Journal of Science and Medicine in Sport 26 (2023) S40–S45

Contents lists available at ScienceDirect

Journal of Science and Medicine in Sport

journal homepage: www.elsevier.com/locate/jsams

Review

Developing AI enabled sensors and decision support for military

operators in the field
Brian K. Russell a,b,⁎, Josh McGeown d, Bettina L. Beard c
a
Sports Performance Institute of New Zealand, Auckland University of Technology, New Zealand
b
Ambient Cognition Limited, Aukland, New Zealand
c
NASA Ames Research Center, Moffett Field, USA
d
Matai Medical Research Institute Inc, New Zealand

a r t i c l e i n f o a b s t r a c t

Article history: Wearable sensors enable down range data collection of physiological and cognitive performance of the
Received 4 June 2022 warfighter. However, autonomous teams may find the sensor data impractical to interpret and hence influence
Received in revised form 17 February 2023 real-time decisions without the support of subject matter experts. Decision support tools can reduce the burden
Accepted 1 March 2023
of interpreting physiological data in the field and incorporate a systems perspective where noisy field data can
Available online 5 March 2023
contain useful additional signals. We present a methodology of how artificial intelligence can be used for model-
Keywords:
ing human performance with decision-making to achieve actionable decision support. We provide a framework
Remote physiological monitoring for systems design and advancing from the laboratory to real world environments. The result is a validated mea-
Cognition performance sure of down-range human performance with a low burden of operation.
Artificial Intelligence © 2023 Sports Medicine Australia. Published by Elsevier Ltd. All rights reserved.
Decision Support

Practical implications by utilizing wearable sensors and ambient detection. This real-time
monitoring of an individual's physiological and cognitive status can po-
• Raw data from wearable sensors and remote physiological monitoring tentially aid in tactical decisions down range. For example, heart rate
is not actionable in an operational setting. monitors allow estimation of an operator's physical performance and
• The burden of monitoring sensors can be reduced with AI models. energy systems during exercise/exertion. However, heart rate alone
• Decision support systems can increase the knowledge, skills and abil- does not facilitate an actionable RT-PSM system due to secondary effects
ities of operators in the field. such as: physical load; body position (e.g., laying vs standing); recovery
• Autonomy can be increased when using decision support tools for from recent exercise; sleep status; emotional state (e.g., relaxed vs
human-machine teaming. afraid); dehydration; digestion of food/ingestion of ergogenic aids
(e.g., caffeine); core temperature (e.g., vasodilation from temperature
or blood loss) and possible viral infection.
The challenge of interpreting human performance data from a RT-
1. Decision-support system
PSM system can be solved by incorporating it into a decision support
system (DSS). A DSS combines sensor data with artificial intelligence
In safety-critical professions (e.g., warfare, aerospace, firefighting,
(AI) models to deliver problem solving, actionable choices and
piloting, space travel, air traffic control, and sports) individuals must
observation.1 This article proposes a framework to determine the devel-
make rapid tactical decisions based on many data sources. Real time
opment of an AI assisted DSS including RT-PSM and reviews common
monitoring of an individual's physiological and cognitive status can po-
goals of DSS, including that the system should:
tentially aid the speed and accuracy of tactical decisions. Recent devel-
opments have combined sensors with artificial intelligence (AI) to
deliver precision medicine with individualized metrics of health and • Be low burden to for operator time, cognitive load and physical
performance. distractions2
Remote Physiological Status Monitoring (RT-PSM) systems use sen- • Be low size, weight and power (SWaP)
sors to measure and calculate health and performance of an individual • Be context aware of the current situation
• Augment the operators' knowledge, skills and abilities (KSA)3
⁎ Corresponding author. • Be based on an appropriate underlying hierarchy of models
E-mail address: [email protected] (B.K. Russell). • Share information from a database of decision-relevant information

https://doi.org/10.1016/j.jsams.2023.03.001
1440-2440/© 2023 Sports Medicine Australia. Published by Elsevier Ltd. All rights reserved.
B.K. Russell, J. McGeown and B.L. Beard Journal of Science and Medicine in Sport 26 (2023) S40–S45

