003). Sleep among shiftworkers-A priority for clinical practice and research in occupational health nursing. American Association of Occupational Health Nursing Journal, 51(10), 418e420. Lerdal, A., Gay, C. L., Aouizerat, B. E., Portillo, C. J., & Lee, K. A. (2011). Patterns of morning and evening fatigue among adults with HIV/AIDS. Journal of Clinical Nursing, 20(15-16), 2204e2216. Lerdal, A., Lee, K. A., Bakken, L. N., Finset, A., & Kim, H. S. (2012). The course of fatigue during the first 18 months after first-ever stroke: A longitudinal study. Stroke Research and Treatment, 2012, article 126275. http://dx.doi.org/10.1155/2012/126275. Liu, L., Rissling, M., Neikrug, A., Fiorentino, L., Natarajan, L., Faierman, M., . Ancoli-Israel, S. (2013). Fatigue and circadian activity rhythms in breast cancer patients before and after chemotherapy: A controlled study. Fatigue, 1(1-2), 12e26. MathWorks, Inc. (2014). MATLABÓ. MathWorks, Inc. Retrieved from http://www.mathworks.com/products/matlab/. Meek, P. M., & Lareau, S. C. (2003). Critical outcomes in pulmonary rehabilitation: Assessment and evaluation of dyspnea and fatigue. Journal of Rehabilitation Research and Development, 40(5 Suppl 2), 13e24. Miaskowski, C., Dodd, M., Lee, K., West, C., Paul, S. M., Cooper, B. A., . Aouizerat, B. E. (2010). Preliminary evidence of an association between a functional interleukin-6 polymorphism and fatigue and sleep disturbance in oncology patients and their family caregivers. Journal of Pain and Symptom Management, 40(4), 531e544. Miaskowski, C., Lee, K., Dunn, L., Dodd, M., Aouizerat, B. E., West, C., . Swift, P. (2011). Sleep-wake circadian activity
321
rhythm parameters and fatigue in oncology patients before the initiation of radiation therapy. Cancer Nursing, 34(4), 255e268. Miaskowski, C., Paul, S. M., Cooper, B. A., Lee, K., Dodd, M., West, C., . Wara, W. (2008). Trajectories of fatigue in men with prostate cancer before, during, and after radiation therapy. Journal of Pain and Symptom Management, 35(6), 632e643. Murphy, K. P. (2012). Machine learning: A probabilistic perspective. MIT Press. Naimi, B., & Voinov, A. (2012). StellaR: A software to translate Stella models into R open-source environment. Environmental Modelling & Software, 38(1), 117e118. Petzoldt, T., & Rinke, K. (2007). Simecol: An object-oriented framework for ecological modeling in R. Journal of Statistical Software, 22(9), 1e31. Petzoldt, T. (2014). Simulation of ecological (and other) dynamic systems. Retrieved from http://cran.r-project.org/web/ packages/simecol/simecol.pdf. Powersim Software AS. (2014). Powersim Studio 10Ò. Power Sim Software ASÓ. Retrieved from http://www.powersim.com/. R Core Team. (2013). R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing, ISBN 3-900051-07-0. URL. http://www.R-project.org/. Rastad, C., Ulfberg, J., & Lindberg, P. (2011). Improvement in fatigue, sleepiness, and health-related quality of life with bright light treatment in persons with seasonal affective disorder and subsyndromal SAD. Depression Research and Treatment, 2011, 543906. http://dx.doi.org/10.1155/2011/ 543906. Richardson, G. P. (2013). System dynamics. In S. Gass, & C. Harris (Eds.), Encyclopedia of operations research and management science (pp. 1519e1522). Kluwer Academic Publishers. Shen, J., Botly, L. C., Chung, S. A., Gibbs, A. L., Sabanadzovic, S., & Shapiro, C. M. (2006). Fatigue and shift work. Journal of Sleep Research, 15(1), 1e5. Sterman, J. D. (2000). Business dynamics: Systems thinking and modeling for a complex world. Boston: Irwin/McGraw-Hill. Thiele, J. C. (2014). R marries NetLogo: Introduction to the RNetLogo package. Journal of Statistical Software, 58(2), 1e41. Vensim, P. (2010). Ventana Systems, Inc. Retrieved from http:// www.vensim.com. Ventana Systems, Inc. (2014). Vensim PLE SoftwareÒ. Ventana Systems Inc. Retrieved from http://vensim.com/. Verzi, S. J., Brodsky, N. S., Brown, T. J., Apelberg, B., & Rostron, B. (2012). An agent-based approach for modeling population behavior and health with application to tobacco use. Sandia National Laboratories. Voss, J. G., Dobra, A., Morse, C., Kovacs, J. A., Danner, R. L., Munson, P. J., . Dalakas, M. C. (2013). Fatigue-related gene networks identified in CD(14)þ cells isolated from HIVinfected patients: Part I: Research findings. Biological Research for Nursing, 15(2), 137e151. Voss, J. G., Dodd, M., Portillo, C., & Holzemer, W. (2006). Theories of fatigue: Application in HIV/AIDS. Journal of the Association of Nurses in AIDS Care, 17(1), 37e50. Wilensky, U. (1997). NetLogo AIDS Model. Evanston, IL: Center for Connected Learning and Computer-Based Modeling, Northwestern University. Retrieved from http://ccl. northwestern.edu/netlogo/models/AIDS. Wilensky, U. (1999). NetLogo. Evanston, IL: Center for Connected Learning and Computer-Based Modeling, Northwestern University. Retrieved from http://ccl.northwestern.edu/ netlogo/. XJ Technologies. (2009). AnyLogicÒ. XJ technologies. Retrieved from http://www.xjtek.com/.
