Basic principles and concept design of a real-time clinical decision support system for managing medical emergencies on missions to Mars
aa r X i v : . [ c s . C Y ] S e p Basic principles and concept design of areal-time clinical decision support system forautonomous medical care on missions to Marsbased on adaptive deep learning
Juan M Garcia-Gomez
Universitat Politecnica de Valencia , Biomedical Data Science Lab , Instituto Universitario de Tecnologias de la Informacion y Comunicaciones (ITACA) , Valencia, Spain
October 15, 2020
Abstract
Space agencies and private companies prepare the beginning of thehuman space exploration for the 2030s with missions to put the firsthuman on the Mars surface. The absence of gravity and radiation,along with distance, isolation and hostile environment are expectedto increase medical events with unidentified manifestations along thecrewmembers. The current healthcare strategy based on telemedicineand the possibility to stabilise and transport the injured crewmemberto a terrestrial definitive medical facility is not applicable in explo-ration class missions. Therefore, full autonomous capability to solvemedical situations will guide design of future healthcare systems on-board.This study presents the basic principles and concept design ofa software suite to bring on-board decision support to help dealingwith medical conditions of the crewmembers, with special attentionto emergency care situations and critical monitoring. MEDEA is anautonomous clinical decision support system that provides real-time dvice to tertiary interventions on-board in space exploration mis-sions. Its basic principles are 1) to give real-time support for medicaldecision making, 2) give patient-specific advice for executive problem-solving, 3) take into account available information from life supportand monitoring of crewmembers, 4) be full autonomous from remotefacilities, 5) continuously adapt predictions to physiological distur-bance and changing conditions, 6) optimise medical decision makingin terms of mission fundamental priorities, 7) take into account med-ical supplies and equipment on-board, 8) apply health standards forlevels of care V, 9) apply ethical standards for spaceflights, and 10)apply ethical standards for artificial intelligence.To fulfil these principles, MEDEA is conceptually designed as asoftware suite consists of four interconnected modules. The main ofthem is responsible to give direct advice to the crew by means of adeep learning multitask neural network to predict the characters of themedical event (e.g. life-threatening, delayability, ethical dilemma, du-ration of therapy, and compatible diagnoses), a classifier of the tertiarymedical intervention and an optimiser of medical action plans. Thismodule is continuously evaluate and re-trained with changing phys-iological data from the crew by an adaptive deep learning module,ensuring fairness, interpretability and traceability of decision mak-ing during the full operational time of MEDEA. Finally, MEDEAwould be semantically interoperable with health information systemson-board by a FHIR module.The deployment of MEDEA on-board of future missions to Marswill facilitate the deployment of a comprehensive preventive medi-cal strategy. Moreover, the advance in technology may represent astepstone of the future quantitative medicine on Earth and on theexpansion of humans throughout the solar system. On early 2021 three spacecraft, Perseverance, Tianwen-1 and Hope, willhopefully arrive to Mars to study its atmosphere and surface with the poten-tial result of detecting clear signs of life [1]. The three spaceflights are explo-ration missions based on robotics, but it is not the orbital window for humanexploration of Mars yet. NASA, other space agencies and private companiesprepare the beginning of the planetary-type missions with humans for the2030s transfer windows [2], with the objective of putting the first human on2he Mars surface after achieving the moon surface again through its orbitalGateway station [3]. This new achievement may open a new era of deepspace journeys and planetary design reference missions, including long-termstays on Mars, Lagrange points and exploitation of near-earth asteroids [4],opening unseen opportunities to humanity.Nevertheless, the absence of gravity and radio-protective geomagneticfields from Earth, along with distance, isolation and hostile environmentsexperimented during long duration space travels may increase systematiceffects on the physiology, biology and behaviour of human beings comparedto those measured during short-term experiments performed on InternationalSpace Station (ISS) in the low Earth orbit (LEO) [5]. With that in mind,NASA applied its continuous risk management platform to identify 30 humanhealth and performance risks for space exploration including the risks ofadverse outcomes due to inflight medical conditions [6].Many of the consequences of human risks in long spaceflights are not fullyunderstood yet and technologies for controlling them are still to be invented.NASA established the Human Resource Programme (HRP) at Johnson SpaceCenter (JSP) in 2005 for investigating the highest risks to astronaut healthand performance by quantifying the likelihood of occurrence, severity of con-sequences, and the extent that a risk can be controlled or mitigated bothinflight and post flight for each type of mission. Specifically, the HRP Pathto Risk Reduction for a Planetary Mission estimates that inflight medicalconditions with potential high impact [7] will only be partially controlled intime for a Mars mission in the 2030s.The accepted approach for the healthcare of the crew on a space mis-sion follows the occupational medicine prevention strategy [8]. It focuseson reducing the likelihood and severity of medical events by primary, sec-ondary and tertiary interventions. With primary interventions, likelihoodof risk factors is reduced by careful selection of crewmembers. Therefore,priority is currently given to astronauts with low coronary artery calciumand low Framingham risk scores over those with higher risk levels. Besides,secondary interventions are also applied as countermeasures of the effects ofenvironmental factors in space. In LEO missions, countermeasures to the ef-fect of microgravity and isolations are carried out by routines on treadmills,resistive devices and cycles. Finally, tertiary interventions are triggered totreat illness or injury under emergency situations requiring advance life sup-port, transitional care or ambulatory care. Given the difficulty to provide3ull healthcare support to space, current LEO and lunar missions have followthe paradigm stabilise and transport the injured crewmember to a terrestrialdefinitive medical care facility (DMCF), that is not applicable to explorationclass missions.Deploying a medical strategy for controlling inflight medical conditionsin exploration class missions will need to deal with limitations imposed bydeep space hazards. Three are the major restrictions affecting healthcareof mission crewmembers: physiological disturbance, communication latencyand mission length: • Physiological disturbance includes radiation-induced changes, alterednutritional status, neurovestibular deconditioning, cardiovascular de-conditioning, bone and muscle loss, renal-stone formation, plasma-volume shifts, spaceflight-associated neuro-ocular syndrome (SANS),and altered immunity due to microgravity, radiation and isolation,among others [9]. Multiple biomedical experiments have been per-formed on-board of the ISS . For example, the NASA Twins Studyidentified multiple on-board spaceflight-specific changes, including de-creased body mass, telomere elongation, genome instability, carotidartery distension and increased intima-media thickness, altered ocularstructure, transcriptional and metabolic changes, DNA methylationchanges in immune and oxidative stress-related pathways, and gas-trointestinal microbiota alterations [10]. Nevertheless, there are noevidences of the effects of long distance travels in humans yet. Hence,healthcare strategies, including clinical decision making, will need to befull adaptive to continuous changes, to what clinical symptoms will ap-pear, what new clinical complications will arise and how environmentalconditions will affect the clinical conditions of the crewmembers. • Communication latency between crew and mission control will increasewith distance to Earth. From three to six seconds delay for a roundtrip communication from ISS to Earth, any deep space journey to near-earth asteroids and planetary missions will take several minutes for afull bidirectional communication. This precludes asking immediate in-quires to a telemedicine service similar to what ISS uses to consultmission surgeon at mission control. Hence, autonomous