• Provide information that aligns with the decision-maker's expecta- In the next section we discuss the application of artificial intelligence
tions and mental model4 (AI), as a systems-based computational tool that can learn the complex
• Support judgments rather than replace them relationships between system inputs and outputs without explicit
• Operate at a high level of trust.5 programming.22

2. Evolution of decision-making models toward a systems perspective 3. Artificial intelligence

Decision-making is a problem-solving activity that yields a solution. A model in the context of computer science refers to software code
Several Naturalistic Decision Making (NDM) models have been that performs a function. AI models can achieve human level accuracy
developed,6 one of which is referred to as the Recognition-Primed Deci- performing certain cognitive tasks. These models range from simple
sion Model (RPM).7 Briefly, the RPM hypothesizes that people use their classification models (e.g., cats vs dogs), regression models calculating
prior experiences to build a repertoire of patterns, including training, continuous variables (e.g., human activity recognition from acceleration
not just real-world experiences. These patterns include relevant cues or heart rate from plethysmograph waveforms) to more complex com-
and suggest typical types of reactions for that type of situation. If a com- putations including probabilistic decision making. Broadly, AI models
mander, for example, needs to make a decision, they quickly match are split into supervised or unsupervised learning algorithms. Super-
the situation to learned patterns and visualize how that same course vised AI models require labeled data to learn patterns that differentiate
of action will play out in the current context. When there is a clear categories (e.g. cat vs dog). Unsupervised models determine correla-
match and the mental simulation suggests that a previously used so- tions which assist with labeling and dissemination in decision making
lution would work in the current context, the commander will carry systems.10
out that action. Therefore, decision-making in-the-field is a blend of Deep learning is a subset of AI that automatically learns complex re-
intuition and analysis. lationships between high dimensional inputs to determine the most im-
In 1995, John Boyd 8 first presented the now well-recognized portant features related to the outputs of interest.11,12 Deep learning
OODA loop of human decision making (Fig. 1). The OODA loop con- models are commonly described as opaque, or “black box” models as
cept describes decision-making as a cycle between observations, they offer a high degree of predictive accuracy at the expense of
orienting, deciding and acting. Basically, as a situation evolves the explainability of the underlying relationships between inputs and out-
decision-maker observes, with attention primarily focused on the puts. Trust is important when using predictive models and typically in-
current problem. Orientation focuses observations based on prior ex- volves explainability and validation results. Table 1 provides a subset of
perience and knowledge drives the focus. It is here that the observa- computational models ordered by explainability. AI and decision sup-
tions are processed. Decisions and actions are based on the current port tools requiring trust have been used successfully in fields such as
information and knowledge of possible choices. The tempo of the bioinformatics13,14 and neuroradiology.15,16
OODA cycle is critical in situations where the goal is to out-think an In the subsequent sections of this discussion, we describe a frame-
opponent (e.g., during a battle, on a football field, medicine-where work of how AI/deep learning can aid in the exploration and develop-
the opponent is the illness or disease). ment of a DSS.
Other recent research and theoretical endeavors have focused on
human “anomaly resolution” by individuals and groups. For example, 4. Finding the signal in environmental complexity with AI
Woods9 defined an iterative process composed of recognition, trouble-
shooting, response, and contingency management. To resolve an anom- A systems theory perspective assumes the whole is greater than the
aly, it must first be detected. This involves recognizing anomalous data sum of its parts. Ahn17 described the systems perspective succinctly by
or an anomalous situation. The source of the anomaly must then be di- stating “the systems perspective is rooted in the assumption that the
agnosed through troubleshooting. This is where prior experience or forest cannot be explained by studying the trees individually”. A defin-
training has established expectations that focus attention to what is rel- ing feature of the systems perspective is how environmental factors
evant. The initial interpretation and response affects how the anomaly that introduce variability within the system are viewed. Rather than
will progress. viewing this variability as nuisance noise to be methodologically

Fig. 1. Boyd's OODA loop, a cognitive model for decision making.

S41
B.K. Russell, J. McGeown and B.L. Beard Journal of Science and Medicine in Sport 26 (2023) S40–S45

Table 1
Explainability listed by examples of computational approaches.