Nurs Outlook 62 (2014) 313e321
www.nursingoutlook.org
A systems science approach to fatigue management in research and health care Kathryn A. Lee, PhD, RN, FAAN, CBSMa,*, Oliwier Dziadkowiec, PhDb, Paula Meek, PhD, RN, FAANc b
a School of Nursing, University of California, San Francisco, San Francisco CA Center for Nursing Research, College of Nursing, University of Colorado at Denver, Aurora, CO c College of Nursing, University of Colorado at Denver, Aurora, CO
article info
abstract
Article history: Received 25 June 2014 Revised 2 July 2014 Accepted 3 July 2014 Available online 9 July 2014
The purpose of this article was to highlight innovative analytic approaches in symptom science. Complex systems modeling is discussed using fatigue as an example. Fatigue is a common symptom among individuals of any age. It can be acute or chronic, and it can vary across the day and on weekends compared with weekdays. Fatigue can overlap with other symptoms, has many dimensions, and impacts daily function as well as society at large. With the complexities surrounding symptom science, innovative models are needed to advance our understanding of factors within the person, contextual and situational factors, and workplace or health care system factors that impact the symptom experience. Advances in methodologies, such as complex systems modeling, allow for more innovative methods to study the complexities of the symptom experience, design better ways to intervene and manage symptoms, and ultimately improve outcomes related to symptom management, quality of life, and health care utilization.
Keywords: Agent-based modeling Big data Complex systems modeling Markov model Precision medicine System dynamics
Cite this article: Lee, K. A., Dziadkowiec, O., & Meek, P. (2014, OCTOBER). A systems science approach to fatigue management in research and health care. Nursing Outlook, 62(5), 313-321. http://dx.doi.org/ 10.1016/j.outlook.2014.07.002.
Introduction Fatigue is a common symptom that is experienced across the life span among different populations (Barroso & Voss, 2013; Hallowell, 2010; Gay, Lee, & Lee, 2004; Miaskowski et al., 2008). Fatigue can be related to either mental or physical exertion or both. It can be an acute cyclic phenomenon in healthy workers, and it can be a more chronic distressful experience for people living with a health problem. For most healthy children and adults, exertion accumulates during the day, creating a pattern of higher fatigue severity in the evening and relief from fatigue by resting or getting a
good night of restorative sleep (Lerdal, Gay, Aouizerat, Portillo, & Lee, 2011). Without an opportunity to rest or get adequate sleep, which may occur with long work shifts, prolonged physical labor, or mentally challenging tasks, fatigue accumulates, and quality of life suffers (Hallowell, 2010). In individuals with a health problem, fatigue can also be related to their disease condition, medical therapy, or side effects of medications (Lerdal, Lee, Bakken, Finset, & Kim, 2012; Meek & Lareau, 2003; Voss, Dodd, Portillo, & Holzemer, 2006). Complicating the experience of fatigue for both healthy and ill populations is its overlapping features with other symptoms such as anxiety or depression, excessive daytime sleepiness, lack of motivation,
* Corresponding author: Kathryn A. Lee, School of Nursing, UCSF, Box 0606, San Francisco CA 94143-0606. E-mail address: [email protected] (K.A. Lee). 0029-6554/$ - see front matter Ó 2014 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.outlook.2014.07.002
314
Nurs Outlook 62 (2014) 313e321
cognitive impairment, or feeling overwhelmed by accumulating daily stressors (Aouizerat, Gay, Lerdal, Portillo, & Lee, 2013). In symptom science, it is fully acceptable to use a person’s self-report to assess the symptom of interest. In other words, the person is the “gold standard” for valid and reliable measurement. To some extent, using validated measures of fatigue only serves to complicate an already complex phenomenon. These fatigue measures are often multidimensional in nature and include impact on physical and mental function over time rather than current experience of fatigue’s frequency or severity. Many measures of fatigue conceptually overlap with measuring other symptoms (Aouizerat et al., 2013) including items on depression, low motivation, or sleepiness (Shen et al., 2006). This makes intervention trials even more complex to design and test because of the varying time frames and the various dimensions of symptoms assessed. A highly effective intervention may change the severity of fatigue but not its frequency, or it may alter the distress associated with the feeling but not the severity. Using the wrong measure or wrong time frame for measurement will only yield false-negative results. Fatigue measures and fatigue items within the Patient Reported Outcomes Measurement Information System (PROMIS) measures (Bevans, Ross, & Cella, 2014) should be evaluated by researchers for dimensions and time frames of relevance to the study design and population of interest. Fatigue is only one example of just how complex symptom research can be. What is discussed in this article can be applied to a multitude of other symptoms. It can also be applied to more complex phenomena such as symptom clusters (Dodd, Miaskowski, & Lee, 2004). Complex systems science has been applied to other areas of health research such as nutritional security and obesity (Hammond, 2009; Hammond & Dube´, 2012). In this article, current knowledge about the symptom of fatigue is used to describe how complex systems modeling approaches can facilitate our understanding of symptom science and gaps in knowledge.