real-time de-cision making based on-board health information system will need to https://nebula.esa.int • The length of the space missions are going to be extended from a currentmaximum of months to several years. Given the high cost of medicalpayload, crewmembers will arrange limited supplies, equipment andresources during mission, so optimisation will be needed on every day-by-day medical decision making. The major condition that missionlength implies is the non-return to permanent health facilities in caseof requiring advanced healthcare. Therefore, it will be mandatory toprovide healthcare autonomy to the mission. Moreover, mission lengthhas a cumulative effect in the physiological disturbance of human be-ings, increasing probability of health failure and clinical severity withtime.As Hamilton et al. stated on [8], the current tertiary intervention cannotlonger draw on a close DMCF at an effective time, so medical design shouldevolve into an autonomous treat to final resolution capability , which repre-sents a significant challenge to space medicine and mission designers. Giventhe major restrictions affecting healthcare of mission crewmembers, in thispaper we propose a real-time clinical decision support system based on adap-tive deep learning as technological solution of tertiary medical interventionsfor Mars-type missions. Our study establishes ten basic pillars and proposesa concept design based on a software suite composed of the next four mod-ules: autonomous real-time clinical decision making , space adaptive learning , semantic interoperability and ethical & legal functional support . This con-cept design gives answer to the system-based strategic vision conceived byWilliams, Hamilton, Doarn and others [11, 8, 12] by means of the last decadeadvances in Biomedical Data Science and Artificial Intelligence (AI). The routine and emergency medical operations in the human flight missionsare currently managed by the flight surgeons from the mission control centre.Crewmembers have an integrated checklist on-board [13] that describes theroutine and emergency medical operations procedures and hardware associ-ated with crew health. Nevertheless, these procedures based on telemedicineand prompt evaluation for tertiary medical interventions should change for5xploration and planetary missions. Then, the design of Mars missions in-cludes re-thinking the life support and medical management for remote hu-man risks management. In this way, American and European space pro-grammes have detected unresolved technologies to manage medical interven-tions on-board of spaceflights in long exploration missions such as Mars-typemissions.The HRP by NASA detected the need of having the capability to providecomputed medical decision support during exploration missions as part ofthe detected gap to identify new capabilities that maximise benefit and/orreduce costs on human system/mission/vehicle resources . By solving thisgap, NASA is willing to develop continued monitoring of biomedical signalsand images, improve medical capability technology for unique spaceflightneeds, provide medical care in a progressively Earth independent fashion anddemonstrate the integration of the new procedures and technologies with theon-board processes.Besides, the European strategy towards human exploration of space de-veloped during the THESEUS FP7 project in 2012 [14] detected insufficienton-board expert systems / decision support systems for medical diagnostics asspecific key issue for the space medicine expert group. In 2013, the ASSISTreport by ESA/ESTEC encountered a major limitation in the data gatheringfor remote monitoring by telemedicine. Specifically, they reported the bot-tleneck of acquiring and transmitting too much data not always immediatelyand easily available and the difficulty to interpret them, so 75% of validatorsconsidered useful to incorporate additional functionalities such as the au-tomatic identification of indicators with high sensitivity and the automaticidentification of values for the break-even. Then, CSEM, Airbus and MEDESevaluated in 2016 the benefit of data mining for an autonomous medical mon-itoring/diagnostic system (AMIGO) in long-term spaceflight and non-spacerelated applications. In their studies, they estimated higher medical risk dueto expected human tasks with EVA during asteroid missions for sample ex-traction and planetary surface exploration for the Mars-type mission. Theyperformed two proof of concept on subjects with de novo cardiac arrhyth-mia and sleep apnoea, respectively. For their computational experiments,they applied linear classifiers, quadratic classifiers, Gaussian mixture mod-els, hidden Markov models, artificial neural networks, k-nearest neighbours, https://humanresearchroadmap.nasa.gov/gaps/gap.aspx?i=642 https://humanresearchroadmap.nasa.gov/gaps/gap.aspx?i=716 .McGregor in 2013 [15] proposed a real-time platform call Artemis for on-line health analytics during spaceflight by monitoring the astronauts’ physi-ological signals on-board as well as sending the signals to mission control formedical support at each stage when communication is available. Prysyazh-nyuk et al. in 2017 [16] tested by analogs if Artemis was able to supportthe implementation of the IBMP classification model of health states gravi-ties in four functional states (physiological normal, prenosological, premorbidand pathological) based on the discriminative Heart Rate Variability (HRV).Moreover, in 2017 McGregor and the IBMP [17] simulated the integrationof the Artemis platform and the Cosmocard device acquiring electrocardio-grams from the Russian cosmonauts on ISS. Their results shown limitationsfor a real-time performance due to deferred transmission from the Cosmocarddevice.As well as for medical diagnosis, biomedical signals have been acquired on-board of the ISS to study the effects of microgravity and isolation in humanphysiology. EveryWear by ESA monitored the ECG, tonometry, and temper-ature of astronaut Thomas Pesquet from Nov 2016 to May 2017 by wearablesensors connected to an iPad. The Neurolab Spacelab mission on 1996 stud-ied the effects of weightlessness on the brain and nervous system [18]. Petit etal. in 2019 [19] studied the relationship of sleep pressure EEG markers duringwakefulness in astronauts throughout 6-month space mission. Gemignani etal. in [20] performed what we may consider the first adaptive data-drivendecision support system for space medicine. They defined dynamical thresh-olds on high-density EEG to detect sleep spindles with a false positive rateof 5% to deal with the variability of spindle characteristics as a function ofsleep depth and subject. Besides, the Airway Monitoring ISS investigationstudied the inflammation and reduced pressure on pulmonary nitric oxideturn-over due to microgravity and other ESA experiments such as DNAm-AGE, ICELAND and IMMUNO have also studied the effect of spaceflightsat genomic, microbiome and immunological levels.Up to now, ultrasonography is the main diagnostic imaging technique https://physionet.org Institute of Biomedical Problems (IBMP) of the Russian Academy of Sciences. and the ECHO experimenttried to solve the operator-dependence by remote controlled ultrasounds op-erations. Although, this solution is feasible for LEO missions, long latencyin exploration and planetary missions will require autonomous operation on-board. Aravindhan et al. proposed in [21] a Raspberry Pi solution for onlinehealth diagnosis to operate during space tourism and future Mars colonisa-tion by illustrating its potential application to fundoscopies when suspicionof visual impairment due to intracranial pressure.HRP have identified 32 physiological, medical, and behavioural risks as-sociated with long-duration spaceflights [22]. Linked to that, Davis et al. presented in [23] a risk management system based on the acceptable levels ofrisk for each mission type with the goal of guiding research efforts and mis-sion planning through the probability of medical adverse events, uncertaintyof outcomes, impacts, costs and benefits of mitigation actions along withrelated current and future work. More recently, Mindock et al. in [24] de-fined a connected map of contributing factors and the medical risks, whereasRomero and Francisco in [6] identified 100 hundred probable health problemsthat may affect mission success and classified the medical risks in five haz-ards of spaceflight: altered gravity, radiation, distance, isolation and hostileenvironment. Taking into account that medical care will be limited by mass,volume, and power constraints and that life support will represent the 40% ofwet mass in exploration-class spaceflight [25], one of the firsts uses of risk as-sessments presented above was the list of medical resources on-board. Withthat objective in mind, Antonsen et al. [26] designed a tradespace analysistool to score resources, tools, and skillsets required for exploration missions.As we have seen, current research is focused on identifying adaptationeffects to spaceflights. Wherease biomedical signal monitoring is almost rou-tine in astronauts, current solutions are solved by telemedicine but few ad-vances have been made on autonomous decision making. Besides, medicalrisks, diseases, factors, mitigations and consequences identified in the studiesfor healthcare management constitutes the key knowledge to plan researchdirections for designing human spaceflight missions to Mars and asteroids.Hence, our proposal of the clinical decision support system focuses on thespecific requirements that crewmembers will have to deploy autonomous ter-tiary medical interventions and it attends the limitations and state of the artpresented above. http://youbenefit.spaceflight.esa.int/ultrasonography-without-borders/ Basic principles and concept design of MEDEA