Explainability Type of model Example type Example application

Highly transparent Linear algebra linear model Modeling muscle respiration40

Calculus Differential equation Skin surface temperature modeling41
Statistics ANOVA Effectiveness of different treatments42
Fuzzy logic Fuzzy model Diagnosis of cardiac hypertrophy43
Bayesian models Conditional probability Diagnosis of medical condition44
Tree based Random Forest Prediction of energy expenditure45
Component Analysis Principal Component Analysis Facial recognition4
Feature-based machine learning Support Vector Machine Predicting Ion exposure form cognitive performance46
Artificial intelligence Convolutional Neural Network Classification of human activity19 regression analysis of ground reaction forces47
Highly opaque Artificial intelligence Recurrent Neural Network Prediction of cognitive load from speech48

controlled or removed, a systems perspective views this “noise” as a smooth surface. Russell19 showed these waveforms could be used by
crucial component required to predict behavior within the system. Tra- an AI model to detect signals such as surface type, slope, activity type
ditional laboratory assessment protocols require an intervention with a and fatigue.20 The environmental signal was noise for activity recogni-
measurable outcome (e.g., jump height or Stroop Effect). Commonly, tion but the signal of interest when detecting surface type.
field studies will utilize a battery of tests and determine which were
sensitive to the protocol load after completing the protocol. However,
5. A hierarchical system of models
these interactive tests are not practical for operators in the field as
they are distracting and require cognitive and physical interaction. Al-
Implementation of decision support typically involves multiple AI
ternatively, ambient detection observes an individual's performance at
models. Crump21 discusses an AI stack for implementation of an ethical
a task to compute an output value, e.g., the timing from keyboard use
medical decision support system based on the ISO 7 layer stack used in
can be combined with an AI model to determine cognitive fatigue.18 A
communications engineering. A decision support system is imple-
decision support tool can integrate this information to reduce interac-
mented in an hierarchical architecture, such as the OODA loop, with
tion and only require cognitive burden when guiding attention to an
computed information being passed up and feedback loops generating
anomaly and assisting with decision support.
new data for lower stages. Each stage of the decision making process
Consider the example waveforms presented in Fig. 2, from Russell,19
can be considered as a hierarchy of opportunity for computational
illustrating an acceleration waveform of a person running outdoors
tools to assist the human operator. Observation is automated through
down an incline on a smooth sealed road versus a dirt track. The
sensor data collection including ambient detection22 to reduce the
shape and silhouettes of each waveform demonstrate that running on
need for additional activities. AI models can assist with orientation by
a dirt track introduces more complex gait mechanics compared to a
interpreting data, often better than a human, by tirelessly observing
time series data such as accelerometery and determining higher order
features such as fatigue. The decision state is typically modeled with a
Bayesian approach to probabilistically determine which information is
required to differentiate between decisions using the highest positive
predictive value. Actions are chosen by an expert systems model
(Fig. 3).
The following section provides practitioners with a systematic
methodology for identifying valid, non-invasive performance measures
for the prediction of real-world task performance decline so that useful
countermeasures, such as decision support or information integration,
may be proactively provided.

6. Selection of valid performance measures

6.1. Measure dependencies

An individual's ability to perform in a range of safety-critical tasks

can be fatigued by stressors which negatively affect physical and cogni-
tive abilities. Stressors shown to affect performance include exposure to
hot or cold temperatures,23 pain,24 physical inactivity,25 isolation,26
time pressure,27 concussion,28 high workload29 and sleep deficits, to
name a few.
The physical or cognitive ability affected (e.g., visuo-motor responses,
sustained attention, selective attention, higher level problem-solving) de-
pends on the stressor type, duration, and intensity. As examples, time
pressure can result in overlooked critical information,30 concussions
have an association with reduced mental flexibility,31 and sleep depriva-
tion has been shown to affect concentration.32 To further complicate mat-
ters, each individual reacts differently to stressors.33,34
Fig. 2. IMU waveforms for running down slopes of various terrain types (sealed road vs Many studies have explored methods for detecting, monitoring,
dirt track),19 and aiding performance. However, poor transferability of these

S42
B.K. Russell, J. McGeown and B.L. Beard Journal of Science and Medicine in Sport 26 (2023) S40–S45

Fig. 3. Hierarchy of computational models to the OODA decision making loop.