Types of Fatigue and Potential Mechanisms Occupational Fatigue In occupational health and safety, fatigue is primarily focused on physical capacity and work productivity (Hallowell, 2010). Fatigued workers are less productive, but they are also at particularly high risk for accidents and errors in the work environment and in the home setting (Caruso, 2014; Lee & Lipscomb, 2003). Policies are usually in place to allow for adequate rest periods to minimize fatigue and keep everyone safe, but there are no assessments of how fatigued a worker may be going to and from the workplace. This type of fatigue in a healthy employee can endanger the worker’s health
and well-being and impact social and family interactions. There are societal implications as well. Fatigue of workers puts society at risk, particularly for patients getting 24-hour hospital care and for passengers aboard planes, buses, trains, or ships. Occupational fatigue is of great concern to nurses and hospital administrators as well as labor unions and insurance companies (Han, Trinkoff, Storr, & Geiger-Brown, 2011). Adding to the complexity of occupational fatigue is its overlapping features with stress, symptoms of anxiety or depression, or excessive sleepiness from short sleep duration. Most individuals, and their clinicians or workplace supervisors, cannot make distinctions between these confounding features (Akerstedt, Fredlund, Gillberg, & Jansson, 2002; Shen et al., 2006). Thus, the potential mechanisms for occupational fatigue are complex in nature and have implications for the health of individuals and society (Geiger-Brown et al., 2012). Figure 1 is an example of the complexity of fatigue in occupational health (Hallowell, 2010). Better analytic models for this type of fatigue will improve our understanding of mechanisms to target for more precise interventions to reduce the morbidity and mortality associated with occupational fatigue.
Disease-related Fatigue Fatigue is epidemic among children and adults with health problems. Disease-related fatigue is primarily focused on physical and mental function and limitations that either affect adherence to medical therapy or reduce quality of life. Clinicians often see fatigue as a subjective indicator of poor health or inadequate nutrition. This fatigue complaint is often nonspecific and requires the clinician to explore many possible culprits that can be remedied with medical treatment, such as obesity, anemia, thyroid dysfunction, depression, pain, sleep disorders, or physical deconditioning and lack of exercise. Genetic studies are also finding associations between fatigue and cytokine genetic markers (Lee, Gay, Lerdal, Pullinger, & Aouizerat, 2014; Miaskowski et al., 2010; Voss et al., 2013). Fatigue can be a precursor of infection or a warning sign of inflammatory processes during impending exacerbations of immune disorders like rheumatoid arthritis or multiple sclerosis. In most studies of patient populations that include laboratory values for anemia, like hemoglobin or hematocrit, or include clinical indicators of disease progression like viral load with HIV infection, there are usually weak associations with fatigue. Other factors such as socioeconomic status or age and sex are more strongly correlated with fatigue than clinical indicators (Aouizerat et al., 2013). If fatigue is not part of the disease process, it is often a side effect of medical therapy used to treat a health problem. An example of disease-related fatigue is cancer, with interventions that include chemotherapy and radiation (Lui et al., 2013; Miaskowski et al., 2011). Fatigue may be present as part of the initial cancer
Nurs Outlook 62 (2014) 313e321
315
Figure 1 e Occupational Fatigue Management Model (Hallowell, 2010). Reprinted with permission from M. Hallowell.
diagnosis, but treatment either worsens the fatigue severity or frequency or changes how fatigue is perceived as bothersome or highly distressing. Further complicating the fatigue symptom experience is the use of over-the-counter or prescription medications for sleep or the use of stimulants to manage daytime fatigue. These substances may have side effects of daytime sleepiness, fatigue, or inertia, and stimulants can linger and cause insomnia at bedtime. Fatigue can also impede recovery after surgery or acute illness onset such as in stroke (Lerdal et al., 2012). Figure 2 is one example of the complexity of fatigue in HIV disease (Voss et al., 2006). Better analyses and modeling of this type of fatigue and its mechanisms can improve our understanding of the fatigue experience and provide more precise clinical interventions to improve quality of life.
Environmental and Sociocultural Influences on Fatigue With both occupational fatigue and disease-related fatigue, there can be external environmental and social factors that the individual, the researcher, and the clinician may fail to consider. These external factors can exist alone but often interact with what is going on inside the person, physiologically or psychologically.
Work shift schedules and airline flights to different time zones are some of the more obvious factors (Geiger-Brown et al., 2012; Gander et al., 2013). However, for someone with diabetes or irritable bowel disease, the fatigue can be worsened by travel or a change in schedule. Other factors may be less obvious, such as appropriately timed light exposure. Seasonal variations in day lengths can precipitate mood disorders or “winter depression” because of a lack of sunlight, but poor or dim lighting at work or home can result in fatigue, insomnia, or daytime sleepiness (Baars, Gans, & Ellis, 2008; Rastad, Ulfberg, & Lindberg, 2011). A night shift worker who tries to sleep during the day gets little exposure to daylight. A person living with a chronic illness may also get inadequate light during the day because of physical or social limitations (Lui et al., 2013). Environmental noise pollution (Blomkvist, Eriksen, Theorell, Ulrich, & Rasmanis, 2005; Konkani, Oakley, & Penprase, 2014) and overcrowding conditions can also be fatiguing (Hallowell, 2010). These types of environmental stressors contribute to the fatigue experience on many levels, regardless of the underlying physiological mechanism or occupational influences. In addition, exposures to these environmental stressors on a chronic basis may be the sole source of fatigue. Confounding both occupational fatigue and disease-related fatigue is the environmental consideration of timing and distance. The distance to a
316
Nurs Outlook 62 (2014) 313e321
Figure 2 e Multidimensional Model of HIV-related Fatigue (Voss et al., 2006). Reprinted with permission from Dr. J. Barroso.