Primary and secondary can be operated in advance and from mission controlas it is currently designed for ISS missions. Nevertheless, long distance inexploration and Mars missions requires autonomy to decide how to react toa medical situation. Here we study how to bring on-board medical decisionsupport to the physician and crewmembers during medical conditions, withspecial focus on emergency care situations and critical monitoring. Giventhe limitations imposed by deep space hazards, a clinical decision suppportsystem for exploration and Mars missions may fulfil the next ten basic prin-ciples: • Give real-time support for medical decision making • Give patient-specific and quantitative advice • Take into account available information from life support and monitor-ing of crewmembers • Be fully autonomous from remote facilities • Continuously adapt predictions to physiological disturbance and chang-ing conditions • Optimise medical decision making in terms of mission fundamentalpriorities • Take into account medical supplies and equipment on-board • Apply health standards for levels of care V • Apply ethical standards for spaceflights • Apply ethical standards for artificial intelligenceTherefore, MEDEA is designed as a comprehensive computational suite todeploy personalised Clinical Decision Support for exploration and planetarymissions. It should operate continuously to react to emergency and unseenmedical situations giving support for clinical decision making by quantitativeadapted predictions to individual crewmembers profiles. The suite is com-posed of four software modules physically distributed on-board or on Earth9acilities to provide four main functionalities of the system: autonomous de-cision making (on-board), space adaptive deep learning (on Earth), semanticinteroperability (on-board) and ethical & legal functional support (on-boardand on Earth). Figure 1 shows the details of the four modules that im-plements the principles enumerated above to support tertiary medical in-terventions and how the modules are interconnected among them and theinformation systems on-board.The module
Autonomous real-time CDSS for tertiary medical care is theone that directly interacts with the crewmembers for supporting medical de-cision making by means of AI-models during medical emergencies and criticalmonitoring under the paradigm treat to final resolution . To do that, it re-ceives the health status of the crewmembers from the on-board health caresystem and the medical diagnosis systems through the semantic interoper-ability module. The AI-models are continuously adapted to the changingspace conditions by the
Space adaptive deep learning module. Both CDSSand deep learning performances are continuously verified by the
Ethical andlegal functional support module. In the next four sections, we provide thetechnological solution of each module of the MEDEA suite.
Romero et al. in [6] compiled the most common hundred medical condi-tions in space derived from the ISS Medical Checklist, scientific research andoccupational health statistics. The stratification of this list serves to HRPfor planning mitigation actions for the thirty human system risks. Medi-cal conditions in spaceflights may be occupationally-induced conditions oridiopathic illnesses. Crew activities confined in a spacecraft extravehicularactivities (EVA) and surface explorations may increase the probability ofinjuries and trauma, that may derive on emergency medical situations. Al-though presentations and frequencies of medical conditions during prolongedstays may change for Mars missions, the closest reference given by medicalreports from ISS [27] revealed that 46% of crewmembers expressed an eventdeemed notable, being skin rashes and hypersensitivities (40%, 1.12/flightyear) along with upper respiratory symptoms (0.97/flight year), the mostreported events. Moreover, artificial life support added to space-specific con-10 o return to a definitive medical care facilityTx/Rx latency of several minutesFrom 400 days to 3 years mission Diverse crewmembersDiverse clinical profiles Diverse sex & genderUp to 1 physician in the mission
Medical prevention strategyPrimary preventionSecondary preventionTertiary prevention
Level of care VBasic life supportFirst-aid capabilityClinical diagnosisAmbulatory careNo inmediate return to EarthNo RT telemedicineAdvance cardiac life supportAdvanced trauma life supportAdvanced Life SupportBasic surgical carePalliative careCrew selection
LEO: No interventions“Stabilise and transport” policy
Treatment resources by diagnosis
Autonomous real-time CDSS for tertiary medical care
On-board health care system Medical diagnosis systems
MARS: Evolve into an autonomous“Treat to final resolution” policy
Countermeasures Idiopathic illnessesSubclinical diseasesOccupationally-induced medical eventsDisturbed physiology
Mission fundamental priorities & Level of care VCompatible diagnoses Optimisation of action plansStandardised taxonomy of risk factorsMedical operations - procedure libraryOn-board pharmaceutical supplies
MicrogravityRadiation Isolation
Advance Life Support careEmergency careTransitional careAmbulatory carePaliative careClassification of tertiary care interventions
Space adaptive learning
Dataset shift and variability characterisationContinual deep learning
On-board facilities
On-board device standard protocols
Semantic interoperability
Medical standards formats
Mission Fundamental Priorities1- vehicle survival2- health & safety of crew3- mission success4- payload success
Data protection
Ethical & legal functional support
On-board & on Earth facilitiesOn-board facilitiesOn Earth facilities
Fairness Interpretability TraceabilityMulti-task prediction modelingLife-threatening situationDelayabilityEthical dilemmaDuration of therapy
Figure 1: Concept design of MEDEA composed of four main subsystems for autonomous decision making , space adaptive deep learning , semantic inter-operability and ethical & legal functional support .11itions and isolation increase the onset of conditions such as space adaptationsyndrome, headaches, gastrointestinal distresses, degradation of the immunesystem, infectious processes, sleeplessness and depression, among others.Added to that, idiopathic illnesses during 3-4 years of a Mars-type missionare willing to appear more than in a LEO mission due to: prolonged stays ofthe same crewmembers, variability of tasks during exploration missions [28],completely absence of gravity, exposure to radiation, and increase in thenumber of astronauts from private and public space programmes and highervariability of medical profiling [9].Current design of exploration-type and Mars-type missions plans to bookone seat of the spacecraft for a physician acting as Chief Medical Officer(CMO) in every spaceflight crew [29, 8]. Although this may enhance missionsafety, long latency to communicate from mission control involves isolation onmedical decision making to solve emergency situations. A real-time clinicaldecision support system may give processed knowledge related to patients’conditions to CMO during on-board emergency situations. Moreover, theresult of the system may assist crewmembers in the critical case CMO isunavailable.Furthermore, current design of clinical decision support systems shouldfollow the next four caveats by Shortliffe and Sepulveda [30]: • Black boxes are unacceptable • Complexity and lack of usability thwart use • Delivery of knowledge and information must be respectful • Scientific foundation must be strongThis module performs three sequential functionalities to give a completesupport for dealing with medical situations: multitask prediction modelling,classification of type of tertiary medical intervention and optimisation of theaction plan. Each functionality and its design are described below.