methods from simulation to the field suggests that either different who is piloting a robotic arm located on the outside of the ISS. The astro-
cognitive processes are being assessed or that the complexity of op- naut is being exposed to microgravity, a high workload, possible vision
erations have not been captured. In many cases, technology selection changes, elevated CO2 in the cabin atmosphere, an audience of millions,
is based on cost, availability, or successful application in another do- etc. All stressors should be identified.
main. Little attention is given to determining how changes in activity
or sensor profiles compare to the individual's (or group's) tasks at 6.4. Align stressors with critical tasks
hand, what cognitive or motor abilities are needed to accomplish
those tasks, the stressors currently being experienced, or how per- Next, determine how performance has been compromised with ex-
formance or those tasks are affected by the individual or combined posure to each stressor alone, and in concert.35 This alignment is based
stressors. on the published, scientific literature. For example, elevated CO2 has
been shown to affect visuo-motor abilities needed to adjust the robotic
6.2. Define safety-critical tasks arm position.36 Therefore, sensors indicating elevated CO2 could
communicate with technology that adaptively provides additional
Each position in a profession has specific roles and responsibilities. visuo-motor guidance. As another example, imagine a scenario where
Task analysis, performed by a trained human factors specialist, identifies a single fighter pilot must navigate to a specific target in a well-hidden
the tasks that must be supported for a particular position. If tasks in- location to eliminate a threat. To find the location may take several
clude safety-critical mental effort or critical decisions, then a cognitive hours at which time he will be running low on fuel, and he is not
task analysis (CTA) should be performed. CTA uncovers what the indi- equipped with a water canteen. Both time–pressure and dehydration
vidual needs to know (e.g., information needs) and strategies used to have been shown to affect spatial navigation, even in expert pilots.37
think through the decision. In addition to helping define a sensible per- Whenever possible, the correlation between the field stressors and the
formance metric, CTA is invaluable for training development, interface specific real-world task identified should be validated.
design and procedure development. In summary, knowledge of specific safety-critical tasks and known
Once a complete CTA has been performed, pinpoint those tasks that stressors that affect performance on that task can help define a prag-
are performed during safety-critical operations. For example, a TRACON matic performance metric. It is likely that a combination of measures
arrival air traffic controller must often integrate visual information from will be needed. Additionally, knowledge about potential degraded per-
a variety of displays (e.g., arrival rates, aircraft speeds, wind, the time, formance can be used a priori to make design decisions and to deter-
projected arrivals). Therefore, visual information integration is a candi- mine effective countermeasures.
date ability to monitor for performance decline and/or for which to de-
velop technological support. 7. AI research and development framework

6.3. Define stressors A challenge when developing AI models with wearable sensor data
is how to efficiently and effectively handle the sheer volume of data
The specific stressors the population will be exposed to prior to and that these sensors provide. The researcher and developer have choices
during the time-frame of the critical decisions should also be deter- on what sensor types to use, how frequently data should be sampled,
mined. Imagine an astronaut on the International Space Station (ISS) stored, transmitted, or analyzed. Often the raw sensor data is too

S43
B.K. Russell, J. McGeown and B.L. Beard Journal of Science and Medicine in Sport 26 (2023) S40–S45

Table 2 use operational tasks as a stimulus to replace protocols. AI can use envi-
Research protocol for AI enabled models. ronmental noise as signal and expert decision making can bring com-
Stage Decision or task Description plex probabilistic decision-making down range increasing the
1 Determine mission Outputs required to be valuable to the operator
knowledge skills and abilities of operators.
outputs and the mission. Establish how these Adapting to field conditions has challenges that, if overcome, prom-
requirements will be quantified. ise to increase performance, without adding additional burden to re-
2 Select Clinical Used for AI model training and validation mote autonomous operators down range. Field noise is both a
Assessment Battery
challenge and a stimulus for the ambient detection of response to the
3 Determine field Representative data for different scenarios,
protocol for data sufficient collection of assessment battery to stimulus when assessing performance or other parameters. Data collec-
collection enable high resolution of AI training tion for research with AI modeling requires special field protocols. This
4 Data Collection Tools To be used by research staff on field, that may be systems approach offers the promise of decision support tools aiding
fatigued and distracted with intermittent power, the warfighter operating autonomously down range in the near future.
data logging and communications
5 Label Data Add labels to data to enable AI model to learn
6 Protocol study Perform prototype study on low subject Funding information
numbers to prove functionality of the protocol,
tools and computational models. No grants or financial contributions were made for this paper.
7 Large study Perform field trial to validate the research.
8 Deploy Deploy fieldable tool
Declaration of interest statement