scheduled treatment (Lerdal et al., 2012), or the stress of commuting to work can influence fatigue (Han et al., 2011), with added complexity for navigating bus connections, day care or school schedules, and time of day contributing even more complexity to the fatigue experience. In addition, the common cultural experience of a daytime nap is disappearing from many societies because of longer commute distances. There is very little consideration given to scheduling medical appointments or work to allow an individual to awaken without alarm clocks and have sufficient time to bathe and dress for the appointment or workplace. There is also very little consideration for how to best schedule a nap as a preventive strategy for fatigue (Hartzler, 2014; Kassardjian et al., 2013). Finally, societal pressures, cultural attitudes, and stigma are part of the external environment that can add to the complexity of understanding and managing fatigue (Han et al., 2011). Today’s global marketplace and 24-hour access to social media place pressure on individuals and corporations to be “on call” at any time, even on days off. The workplace may have a culture that promotes a healthy lifestyle with scheduled meal and rest breaks and with space for social interactions outdoors, but the workplace may also have a culture of hiding fatigue and have expectations that taking stimulants is the solution rather than demanding rest breaks or an adequate night’s sleep before coming to work (Han et al., 2011; Hallowell, 2010).
Innovations in Computational Models for Symptom Science With the multifactorial influences on individuals’ perceptions of their symptom experience, computational models are needed in nursing science to handle the dynamic interactions between person, environment, and health and illness. In this final section, we describe three types of computational models that could be considered relevant with high potential for application to symptom science.
System Dynamics System dynamics modeling is a computer-aided, theory building, and decision-making approach used to understand dynamic problems arising from complex (dynamic) systems (Richardson, 2013). Dynamic systems are characterized by circular casualty, feedback, interdependence, and interaction (Richardson, 2013). The SD approach is especially informative when a symptom’s cause and effect might be removed from each other across time, have multiple feedback loops, cause unforeseen ripple effects, or is resistant to change (Sterman, 2000). For example, fatigue might be caused by sleep disturbance or high stress at work, or excessive fatigue may reduce physical activity and create muscle deconditioning that worsens fatigue or
Nurs Outlook 62 (2014) 313e321
leads to depression. It is also possible that stress at work might counteract the positive aspects of physical exercise, making it more difficult to recover from fatigue. If variables are assessed at a single point in time and subjected to a multiple regression types of models, the investigator can only test what is set up as the dependent variable (effect) and independent variables (predictors). Any type of feedback loop, rates of change, or ripple effects will be missed with more traditional logistic or linear regression models. Building an SD model starts with defining the “dynamics” of the problem with graphs over time (Richardson, 2013) and formulating a dynamic hypothesis (Sterman, 2000) through the construction of a conceptual map or model of the problem or symptom under investigation. The conceptual map represents an endogenous, “arising from within” explanation of the problem. Two particularly important conceptual maps are (a) stock and flow diagrams and (b) causal loop diagrams. Figure 3 is an example of a stock and flow diagram, showing two flows (“becoming fatigued” and “recovering from fatigue”) as well as one stock (“population with fatigue”). The “stock” represents the current state of the system, and the “flows” change the state of the system by increasing or decreasing at a certain rate. In the simplest form, this represents a single differential equation, and as more stocks and flows are added, as well as other variables affecting the system, it quickly becomes a set of differential equations. Figure 4 is a causal loop map or model that depicts an SD model that is more circuitous. With an increase in the severity or frequency of sleep disturbance, fatigue and stress also increase. Stock and flow diagrams are useful for identifying inflows and outflows in a system, which are a building block of an SD model. However, it is difficult to show how other variables influence, or are influenced by, a rate of change for people becoming fatigued or recovering from fatigue. Causal loop diagrams are also important for identifying potential feedback loops in the system, but if disconnected from stocks and flows, it is difficult to determine their influence on the stock of a population with fatigue or how they then influence the rate at which a person becomes fatigued or recovers from fatigue. Hence, a complete SD model, which incorporates a stock and flow model with a causal loop model, is shown in Figure 5. After constructing the conceptual models, a formal mathematical model is constructed using coupled nonlinear
Figure 3 e Stock and Flow Diagram. There are two flows (“becoming fatigued” and “recovering from fatigue”) as well as one stock (“population with fatigue”). Stock represents the current state of the system, and the “flows” change the state of the system by increasing or decreasing at a certain rate.