When a medical emergency arises, a prompt prediction should be carried outto decide should plan to treat in the draft list of the medical conditions for Mars missions presented byNusbaum et al. in 2019 [31]. Additional 22 conditions were plan to treatwith conditions and 34 of them should not plan to treat . Given that largeamount of prediction values and varying conditions in deep space, the bestconfiguration for the Task
We can distinguish among five different types of tertiary medical interven-tions on spaceflight: advance life support care, transitional care, ambulatorycare, palliative care and emergency care. The classification of the medi-cal intervention for any medical situation can be directly mapped from thefive predictions carried out by the multitask prediction modelling . Giventhat space medicine shares many attributes with extreme conditions and en-vironments where emergency medicine operates, we propose following theapproach applied by Ferri et al. [33] for classifying emergency medical callincidents. Hence, the mapping can be defined by a panel of over 20 physiciansand mission designers using a Delphi methodology [40].14 .3 optimisation of action plans
Once a compatible diagnosis and the type of tertiary intervention have beenassigned, it is time to apply a set of medical action to restore the healthand performance of the crewmembers. They should be compatible with theLevel of Care V, that entails basic and advanced life support, first aid, clini-cal diagnosis, imaging, ambulatory care, telemedicine, sustainable advancedcardiac life support, advanced trauma life support, basic surgical care andpalliative care without immediate return to earth capability [8]. Therefore,the medical actions available in exploration and planetary missions wouldbe an upgraded version of the routine and emergency medical operationsincluded in the ISS Integrated Medical Group Medical Checklist [13].Then, a complete medical action plan should be tailored to the medicalsituation trying to optimise the mission fundamental priorities (i.e. 1: vehiclesurvival, 2: health & safety of the crew, 3: mission success, and 4: payloadsuccess) under the constraints imposed by the restricted medical equipmentfor diagnoses and a limited amount of medical supplies for treatment. Forthis functionality, we intend to optimise the sequence of actions using a re-inforcement learning approach [41] continuously checked by a real-time alertwhen the action plan diverts from the mission fundamental priorities.Emergency events are the situations where an on-board clinical decisionsupport may be more needed by the crew. Nevertheless, it cannot be isolatedfrom the full tertiary interventions deployed for the medical prevention strat-egy. Of most interest is to evaluate if multitask prediction models developedfor medical emergencies on Earth [33] can be transferred to on-board decisionmaking with high rates of accuracy. Doing that, the development of clinicaldecision support systems to help on-board medical interventions may takeadvantage of massive biomedical data analysis performed on Earth.
Deep learning [42] is the novel technology with more success for mimickinghuman decision making [43] from complex types of data, such as involvinghigh dimensional and multimodal data [44], sequences [45] and unstructureddata [46]. Given the modular architecture proposed in 4.1, we suggest per-forming an independent learning process of each task-independent subnet-work by the Adam stochastic optimisation algorithm [47] with a weigh decay15erm to promote regularisation [48] followed by their ensemble as looselycoupled models [49]. Moreover, to evaluate the model performance and tunehyperparameters without biases maximising available re-use data we followthe robust methodology proposed by Kohavi [50].Nevertheless, the main potential limitation to design prediction modelsfor medical decision-making during exploration missions is the continuousmedical dataset shift [51] of the on-board cases produced by the physiologicaldisturbance. Dataset shift was first described in [52] and defined by Moreno-Torres et al. [51] as the situation in which the training and test data followjoint distributions that are different. Dataset shift occurs when the dataexperience a phenomenon that leads to a change in the distribution of asingle feature, a combination of features or the output boundaries.In medical prediction problems, where the output (e.g. the disease)causally determines the values of the features (e.g. symptoms), there aretwo types of dataset shift that may appear independently or at the sametime. First, prior probability shift refers to changes in the distribution of theoutput variable. In space medicine it is observed how the prevalence of somemedical conditions increase due to the specific environment and activities ofthe crewmembers. Then, the frequencies of arrhythmia, headache, dermati-tis, respiratory infection, and renal stone formation, among other medicalevents, are increased in space with respect to on Earth. We can expect thatprobability shift of medical conditions in space will continuously change giventhe cumulative effect of space influence in the physiological disturbance andthe non-stationary environment of exploration missions. Second, conceptdrift (a.k.a. concept shift) happens when the representation of the inputsconditioned to the outputs of a predictive model changes in test cases withregard to training cases. In medical applications, this may happen whenthe observation of symptoms manifesting diseases changes during operationswith respect to the data distributions learned during the design of the pre-diction model. Given the effect of microgravity, radiation and isolation inhuman bodies, we can expect a major concept drift in the biomedical datagenerated in long-term space missions.Moreover, designing an effective data-driven decision support system forspace exploration missions will require a continual update of prediction mod-els to the lastest registered data. That can be solved by several alternativeapproaches: a) perform retraining using the complete historical dataset, b)perform continual learning of models’ parameters including periodically the16ew test cases [53], c) select the most robust models adapted to impreciseenvironments [54]. Given that currently there are no registries of biomedicaldata of humans from space exploration missions, our choice is the continualdeep learning approach that avoids access to historical multisource data, al-lows using data from ISS and Earth at present time and produces predictionmodels continuously adapted by cases which features are conditioned by thedisturbance physiology effects of space.Moreover, a careful monitor of biases affecting data representation shouldbe carried out to ensure quality and performance of updated models [55]. Themethodology developed in [56] based on non-parametric statistical manifoldsmay be useful to calculate the dynamics of temporal variability of biomedi-cal data from crewmembers, including continuous temporal trends, seasonalbehaviours and abrupt changes produced by dataset shift effects.The most challenging step in the design of the MEDEA system is indeedthe space adaptive learning. Added to the lack of data from deep space,continual learning is nowadays an open topic still to be solved for terres-trial scenarios. Nevertheless, space exploration missions involve a dynamismdifficult to compare with other situations. Hence obtaining a successful con-tinual learning of on-board data-driven clinical decision making is the mostsalience hypothesis of the MEDEA concept design. With this approach weexpect that the evidence of tomorrow will help us further develop and buildsmart medical systems to address those yet undiscovered challenges of long-duration, long-distance spaceflight [57] . This module is in charge of exchanging unambiguous information with thecomputer systems for the health maintenance of the crewmembers. Spaceagencies have addressed exchange of data in multiple vendors environmentsby definition of interoperability protocols, such as STEP-SPE. Whereas spaceagencies were focused on the exchange of information among space environ-ment analysis tools [58], medical informatics has focused the attention ondeploying semantic interoperability in healthcare organisations, that goessome steps further in the exchange of information among heterogeneous sys-tems. Specifically, semantic interoperability is defined as the transmission ofdata along with the required knowledge to understand it, by sharing clinicalconcepts described in a reference model using a binding medical terminology17hared vocabulary [59]. This may allow sending information to buses of datawithout assuming that every receptor needs to know in advance its semantic.Then, we propose to adopt the Fast Healthcare Interoperability Resources(FHIR) to exchange medical data on-board between our clinical decision sup-port system and the on-board healthcare systems and medical diagnosis sys-tems. FHIR was designed for exchanging electronic health records (EHR) bythe Health Level Seven International (HL7) organisation and it is supportedby the American Medical Informatics Association. The idea is to organize en-tities, such as patients, observations, measurements and diagnoses, as FHIRresources specified by profiles (clinical concepts) with U.S. Core Data forInteroperability (USCDI) elements written in SNOMED Clinical Terms. Inthat way, medical records of the astronauts will be consistent with the withthe international standards followed by the providers of medical informationtechnologies.Although a FHIR-based system would provide a full semantic interoper-ability with health information systems, given that current on-board com-puter systems do not follow interoperable standards yet or they are basedon industrial standards from aeronautics, the semantic operability modulemay also include connectors to the specific on-board systems with adaptersto their data structures.