verbose for system limitations including battery power, storage capacity B. L. Beard is employed by NASA Ames.
or communication bandwidth. The trade-offs require local analysis to J. McGeown is implied by The Matai Institute.
deliver higher level information for onward transmission, e.g., a plethys- B. K. Russell is a research associate of SPRINZ, and contractor to NASA
mograph (PPG) 100 Hz 2 channel signal from a wrist device is used to Ames
calculate and store blood oxygen (SpO2) versus heart rate at 1 Hz.
Validation of opaque AI models requires black box testing under Acknowledgements
many representative scenarios to guarantee performance in uncon-
strained environments. With AI development these data are required The work presented here was informed by research into clinical
at the beginning of the research to train the model and also at the end decision support systems for deep space missions by the NASA Human
of the project to validate performance. AI enabled human performance Research Program, Exploration Medicine Capability, NASA Ames
monitoring requires a method to gather data in the field to both train Research Facility.
the AI model and validate its performance. This ‘black box’ approach dif-
fers from reductionist science in that it asks the question: Transparency declaration
“Can an AI model accurately predict outputs with a given dataset
This article is published as part of a Supplement entitled ‘Macquarie
and, if so, what data is required to find the signal?”
University [MQ] Military Biomarker Symposium 2022’ supported by
A process is required to develop an AI model including input data the Defence Science and Technology Group (DSTG). Funding for the
representing real world use cases, outputs valuable for intended use, Symposium was provided by Office of Naval Research Global through
and comparison gold standards to both train the model and validate the Collaborative Science Program. Additional funding was provided
performance. Field considerations for the team collecting data in remote by the NSW Defence Innovation Network, Smartabase/Fusion Sport,
environments should take into account: spare parts due to lack of resup- and Vicon.
ply, fatigued and distracted research assistants, safety, environmental
noise (e.g., weather, terrain including slope and surface types, obstacles, References
non-uniform inclines and surfaces), operational noise (e.g., fatigue, 1. Sutton RT. An overview of clinical decision support systems: benefits, risks, and strat-
sleep deprivation, operational distractions such as contact with the egies for success. NPJ Digit Med 2020:1-10.
enemy or wild life) and performance selection such as self-pacing and 2. Friedl KE. Military applications of soldier physiological monitoring. J Sci Med Sport
2018;21(11):1147-1153. doi:10.1016/j.jsams.2018.06.004.
activity type (e.g., walking, running or laying during rest periods), sen- 3. Beard BL, Russell BK, Krihak M et al. Supporting Crew Medical Decisions on Deep Space
sors to represent fieldable solution options (SWaP), additional sensors Missions: A Real-time Performance Monitoring Capability, Leiden, The Netherlands, Int
needed for logistics, and protocols to generate required loads (physio- Assoc Adv Space Saf., IAASS, 2021.
4. Medlock S, Wyatt JC, Patel VL et al. Modeling information flows in clinical decision
logical, centraland psychological).20,38–40 support: key insights for enhancing system effectiveness. J Am Med Inform Assoc
This method is recommended with low subject numbers initially, N, 2016;23(5):1001-1006. doi:10.1093/jamia/ocv177.
as the logistics of field experiments is complicated and data analysis and 5. Frost SA. A briefing on metrics and risks for autonomous decision making in aero-
space applications. NASA Tech Memo 2018:1-13.
processing is time consuming. Pilot experiments will inform the proto-
6. Lipshitz R, Strauss O. Coping with uncertainty: a naturalistic decision making analy-
col and validating AI models with small N numbers allows the approach sis. Organ Behav Hum Decis Process 1997;69(2):149-163.
to be proven. It is recommended that generalizability to larger popula- 7. Klein G. The reinvention of decision making. Hum Factors 2008;50(3):456-460.
tions is considered an important second step. Of particular note is Step 8. Boyd JA. Patterns of conflict, Discourse on Winning and Losing, Maxwell Air Force
Base: Air University Press, 2018. p. 19.
5 in Table 2 as this is unique to AI modeling where data are labeled 9. Woods D, Hollnagel E. Joint cognitive systems, Jt Cogn Syst Patterns Cogn Syst Eng.
with additional information to enable the training step where the AI Boca Raton, Taylor and Francis, 2006. p. 69-96.
model learns iteratively by measuring its predictive accuracy against 10. Davis Z. Artificial intelligence on the battlefield. Inst Natl Strateg Secur 2019;8(2):117-
131.
the labels.
11. Esteva A, Robicquet A, Ramsundar B et al. A guide to deep learning in healthcare. Nat
Med 2019;25(1):24-29. doi:10.1038/s41591-018-0316-z.
12. Harrison CJ, Sidey-Gibbons CJ. Machine learning in medicine: a practical introduction
8. Conclusion to natural language processing. BMC Med Res Methodol 2021;21(1):1-18. doi:10.
1186/s12874-021-01347-1.
13. Mishra R, Li B. The application of artificial intelligence in the genetic study of
Remote physiological monitoring can be useful when incorporated
Alzheimer’s disease. Aging Dis 2020;11(6):1567-1584. doi:10.14336/AD.2020.0312.
into a DSS enabled with a system of AI models. This approach enables 14. Milanesi L, Romano P, Castellani G et al. Trends in modeling biomedical complex sys-
a lower number of sensors, with lower SWaP, as ambient sensing can tems. BMC Bioinformatics 2009;10(SUPPL 12):1-13. doi:10.1186/1471-2105-10-S12-I1.