317
Figure 4 e Causal Loop Diagram. As one state increases, the other states also increase. Hence, with an increase in the severity or frequency of sleep disturbance, fatigue and stress also increase.
differential equations and then simulated over time. For example, a complete SD model could be used to study factors that contribute to developing fatigue (e.g., stress, lack of physical activity, depression, or poor sleep) and factors that might facilitate recovering from fatigue (e.g., stress management, exercise, antidepressants, and cognitive behavioral therapy) to see how these factors affect each other. Researchers can then design an intervention based on manipulating parameters in the SD model to decrease rates of fatigue over time. There are a number of software options for SD modeling. The most popular are Stella (Isee Systems, 2014) Vensim (Vensim, 2010; Ventana Systems, Inc, 2014), and Powersim (Powersim Software AS, 2014). For more advanced users, there are open-source options such as simecol (Simecol, Dresden, Germany) (Jones, 2007; Petzoldt & Rinke, 2007) and StellaR packages in R (Naimi & Voinov, 2012). SD can also be done as a secondary feature in some agent-based modeling
Figure 5 e Complete SD Model. The polarity of the loop depends on the number of positive arrows relative to the number of negative arrows.
318
Nurs Outlook 62 (2014) 313e321
(ABM) software packages (NetLogo, Northwestern Institute on Complex Systems, Evanston, IL) (Thiele, 2014; Wilensky, 1997, 1999).
Agent Based Modeling (ABM) Similar to SD modeling, ABM is a computer-aided mathematical simulation method. The theoretical foundation of ABM is based on discrete mathematical models, including the Turing machine, cellular automata, and Conway’s Game of Life (Epstein, 2007; Epstein, Parker, Cummings, & Hammond, 2008; Gotts, 2009). ABM models the interactions among autonomous agents (i.e., individuals). Each autonomous agent or person can have multiple demographic characteristics and be assigned behavioral preferences that can be modeled over time (Verzi, Brodsky, Brown, Apelberg, & Rostron, 2012). ABM can be used to examine the consequences of behaviors over time. ABM would be particularly useful for symptom research because it also handles the complexities of groups and environments. ABM can include groups of individuals with certain characteristics, such as night shift workers, groups living in certain regions (above the Arctic Circle or in humid coastal regions), or work settings (intensive care, labor and delivery, etc.). ABM models could be used to understand how a health-related decision or behavior of an individual agent (night shift nurse) could lead to the development of chronic fatigue. This may be particularly useful for research on symptoms when there is initially no complaint of the symptom.
Unlike more traditional regression models, which might tell us context-level predictors of fatigue, an ABM can help us understand how individual characteristics and behaviors over time might contribute to the development of chronic fatigue in a specific environment in which individual choices are constrained by the characteristics of that environment. Similar to SD models, there are a number of software choices for creating ABM models. Two popular commercial options include NetLogo (Wilensky, 1997, 1999) and AnyLogic (XJ Technologies, 2009). For more advanced users, there are open-source options, such as simecol (Petzoldt, 2014) and RNetLogo (Thiele, 2014) packages in R (R Core Team, 2013). A screenshot example of an ABM is shown in Figure 6. This model (Wilensky, 1997, 1999) shows how a population of individuals infected with HIV would change over time based on their average coupling tendency, average commitment, average condom use, and average test frequency. Researchers can then manipulate the parameters to observe their impact on potential infection rates (74.5% infected in the example given the four parameters being modeled).
Markov Models Markov models (MMs) are a group of stochastic time series models based on the idea of a Markov property, which states that a future state of a process depends only on the present state and not on any previous states (Murphy, 2012). MMs are especially useful for understanding complex time series and have been widely used in the area of machine
Figure 6 e Screenshots of an ABM Using NetLogo to Model Rates of HIV Infection Over Time. Reprinted with permission from U. Wilensky, 1997. Setting parameters for four key variables (coupling tendency, time in committed relationship, condom use, and HIV testing frequency) using NetLogo software allows the researcher to observe how each of these parameters would impact HIV infection rates.
Nurs Outlook 62 (2014) 313e321
learning and artificial intelligence. MMs could be applied to understand the stage of disease development from data contained in large databases such as electronic health records (EHRs). The symptom experience in many types of chronic illness populations, including HIV/AIDS, cancer, mood disorders, and type 2 diabetes mellitus, could be modeled to examine patterns over time to identify disease development stages or potential unobserved variables influencing disease progression. This can eventually provide clues to potential interventions and data for patient teaching and anticipatory preparation for future health care needs. MMs can estimate stages of disease or symptom progression or can show multiple ways in which a disease or symptom could develop. A two-state MM example is seen in Figure 7; the circles represent states, and the number of arrows represents probabilities. For example, in a healthy state, the probability that a person will stay in a healthy state is 0.6, whereas the probability is 0.4 that the person will develop fatigue at the next state transition. These are called transition probabilities. With wider use of EHRs, nurse researchers can “train” MMs (in this case hidden MMs [HMMs]) to recognize symptom patterns, levels of pain, anxiety, or fatigue severity for example. Training HMMs consists of performing parameter estimation either on a subsample or full data in order to use these parameter estimates to classify new cases. For example, if we discover the process of becoming fatigued and have trained our HMM, we can use our parameter estimates to classify new cases as being at a certain hidden state or estimate the probability of moving from their current state to a different state. As seen in Figure 8, HMMs allow a researcher to group patients in an EHR dataset by these patterns or states and then design more precision-based interventions to test in future symptom research. An HMM can provide additional detail to better understand whether someone is moving from a healthy state to an unhealthy fatigued state. For example, we might know that there is a higher chance that someone starts in a healthy state, and we might also know the probabilities for feeling stressed, sleepy, or normal while in either a healthy state or fatigued state. These are called emission probabilities. We can easily tell which state someone is in, and these two states are
Figure 7 e Example of a Two-state Markov Chain. Circles represent states, and the numbers of arrows represent probabilities. In a healthy state, the probability that one will stay in a healthy state is 0.6, thus the probability that one will develop fatigue at the next state transition is 0.4.