Several ethical and legal concerns must be addressed correctly for a properday-by-day operation of MEDEA as on-board clinical decision support sys-tem. First, the massive use of biomedical data of the crewmembers in thecontext of a unique environment of deep space requires the correct man-agement of regulation for data protection and privacy. Emphasis on theconfidentiality of astronaut clinical data has resulted in missed opportunitiesto understand human physiological adaptations to space [60]. Laws adoptedto take into account the digital era, such as the General Data ProtectionRegulation (GDPR) [61] implemented by the European Union on 2018, givecontrol of data to the individual. Our premise is that the new regulations fordata protection may be a mechanism to avoid the loss of valuable biomedicalinformation from deep space environments.Another relevant issue is the application of ethical criteria when givingreal-time decision support for solving medical events. To this, mission fun-18amental priorities and available resources for Level of Care V may guide theoptimisation of the medical action plan.More generally, given that we are proposing a data-driven decision sup-port system, independently of the medical certificates requested to the crewmedical officers [9], our proposal may also adopt solutions for the next addi-tional issues: • Fairness to avoid reproducing any pattern of discrimination due to prej-udices or bias [62] • Interpretability to enable a correct explanation of medical predictionsand decisions by human experts [63] • Traceability to achieve a comprehensive examination of responsibilitiesof any medical decision making at any time, ensuring the currency ofthe knowledge base and that it is safe to use [30]The inclusion of an ethical and legal framework as a module of MEDEAmay ensure the practical implementation of solutions for the five issues de-scribed above. A careful validation of its performance through computa-tional simulations and analogues may guarantee the correct functionality ofthe MEDEA system with respect to the human well-being.
Our approach may give real-time decision support for continuously changingmedical situations in a way that can be managed in long-duration spaceflight,with special attention to emergency medicine and critical monitoring. Stew-art et al. [22] identified traumatic injury as one of the most relevant emer-gency situations in space exploration given the expected incidence and conse-quences to the mission fundamental priorities. Additionally, it is largely un-known the cardiovascular and immunological effects of long-duration space-flight on wider spectra of medical profiles of astronauts. Kuypers et al. in [29] include two lists of health concerns due to specific space conditionsand medical emergencies, respectively. The first list identifies medical as-pects such as cancer, cataracts, immunologic changes, decreased red bloodcell mass, bone and mineral loss, muscle atrophy/loss of strength, vestibu-lar/sensorimotor changes, cardiovascular changes, hyperopic vision shifts,mental health problems, bacterial growth, water and air contamination or19egradation, and other deficiencies. The second list includes wounds, burns,contusions, sprains, fractures, cardiac dysrhythmias, orthostatic intolerance,pneumonitis, persistent latent viral reactivation, anaphylactic reactions, der-matologic cellulitis, dermatitis, space motion sickness, gastroenteritis, consti-pation, renal stones, urinary tract infections, acute urinary, retention, crownfracture, dental infections, abscess corneal abrasion, corneal infection, foreignbodies, depression, anxiety, exposure to toxins, acute radiation illness anddecompression sickness. Besides, Stewart et al. in [22] and Komoroski andFleming in [64] focused on the medical emergency procedures of the criticallyill and injured on extreme conditions and environments. Nevertheless, weare not aware of any scientific paper or experiment report about informationtechnology specialised in emergency medicine in space, being this study thefirst that propose basic principles and a concept design of a clinical decisionsupport systems to tackle autonomous tertiary medical interventions.Estimated in 2010 and 2019 the cost of their first mission to Mars at 6Billion USD and 10 Billion USD, respectively. Added to the loss of a life,a health problem in the crew may put in danger the rest of the crewmem-bers, jeopardise the mission and loss of the vehicle and payload.
History hasshown that during the exploration of frontiers on Earth, human physiologicmaladaptation, illness, and injury have accounted for more failures of expe-ditions than any single technical or environmental factor [65] . Therefore, itis critical to rely on robust fault-tolerant solutions to deploy tertiary medicalinterventions in the most autonomous fashion during space exploration mis-sions, where evacuation and synchronous communication to mission controlis not a reliable option.Clinical decision support systems may provide real-time advice tailoredto the health problem and optimised in terms of the mission fundamentalpriorities and in compliance of the highest ethical and legal conditions. Thesupport for a wide set of medical conditions must be deployed to deal withmedical emergency events and nominal health issues as well, in light of theexpected increase of astronauts profiles and civilian spaceflight [9]. More-over, adaptive learning must guarantee predictive models updated to thecumulative and variable disturbance of space effects in human physiology.The autonomous clinical decision support system developed in MEDEAmay be directly transfer to medical applications on Earth. Specific scenariosmay have specific requirements similar to space exploration missions. Isola-tion and extreme conditions may appear on deep sea exploration, Artic and20ntarctic missions, and isolated communities on desserts and forests. Theglobal market of clinical decision support systems for general and specialisedmedicine for the year 2028 has been estimated at 3.5 Billion USD. The de-velopment of quantitative medicine will go hand-in-hand with the design ofthis technology, so robustness and adaptation required for operating in spacewill speed up this process.Sooner or later humans will expand along solar systems so spaceflightamong planets, asteroids and space stations will become frequent. Developingadaptive clinical decision support systems will be a keystone for deliveringmedical care on-board, opening the career of healthcare in space. It is notpossible to estimate the economic impact of healthcare after humans expansebut we can grossly understand its dimension by the size of the current globalhealthcare market calculated at 12 Trillion USD.Getting MEDEA on-board of a spacecraft is not going to be easy. Initially,the four modules will have to be designed at the same time as the missionsto Mars are designed for the launch windows by 2030s. Therefore, real-time integration with health information systems on-board may representa challenge on its own. Moreover, the initial versions of predictive modelswill not be able to be trained from any real data acquired during previousplanetary missions, so we will have to use medical data from Earth andISS relying on the continual adaptive learning to get relevant knowledgefrom the physiological conditions that astronauts will experiment on-board.Stakeholders and public opinion will only trust MEDEA if we guaranteeits ethical and legal compliance shown by a clear fairness, interpretabilityand traceability capabilities. The principles and concept design describedin this study may serve as the basis to implement a complete and qualifiedclinical decision support system to operate in space exploration missionsbefore 2030s.