S44
B.K. Russell, J. McGeown and B.L. Beard Journal of Science and Medicine in Sport 26 (2023) S40–S45

15. Pedersen M, Verspoor K, Jenkinson M. Artificial intelligence for clinical decision sup- 33. Cowings P, Toscano W, Deroshia C et al. Converging indicators for assessing individ-
port in neurology. Brain Commun 2020;1:1-11. doi:10.1093/braincomms/fcaa096. ual differences in adaptation to extreme environments. Aviat Sp Environ Med 2007;78
16. Farzaneh N, Williamson CA, Gryak J et al. A hierarchical expert-guided machine (5):195-215.
learning framework for clinical decision support systems: an application to traumatic 34. Fillingim RB. Inidividual differences in pain: understanding the mosaic that makes
brain injury prognostication. Npj Digit Med 2021;4(1):1-9. doi:10.1038/s41746-021- pain personal. 2018;158(Suppl 1):1-18. doi:10.1097/j.pain.0000000000000775.
00445-0. Individual.
17. Ahn AC, Tewari M, Poon CS et al. The limits of reductionism in medicine: could sys- 35. Beard BL. Characterization of how CO2 level may impact crew performance related to
tems biology offer an alternative? PLoS Med 2006;3(6):0709-0713. doi:10.1371/ the HSIA risk, NASA Technical Meorandum, NASA/TM-20205011433, 2020.
journal.pmed.0030208. 36. Manzey D, Lorenz B, Plojakov V. Mental performance in extreme environments: re-
18. Parekh V, Shah D, Shah M. Fatigue detection using artificial intelligence framework. sults from a performance monitoring study during a 438-day spaceflight. Ergonomics
Augment Hum Res 2020;5(1). doi:10.1007/s41133-019-0023-4. 1998;41(4):537-559. doi:10.1080/001401398186991.
19. Russell B, McDaid A, Tscano W et al. Moving the lab into the mountains: a pilot study 37. Gonzalez C. Task workload and cognitive abilities in dynamic decision making. Hum
of human activity recognition in unstructured environments. Sensors 2021;21(2): Factors 2005;47(1):92-101. doi:10.1518/0018720053653767.
654. 38. Van Cutsem J, Marcora S, De Pauw K et al. The effects of mental fatigue on physical
20. Russell B, McDaid A, Toscano W et al. Predicting fatigue in long duration mountain performance: a systematic review. Sports Med 2017;47(8):1569-1588. doi:10.1007/
events with a single sensor and deep learning model. Sensors 2021;21(5442). doi: s40279-016-0672-0.
10.3390/s21165. 39. St. Clair Gibson A, Baden DA, Lambert MI et al. The conscious perception of the sen-
21. Crump C, Schlachta-Fairchild LM. Achieving a Trusted, Reliable, AI-ready Infrastructure sation of fatigue. Sports Med 2003:167-176. doi:10.2165/00007256-200333030-
for Military Medicine and Civilian Trauma Care, 2020:19. doi:10.1117/12.2557514. 00001.
22. Nweke HF, Teh YW, Mujtaba G et al. Data fusion and multiple classifier systems for
40. Meyer RA. A linear model of muscle respiration explains monoexponential phospho-
human activity detection and health monitoring: review and open research direc-
creatine changes. Am J Physiol Cell Physiol 1988;254(4):3-8. doi:10.1152/ajpcell.1988.
tions. Inf Fusion 2019;46:147-170. doi:10.1016/j.inffus.2018.06.002.
254.4.c548.