319
Figure 8 e Example of an HMM. At a baseline measurement, there is a higher probability that a healthy individual will start in a healthy state rather than a fatigued state. The probabilities of feeling stressed, sleepy, or normal while someone is in either a healthy state or fatigued state are called emission probabilities. Fatigued state could then be approximated from the additional information in the model. said to be “hidden and can then be approximated from the additional information in the model. This has now transformed from a two-state Markov chain (MM) into an HMM. MMs can be formulated in MATLAB (MathWorks, Inc., 2014) using a modeling toolkit for a desired MM. Octave (GNU Octave) is an open-source version of MATLAB (Eaton, Bateman, & Hauberg, 2008). There are also a number of MM packages available in R (R Core Team, 2013).
Initial Training in Computational Models All three analytic tools presented in this article require a different skill set than taught in traditional graduatelevel statistics courses. There are many resources available for learning about the theoretical underpinnings for these types of models and how to apply these models to symptom science. Some of these learning resources are also available at no cost. Selflearning is encouraged for individuals with a limited range of programming experience (such as writing simple SPSS syntax) or having a background in elementary calculus. Individuals not experienced in running syntax-based mathematical models are encouraged to take a workshop or a class on one of the discussed approaches.
320
Nurs Outlook 62 (2014) 313e321
The System Dynamics Society (SDS) is an international nonprofit organization with a variety of resources for individuals wanting examples of SD models and wanting to learn about SD (http://www.systemdynamics.org). Similar to SDS, a beginning system scientist can find resources on ABM at a network maintained by ABM researchers (http://www.openabm.org/). There is no central online resource for HMMs, but an online class in machine learning would be beneficial for an introduction to the Markov process, with some elementary illustrations. For more advanced users, Murphy’s Matlab toolbox might be a good starting point (http://www.cs.ubc.ca/ wmurphyk/Software/HMM/hmm.html). For R users, Visser’s “depmixS4” package has a “Github” page that includes sample code and models (https://github.com/ cran/depmixS4).
Conclusions The symptom of fatigue was used in this article as an example to show the complexity of symptom science and the difficulties inherent in assessing and testing interventions designed to improve health outcomes. Other symptoms or symptom clusters could also be examined, including pain, nausea, thirst, tinnitus, insomnia, and dyspnea for example. The three examples of statistical models described in this article hold tremendous promise for increasing our understanding of complex phenomena in symptom research. All models like these will require a time component, whether using crosssectional or longitudinal data. As with any analytic model, the model is only as good as the hypothesis it is testing and the skill of the researcher. The model will not yield correct estimates if the underlying hypothesis is flawed. In addition, as in traditional structural equation or multilevel modeling, researchers should perform sensitivity analysis with ABMs, and model fit statistics should be used to assess the fit of HMMs. SD and ABM offer nurse scientists the potential to develop types of models, whether the data set is small or large. HMMs may perform better with big data than other modeling techniques. These types of models can include many more predictors and determinants of the intensity and distress for a given symptom, symptom cluster, or pattern of symptom experiences over time. This would be especially important in addressing the question of how and whether or not cancer-related fatigue (Barsevick et al., 2013) and HIV-related fatigue (Barroso & Voss, 2013) differ from other types of fatigue like shift work fatigue (Carruso, 2014; Geiger-Brown et al., 2012) or pregnancy fatigue (Gay et al., 2004). Although these analytic techniques are commonly used to understand complex systems of care and estimate potential risk or success with intervention strategies, applying these techniques to symptom science will expand our ability to intervene most effectively and with better precision. In addition, MMs provide a much more quantitative approach for determining a
patient’s improvement or potential for declining over time. This is a novel approach to symptom research that nurse scientists have strongly desired for decades. It is clear that symptom research is in need of these types of analytic techniques, and nurse researchers would do well to learn, collaborate, and incorporate these strategies into their science.