In less than two decades space missions for human exploration of Mars will bedesigned and launched. A key element for their success will be the capabilityto provide autonomous healthcare adapted to the dynamic space conditions.on-board healthcare may be fully redesigned taking into account that low-latency telemedicine and prompt evacuation to Earth will not be feasible21n that new type of human missions. Therefore, autonomy for real-timedecision making will be mandatory to solve medical emergency situationsand necessary to monitor medical status for space induced health conditions.In this study we have introduced the basic principles and concept designof MEDEA, a clinical decision support system to provide real-time advice totertiary interventions on-board of space exploration missions. The presenteddesign applies the current Biomedical Data Science and AI technology tofulfil the fundamental priorities and the level V of healthcare in spaceflights.The design consists of four dependent modules, being the main one the re-sponsible for giving direct advice to the crew by means of a deep learningmultitask network to predict the character of the medical event, a classifierof the tertiary medical intervention and an optimiser of the medical actionplan. The adaptation of prediction model to the changing physiology onspace will be solved by a continual deep learning module and both moduleswill be integrated with health information systems on-board by means of asemantic interoperability FHIR module. Fairness, interpretability and trace-ability provided by the ethical and legal module is expected to ensure bestpractices and trust during the full operational time of MEDEA.The clinical decision support system implementing the MEDEA conceptdesign is expected to give autonomous decision making for the next humanmissions to Mars. Besides, it will represent a stepstone for the future ofquantitative medicine on Earth and a potential healthcare device for theexpansion of humankind throughout the solar system.
Acknowledgements
The author thanks Pablo Ferri, Antonio Felix, Pere Blay, Alberto Conejero,Alberto Gonzlez, Elies Fuster and Carlos Saez for their comments to theconcept design described in this study.
Funding sources
This research did not receive any specific grant from funding agencies in thepublic, commercial, or not-for-profit sectors.22 ompeting interests statement
The author has no competing interests to declare.
References [1] A. Witze, S. Mallapaty, E. Gibney, Countdown to Mars: threedaring missions take aim at the red planet, Nature 583 (7815)(2020) 184–188, number: 7815 Publisher: Nature Publishing Group. doi:10.1038/d41586-020-01861-0 .[2] P. D. Wooster, R. D. Braun, J. Ahn, Z. R. Putnam, Mission designoptions for human Mars missions, International Journal of Mars Scienceand Exploration 3 (2007) 12–28. doi:10.1555/mars.2007.0002 .[3] M. Smith, D. Craig, N. Herrmann, E. Mahoney, J. Krezel,N. McIntyre, K. Goodliff, The Artemis Program: An Overviewof NASA’s Activities to Return Humans to the Moon, in: 2020IEEE Aerospace Conference, 2020, pp. 1–10, iSSN: 1095-323X. doi:10.1109/AERO47225.2020.9172323 .[4] a. T. United States.Congress.Senate.Committee on Commerce, Science,The future of U.S. human spaceflight: hearing before the Committeeon Commerce, Science, and Transportation, United States Senate, OneHundred Eleventh Congress, second session, May 12, 2010, Tech. Rep.Y 4.C 73/7:S.HRG.111-984, WASHINGTON (2011).[5] D. Grimm, Guest Edited Collection: Gravitational biology and spacemedicine, Scientific Reports 9 (1) (2019) 14399, number: 1 Publisher:Nature Publishing Group. doi:10.1038/s41598-019-51231-8 .[6] E. Romero, D. Francisco, The NASA human system risk mitigationprocess for space exploration, Acta Astronautica 175 (2020) 606 – 615. doi:https://doi.org/10.1016/j.actaastro.2020.04.046 .[7] NASA’s Efforts to Manage Health and Human Performance Risks forSpace Exploration, Tech. Rep. IG-16-003, National Aeronautics andSpace Administration. Office of Inspector General (2015).238] D. Hamilton, K. Smart, S. Melton, J. D. Polk, K. Johnson-Throop, Autonomous medical care for exploration class space mis-sions, The Journal of Trauma 64 (4 Suppl) (2008) S354–363. doi:10.1097/TA.0b013e31816c005d .[9] J. Stepanek, R. S. Blue, S. Parazynski, Space Medicine in the Era ofCivilian Spaceflight, New England Journal of Medicine 380 (11) (2019)1053–1060. doi:10.1056/NEJMra1609012 .[10] F. E. Garrett-Bakelman, M. Darshi, S. J. Green, R. C. Gur, L. Lin,B. R. Macias, M. J. McKenna, C. Meydan, T. Mishra, J. Nasrini,B. D. Piening, L. F. Rizzardi, K. Sharma, J. H. Siamwala, L. Tay-lor, M. H. Vitaterna, M. Afkarian, E. Afshinnekoo, S. Ahadi, A. Am-bati, M. Arya, D. Bezdan, C. M. Callahan, S. Chen, A. M. K. Choi,G. E. Chlipala, K. Contrepois, M. Covington, B. E. Crucian, I. De Vivo,D. F. Dinges, D. J. Ebert, J. I. Feinberg, J. A. Gandara, K. A. George,J. Goutsias, G. S. Grills, A. R. Hargens, M. Heer, R. P. Hillary, A. N.Hoofnagle, V. Y. H. Hook, G. Jenkinson, P. Jiang, A. Keshavarzian,S. S. Laurie, B. Lee-McMullen, S. B. Lumpkins, M. MacKay, M. G.Maienschein-Cline, A. M. Melnick, T. M. Moore, K. Nakahira, H. H.Patel, R. Pietrzyk, V. Rao, R. Saito, D. N. Salins, J. M. Schilling, D. D.Sears, C. K. Sheridan, M. B. Stenger, R. Tryggvadottir, A. E. Urban,T. Vaisar, B. Van Espen, J. Zhang, M. G. Ziegler, S. R. Zwart, J. B.Charles, C. E. Kundrot, G. B. I. Scott, S. M. Bailey, M. Basner, A. P.Feinberg, S. M. C. Lee, C. E. Mason, E. Mignot, B. K. Rana, S. M.Smith, M. P. Snyder, F. W. Turek, The NASA Twins Study: A mul-tidimensional analysis of a year-long human spaceflight, Science (NewYork, N.Y.) 364 (6436) (2019). doi:10.1126/science.aau8650 .[11] D. R. Williams, R. L. Bashshur, S. L. Pool, C. R. Doarn, R. C. Merrell,J. S. Logan, A strategic vision for telemedicine and medical informaticsin space flight, Telemedicine Journal and E-Health: The Official Jour-nal of the American Telemedicine Association 6 (4) (2000) 441–448. doi:10.1089/15305620050503924 .