23. Pilcher JJ, Nadler E, Busch C. Effects of hot and cold temperature exposure on perfor-
41. Khanday MA. Numerical study of partial differential equations to estimate thermo-
mance: a meta-analytic review. Ergonomics 2002;45(10):682-698. doi:10.1080/
regulation in human dermal regions for temperature dependent thermal conductiv-
0014013021015841.
ity. J Egypt Math Soc 2014;22(1):152-155. doi:10.1016/j.joems.2013.05.006.
24. Attridge N, Noonan D, Eccleston C et al. The disruptive effects of pain on n-back task
42. Krzywinski M, Altman N. Analysis of variance and blocking. Nat Methods 2014;11(7):
performance in a large general population sample. Pain 2015;156:1885-1891.
699-700. doi:10.1038/nmeth.3005.
25. Prakash RS, Voss MW, Erickson KI, et al. Physical activity and cognitive vitality n.d.
https://doi.org/10.1146/annurev-psych-010814-015249. 43. Dorn GW. The fuzzy logic of physiological cardiac hypertrophy. Hypertension 2007;49
26. Cacioppo JT, Cacioppo S. Social relationships and health: the toxic effects of percieved (5):962-970. doi:10.1161/HYPERTENSIONAHA.106.079426.
social isolation. Soc Pers Psychol Compass 2014;8(2):58-72. doi:10.1111/spc3.12087. 44. Bours MJ. Bayes’ rule in diagnosis. J Clin Epidemiol 2021;131:158-160. doi:10.1016/j.
Social. jclinepi.2020.12.021.
27. Kerstholt JH. Decision Making in a Dynamic Task Environment: The Effect of Time 45. Ellis K, Kerr J, Godbole S et al. A random forest classifier for the prediction of energy
Pressure, Institute for Perception, TNO Soesterberg Newtherlands, 1992. expenditure and type of physical activity from wrist and hip accelerometers. Physiol
28. Schatz P, Pardini JE, Lovell MR et al. Sensitivity and specificity of the ImPACT test bat- Meas 2014;35(11):2191-2203. doi:10.1088/0967-3334/35/11/2191.
tery for concussion in athletes. 2006;21:91-99. doi:10.1016/j.acn.2005.08.001. 46. Prelich MT, Matar M, Gokoglu SA et al. Predicting space radiation single ion exposure
29. Fothergill S, Neal A. The Effect of Workload on Conflict Decision Making Strategies in Air in rodents: a Machine Learning Approach. Front Syst Neurosci 2021:15. doi:10.3389/
Traffic Control, 2008:39-43. fnsys.2021.715433.
30. Anastasios P, Nasser M, Hanoch Y et al. Impact of time pressure on dentists diagnostic 47. Johnson WR, Member S, Alderson J et al. Predicting athlete ground reaction forces
performance. Dentistry 2019;01(31). and moments from spatio-temporal driven CNN models. IEEE Trans Biomed Eng
31. Collins MW, Grindel SH, Lovell MR et al. Relationshp between concussion and neuro- 2019;66(3):689-694.
psychological performance in college football players. Am Med Assoc 1999;282(10): 48. Gallardo-Antolin A, Montera JM. External attention LSTM models for cognitive load
964-970. classification from speech, In: Carlos M-V, Matthew P, Senja P, eds. Stat. Lang. speech
32. Pérez-olmos I, Ibáñez-pinilla M. Night Shifts, Sleep Deprivation, and Attention Perfor- Process, 2019. p. 139-151. (Ljubljana, Slovenia).
mance in Medical Students, 2014:56-62. doi:10.5116/ijme.531a.f2c9.

S45