references
Akerstedt, T., Fredlund, P., Gillberg, M., & Jansson, B. (2002). Work load and work hours in relation to disturbed sleep and fatigue in a large representative sample. Journal of Psychosomatic Research, 53(1), 585e588. Aouizerat, B. E., Gay, C. L., Lerdal, A., Portillo, C. J., & Lee, K. A. (2013). Lack of energy: An important and distinct component of HIV-related fatigue and daytime function. Journal of Pain Symptom Management, 45(2), 191e201. Baars, E. W., Gans, S., & Ellis, E. L. (2008). The effect of hepar magnesium on seasonal fatigue symptoms: A pilot study. Journal of Alternative and Complementary Medicine, 14(4), 395e402. Barroso, J., & Voss, J. G. (2013). Fatigue in HIV and AIDS: An analysis of evidence. Journal of the Association of Nurses in AIDS Care, 24(1 Suppl), S5e14. Barsevick, A. M., Irwin, M. R., Hinds, P., Miller, A., Berger, A., Jacobsen, P., . Cella, D. (2013). National Cancer Institute clinical trials planning meeting. Recommendations for highpriority research on cancer-related fatigue in children and adults. Journal of the National Cancer Institute, 105(19), 1432e1440. Bevans, M., Ross, A., & Cella, D. (2014). Patient-reported outcomes measurement information system (PROMISÒ): Efficient, standardized tools to measure self-reported health and quality of life. Nursing Outlook, 62(5), 339e345. Blomkvist, V., Eriksen, C. A., Theorell, T., Ulrich, R., & Rasmanis, G. (2005). Acoustics and psychosocial environment in intensive coronary care. Occupational and Environmental Medicine, 62(3), e1. Caruso, C. C. (2014). Negative impacts of shiftwork and long work hours. Rehabilitation Nursing, 39(1), 16e25. Dodd, M. J., Miaskowski, C., & Lee, K. A. (2004). Symptom clusters: The new frontier in symptom management research. Journal of the National Cancer Institute Monographs, 2004(32), 17e21. Eaton, J. W., Bateman, D., & Hauberg, S. (2008). GNU Octave Manual Version 3: A high-level interactive language for numerical computations. Bristol, UK: Network Theory Ltd. Epstein, J. M. (2007). Generative social science: Studies in agent-based computational modeling. Princeton University Press. Retrieved from http://press.princeton.edu/chapters/s8277.html. Epstein, J. M., Parker, J., Cummings, D., & Hammond, R. A. (2008). Coupled contagion dynamics of fear and disease: Mathematical and computational explorations. PLoS One, 3(12). http://dx.doi.org/10.1371/journal.pone.0003955. Gander, P. H., Signal, T. L., van den Berg, M. J., Mulrine, H. M., Jay, S. M., & Jim Mangie, C. (2013). In-flight sleep, pilot fatigue and psychomotor vigilance task performance on ultra-long range versus long range flights. Journal of Sleep Research, 22(6), 697e706. Gay, C. L., Lee, K. A., & Lee, S. Y. (2004). Sleep patterns and fatigue in new mothers and fathers. Biological Research for Nursing, 5(4), 311e318. Geiger-Brown, J., Rogers, V. E., Trinkoff, A. M., Kane, R. L., Bausell, R. B., & Scharf, S. M. (2012). Sleep, sleepiness, fatigue,
Nurs Outlook 62 (2014) 313e321
and performance of 12-hour-shift nurses. Chronobiology International, 29(2), 211e219. Gotts, N. M. (2009). Ramifying feedback networks, cross-scale interactions, and emergent quasi individuals in Conway’s game of Life. Artificial Life, 15(3), 351e375. Hallowell, M. R. (2010). Worker fatigue: Managing concerns in rapid renewal highway construction projects. Professional Safety (pp. 18e26), Retrieved from http://www.asse.org/professional safety/pastissues/055/12/018_026_F1Hallo_1210Z.pdf. Hammond, R. A. (2009). Complex systems modeling for obesity research. Preventing Chronic Disease, 6(3), A97. Hammond, R. A., & Dube´, L. (2012). A systems science perspective and transdisciplinary models for food and nutrition security. Proceedings of the National Academy of Sciences of the United States of America, 109(31), 12356e12363. Han, K., Trinkoff, A. M., Storr, C. L., & Geiger-Brown, J. (2011). Job stress and work schedules in relation to nurse obesity. Journal of Nursing Administration, 41(11), 488e495. Hartzler, B. M. (2014). Fatigue on the flight deck: The consequences of sleep loss and the benefits of napping. Accident Analysis & Prevention, 62(1), 309e318. Isee Systems, Inc. (2014). STELLA. Isee Systems, Inc. Retrieved from http://www.iseesystems.com/softwares/Education/ StellaSoftware.aspx. Jones, O. (2007). Simco: A Package to Import Structure Files and Deduce Similarity Coefficients from Them. R package version, 1, 2007. Kassardjian, C. D., Murray, B. J., Kokokyi, S., Jewell, D., Barnett, C., Bril, V., . Katzberg, H. D. (2013). Effects of napping on neuromuscular fatigue in myasthenia gravis. Muscle & Nerve, 48(5), 816e818. Konkani, A., Oakley, B., & Penprase, B. (2014). Reducing hospital ICU noise: A behavior-based approach. Journal of Healthcare Engineering, 5(2), 229e246. Lee, K. A., Gay, C. L., Lerdal, A., Pullinger, C. R., & Aouizerat, B. E. (2014). Cytokine polymorphisms are associated with fatigue in adults living with HIV/AIDS. Brain, Behavior, & Immunity, 40, 95e103. Lee, K. A., & Lipscomb, J. (2003). Sleep