[12] C. Doarn, D. Williams, A. Nicogossian, R. Williams, Medical Care fora Mars Transit Mission and Extended Stay on the Martian Surface,Journal of Cosmology 12 (2010) 3758–3767.2413] Mission Operations Directorate, Operations Division, InternationalSpace Station Integrated Medical Group (IMG) Medical Checklist(2001).[14] Towards Human Exploration of Space: a EUropean Strategy, Tech. Rep.THESEUS Grant agreement ID: 242482, FONDATION EUROPEENNEDE LA SCIENCE (2012).[15] C. McGregor, A platform for real-time online health analytics duringspaceflight, in: 2013 IEEE Aerospace Conference, 2013, pp. 1–8, iSSN:1095-323X. doi:10.1109/AERO.2013.6497382 .[16] A. Prysyazhnyuk, R. Baevsky, A. Berseneva, A. Chernikova, E. Lu-chitskaya, V. Rusanov, C. McGregor, Big data analytics for en-hanced clinical decision support systems during spaceflight, in:2017 IEEE Life Sciences Conference (LSC), 2017, pp. 296–299. doi:10.1109/LSC.2017.8268201 .[17] C. McGregor, O. Orlov, R. Baevsky, A. Chernikova, V. Rusanov, AMethod for the Integration of Real-time Probabilistic Approaches forAstronaut Wellness in Human in the Loop Related Missions and Sit-uations with Big Data Analytics, in: 19th AIAA Non-DeterministicApproaches Conference, AIAA SciTech Forum, American Institute ofAeronautics and Astronautics, 2017. doi:10.2514/6.2017-1096 .[18] R. Bondar, The Neurolab Spacelab Mission: neuroscience research inspace: results from the STS-90, Neurolab Space lab Mission, Archivesof Neurology (2005).[19] G. Petit, A. M. Cebolla, S. Fattinger, M. Petieau, L. Summerer,G. Cheron, R. Huber, Local sleep-like events during wakefulness andtheir relationship to decreased alertness in astronauts on ISS, NPJ mi-crogravity 5 (2019) 10. doi:10.1038/s41526-019-0069-0 .[20] J. Gemignani, J. Agrimi, E. Cheli, A. Gemignani, M. Laurino, P. Al-legrini, A. Landi, D. Menicucci, Pattern Recognition With Adaptive-Thresholds For Sleep Spindle In High Density EEG Signals, Confer-ence proceedings : ... Annual International Conference of the IEEEEngineering in Medicine and Biology Society. IEEE Engineering in25edicine and Biology Society. Annual Conference 2015 (2015) 594–597. doi:10.1109/EMBC.2015.7318432 .[21] A. Aravindhan, G. Laxmikanth, P. Shanmuga S., S. Kamalraj, Medi-cal Diagnosis During Space Tourism And Future Mars Colonization, in:2020 6th International Conference on Advanced Computing and Com-munication Systems (ICACCS), 2020, pp. 557–559, iSSN: 2575-7288. doi:10.1109/ICACCS48705.2020.9074349 .[22] L. H. Stewart, D. Trunkey, G. S. Rebagliati, Emergency medicinein space, The Journal of Emergency Medicine 32 (1) (2007) 45–54. doi:10.1016/j.jemermed.2006.05.031 .[23] J. R. Davis, J. A. Fogarty, E. E. Richard, Human healthand performance risk management-an approach for explo-ration missions, Acta Astronautica 63 (7) (2008) 988 – 995. doi:https://doi.org/10.1016/j.actaastro.2008.02.004 .[24] J. Mindock, S. Lumpkins, W. Anton, M. Havenhill, M. Shel-hamer, M. Canga, Integrating spaceflight human systemrisk research, Acta Astronautica 139 (2017) 306 – 312. doi:https://doi.org/10.1016/j.actaastro.2017.07.017 .[25] J. Gemignani, T. Gheysens, L. Summerer, Beyond astronaut’s capabil-ities: The current state of the art, Conference proceedings: ... AnnualInternational Conference of the IEEE Engineering in Medicine and Biol-ogy Society. IEEE Engineering in Medicine and Biology Society. AnnualConference 2015 (2015) 3615–3618. doi:10.1109/EMBC.2015.7319175 .[26] E. L. Antonsen, R. A. Mulcahy, D. Rubin, R. S. Blue, M. A.Canga, R. Shah, Prototype Development of a Tradespace Analy-sis Tool for Spaceflight Medical Resources., Aerospace medicine andhuman performance 89 (2) (2018) 108–114, place: United States. doi:10.3357/AMHP.4959.2018 .[27] B. Crucian, A. Babiak-Vazquez, S. Johnston, D. L. Pierson, C. M. Ott,C. Sams, Incidence of clinical symptoms during long-duration orbitalspaceflight, International Journal of General Medicine 9 (2016) 383–391. doi:10.2147/IJGM.S114188 . 2628] P. Ahlf, E. Cantwell, L. Ostrach, A. Pline, Mars scientific investigationsas a precursor for human exploration, Acta Astronautica 47 (2) (2000)535–545. doi:10.1016/S0094-5765(00)00092-8 .[29] M. I. Kuypers, Emergency and Wilderness Medicine Train-ing for Physician Astronauts on Exploration Class Missions,Wilderness & Environmental Medicine 24 (4) (2013) 445–449. doi:10.1016/j.wem.2013.03.024 .[30] E. H. Shortliffe, M. J. Seplveda, Clinical Decision Support in theEra of Artificial Intelligence, JAMA 320 (21) (2018) 2199–2200. doi:10.1001/jama.2018.17163 doi:https://doi.org/10.1155/2009/421425 .[33] P. Ferri, C. Sez, A. F.-D. Castro, J. Juan-Albarracn, V. Blanes-Selva, P. Snchez-Cuesta, J. M. Garca-Gmez, Deep multitask ensem-ble classification of emergency medical call incidents combining mul-timodal data improves emergency medical dispatch, medRxiv (2020)2020.06.26.20123216Publisher: Cold Spring Harbor Laboratory Press. doi:10.1101/2020.06.26.20123216 .[34] S. Ruder, An Overview of Multi-Task Learning in Deep Neural Net-works, arXiv:1706.05098 [cs, stat]ArXiv: 1706.05098 (Jun. 2017).[35] R. Caruana, Multitask Learning, Machine Learning 28 (1) (1997) 41–75. doi:10.1023/A:1007379606734 .[36] A. L. Maas, A. L. Maas, A. Y. Hannun, A. Y. Ng, Rectifier nonlinear-ities improve neural network acoustic models, IN ICML WORKSHOPON DEEP LEARNING FOR AUDIO, SPEECH AND LANGUAGEPROCESSING (2013). 2737] A. M. Elkahky, Y. Song, X. He, A Multi-View Deep Learning Ap-proach for Cross Domain User Modeling in Recommendation Sys-tems, in: Proceedings of the 24th International Conference onWorld Wide Web, WWW ’15, International World Wide Web Con-ferences Steering Committee, Florence, Italy, 2015, pp. 278–288. doi:10.1145/2736277.2741667 .