among shiftworkers-A priority for clinical practice and research in occupational health nursing. American Association of Occupational Health Nursing Journal, 51(10), 418e420. Lerdal, A., Gay, C. L., Aouizerat, B. E., Portillo, C. J., & Lee, K. A. (2011). Patterns of morning and evening fatigue among adults with HIV/AIDS. Journal of Clinical Nursing, 20(15-16), 2204e2216. Lerdal, A., Lee, K. A., Bakken, L. N., Finset, A., & Kim, H. S. (2012). The course of fatigue during the first 18 months after first-ever stroke: A longitudinal study. Stroke Research and Treatment, 2012, article 126275. http://dx.doi.org/10.1155/2012/126275. Liu, L., Rissling, M., Neikrug, A., Fiorentino, L., Natarajan, L., Faierman, M., . Ancoli-Israel, S. (2013). Fatigue and circadian activity rhythms in breast cancer patients before and after chemotherapy: A controlled study. Fatigue, 1(1-2), 12e26. MathWorks, Inc. (2014). MATLABÓ. MathWorks, Inc. Retrieved from http://www.mathworks.com/products/matlab/. Meek, P. M., & Lareau, S. C. (2003). Critical outcomes in pulmonary rehabilitation: Assessment and evaluation of dyspnea and fatigue. Journal of Rehabilitation Research and Development, 40(5 Suppl 2), 13e24. Miaskowski, C., Dodd, M., Lee, K., West, C., Paul, S. M., Cooper, B. A., . Aouizerat, B. E. (2010). Preliminary evidence of an association between a functional interleukin-6 polymorphism and fatigue and sleep disturbance in oncology patients and their family caregivers. Journal of Pain and Symptom Management, 40(4), 531e544. Miaskowski, C., Lee, K., Dunn, L., Dodd, M., Aouizerat, B. E., West, C., . Swift, P. (2011). Sleep-wake circadian activity
321
rhythm parameters and fatigue in oncology patients before the initiation of radiation therapy. Cancer Nursing, 34(4), 255e268. Miaskowski, C., Paul, S. M., Cooper, B. A., Lee, K., Dodd, M., West, C., . Wara, W. (2008). Trajectories of fatigue in men with prostate cancer before, during, and after radiation therapy. Journal of Pain and Symptom Management, 35(6), 632e643. Murphy, K. P. (2012). Machine learning: A probabilistic perspective. MIT Press. Naimi, B., & Voinov, A. (2012). StellaR: A software to translate Stella models into R open-source environment. Environmental Modelling & Software, 38(1), 117e118. Petzoldt, T., & Rinke, K. (2007). Simecol: An object-oriented framework for ecological modeling in R. Journal of Statistical Software, 22(9), 1e31. Petzoldt, T. (2014). Simulation of ecological (and other) dynamic systems. Retrieved from http://cran.r-project.org/web/ packages/simecol/simecol.pdf. Powersim Software AS. (2014). Powersim Studio 10Ò. Power Sim Software ASÓ. Retrieved from http://www.powersim.com/. R Core Team. (2013). R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing, ISBN 3-900051-07-0. URL. http://www.R-project.org/. Rastad, C., Ulfberg, J., & Lindberg, P. (2011). Improvement in fatigue, sleepiness, and health-related quality of life with bright light treatment in persons with seasonal affective disorder and subsyndromal SAD. Depression Research and Treatment, 2011, 543906. http://dx.doi.org/10.1155/2011/ 543906. Richardson, G. P. (2013). System dynamics. In S. Gass, & C. Harris (Eds.), Encyclopedia of operations research and management science (pp. 1519e1522). Kluwer Academic Publishers. Shen, J., Botly, L. C., Chung, S. A., Gibbs, A. L., Sabanadzovic, S., & Shapiro, C. M. (2006). Fatigue and shift work. Journal of Sleep Research, 15(1), 1e5. Sterman, J. D. (2000). Business dynamics: Systems thinking and modeling for a complex world. Boston: Irwin/McGraw-Hill. Thiele, J. C. (2014). R marries NetLogo: Introduction to the RNetLogo package. Journal of Statistical Software, 58(2), 1e41. Vensim, P. (2010). Ventana Systems, Inc. Retrieved from http:// www.vensim.com. Ventana Systems, Inc. (2014). Vensim PLE SoftwareÒ. Ventana Systems Inc. Retrieved from http://vensim.com/. Verzi, S. J., Brodsky, N. S., Brown, T. J., Apelberg, B., & Rostron, B. (2012). An agent-based approach for modeling population behavior and health with application to tobacco use. Sandia National Laboratories. Voss, J. G., Dobra, A., Morse, C., Kovacs, J. A., Danner, R. L., Munson, P. J., . Dalakas, M. C. (2013). Fatigue-related gene networks identified in CD(14)þ cells isolated from HIVinfected patients: Part I: Research findings. Biological Research for Nursing, 15(2), 137e151. Voss, J. G., Dodd, M., Portillo, C., & Holzemer, W. (2006). Theories of fatigue: Application in HIV/AIDS. Journal of the Association of Nurses in AIDS Care, 17(1), 37e50. Wilensky, U. (1997). NetLogo AIDS Model. Evanston, IL: Center for Connected Learning and Computer-Based Modeling, Northwestern University. Retrieved from http://ccl. northwestern.edu/netlogo/models/AIDS. Wilensky, U. (1999). NetLogo. Evanston, IL: Center for Connected Learning and Computer-Based Modeling, Northwestern University. Retrieved from http://ccl.northwestern.edu/ netlogo/. XJ Technologies. (2009). AnyLogicÒ. XJ technologies. Retrieved from http://www.xjtek.com/.