[38] M. Schuster, K. Paliwal, Bidirectional recurrent neural networks,IEEE Transactions on Signal Processing 45 (11) (1997) 2673–2681, conference Name: IEEE Transactions on Signal Processing. doi:10.1109/78.650093 .[39] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-trainingof Deep Bidirectional Transformers for Language Understanding,arXiv:1810.04805 [cs]ArXiv: 1810.04805 (May 2019).[40] N. Dalkey, B. B. Brown, S. Cochran, The Delphi method: an experi-mental study of group opinion., Rand Corp., Santa Monica, Calif., 1969,oCLC: 762327691.[41] O. Gottesman, F. Johansson, M. Komorowski, A. Faisal, D. Sontag,F. Doshi-Velez, L. A. Celi, Guidelines for reinforcement learning inhealthcare, Nature Medicine 25 (1) (2019) 16–18, number: 1 Publisher:Nature Publishing Group. doi:10.1038/s41591-018-0310-5 .[42] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553)(2015) 436–444, number: 7553 Publisher: Nature Publishing Group. doi:10.1038/nature14539 .[43] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van denDriessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanc-tot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever,T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, D. Hassabis, Mas-tering the game of Go with deep neural networks and tree search, Nature529 (7587) (2016) 484–489, number: 7587 Publisher: Nature PublishingGroup. doi:10.1038/nature16961 .[44] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. Fei-Fei,28mageNet Large Scale Visual Recognition Challenge, arXiv:1409.0575[cs]ArXiv: 1409.0575 (Jan. 2015).[45] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Se-nior, V. Vanhoucke, P. Nguyen, T. N. Sainath, B. Kingsbury, DeepNeural Networks for Acoustic Modeling in Speech Recognition: TheShared Views of Four Research Groups, IEEE Signal Processing Mag-azine 29 (6) (2012) 82–97, conference Name: IEEE Signal ProcessingMagazine. doi:10.1109/MSP.2012.2205597 .[46] J. Hirschberg, C. D. Manning, Advances in natural language pro-cessing, Science 349 (6245) (2015) 261–266, publisher: Ameri-can Association for the Advancement of Science Section: Review. doi:10.1126/science.aaa8685 .[47] D. P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization,arXiv:1412.6980 [cs]ArXiv: 1412.6980 (Jan. 2017).[48] A. Krogh, J. A. Hertz, A simple weight decay can improve generaliza-tion, in: Proceedings of the 4th International Conference on Neural In-formation Processing Systems, NIPS’91, Morgan Kaufmann PublishersInc., Denver, Colorado, 1991, pp. 950–957.[49] K. Chen, Deep and Modular Neural Networks, in: J. Kacprzyk,W. Pedrycz (Eds.), Springer Handbook of Computational Intelligence,Springer Handbooks, Springer, Berlin, Heidelberg, 2015, pp. 473–494.[50] R. Kohavi, A study of cross-validation and bootstrap for accuracy esti-mation and model selection, in: Proceedings of the 14th internationaljoint conference on Artificial intelligence - Volume 2, IJCAI’95, Mor-gan Kaufmann Publishers Inc., Montreal, Quebec, Canada, 1995, pp.1137–1143.[51] J. G. Moreno-Torres, T. Raeder, R. Alaiz-Rodr´ıguez, N. V.Chawla, F. Herrera, A unifying view on dataset shift inclassification, Pattern Recognition 45 (1) (2012) 521–530. doi:10.1016/j.patcog.2011.06.019 .[52] T. M. Press, Dataset Shift in Machine Learning, library Catalog: mit-press.mit.edu Publisher: The MIT Press.2953] G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, S. Wermter, Continuallifelong learning with neural networks: A review, Neural Networks 113(2019) 54–71. doi:10.1016/j.neunet.2019.01.012 .[54] F. Provost, T. Fawcett, Robust classification for imprecise environments,Machine learning 42 (3) (2001) 203–231, publisher: Springer.[55] A. Rajkomar, J. Dean, I. Kohane, Machine Learning in Medicine,New England Journal of Medicine 380 (14) (2019) 1347–1358. doi:10.1056/NEJMra1814259 .[56] C. Sez, J. M. Garca-Gmez, Kinematics of Big Biomedical Data to char-acterize temporal variability and seasonality of data repositories: Func-tional Data Analysis of data temporal evolution over non-parametricstatistical manifolds, International Journal of Medical Informatics 119(2018) 109–124. doi:10.1016/j.ijmedinf.2018.09.015 .[57] C. R. Doarn, J. D. Polk, M. Shepanek, Health challenges in-cluding behavioral problems in long-duration spaceflight., Neu-rology India 67 (Supplement) (2019) S190–S195, place: India. doi:10.4103/0028-3886.259116 .[58] TEC-T, GSTP Element 1 Develop Compendium of Potential Activi-ties 2017, Tech. Rep. ESA-GSTP-TECT-PL-005452, European SpaceAgency - ESTEC (2017).[59] D. B. Toms, Detailed clinical models and their relation with electronichealth records., http://purl.org/dc/dcmitype/Text, Universitat Politc-nica de Valncia (2016).[60] J. Ball, C. Evans, Safe Passage: Astronaut Care for Exploration Mis-sions. doi:10.17226/10218 .[61] Regulation (EU) 2016/679 of the European Parliament and of the Coun-cil of 27 April 2016 on the protection of natural persons with regard tothe processing of personal data and on the free movement of such data,and repealing Directive 95/46/EC (General Data Protection Regula-tion), Tech. rep., EUROPEAN PARLIAMENT AND OF THE COUN-CIL, Brussels (2016). 3062] N. Oliver, Governance in the era of data-driven decision-making algo-rithms, Women Shaping Global Economic Governance (2019) 171.[63] M. Reyes, R. Meier, S. Pereira, C. A. Silva, F.-M. Dahlweid, H. v. Tengg-Kobligk, R. M. Summers, R. Wiest, On the Interpretability of ArtificialIntelligence in Radiology: Challenges and Opportunities, Radiology: Ar-tificial Intelligence 2 (3) (2020) e190043, publisher: Radiological Societyof North America. doi:10.1148/ryai.2020190043 .[64] M. Komorowski, S. Fleming, Intubation after rapid sequence inductionperformed by non-medical personnel during space exploration missions:a simulation pilot study in a Mars analogue environment, Extreme Phys-iology & Medicine 4 (2015) 19. doi:10.1186/s13728-015-0038-5doi:10.1186/s13728-015-0038-5