Anticipatory Mobile Digital Health: Towards Personalised Proactive Therapies and Prevention Strategies
AAnticipatory Mobile Digital Health:Towards Personalised Proactive Therapiesand Prevention Strategies
Veljko Pejovic , Abhinav Mehrotra , , and Mirco Musolesi University of Ljubljana, Slovenia University of Birmingham, United Kingdom University College London, United Kingdom
Abstract.
The last two centuries saw groundbreaking advances in thefield of healthcare: from the invention of the vaccine to organ trans-plant, and eradication of numerous deadly diseases. Yet, these break-throughs have only illuminated the role that individual traits and be-haviours play in the health state of a person. Continuous patient mon-itoring and individually-tailored therapies can help in early detectionand efficient tackling of health issues. However, even the most developednations cannot afford proactive personalised healthcare at scale. Mobilecomputing devices, nowadays equipped with an array of sensors, high-performance computing power, and carried by their owners at all time,promise to revolutionise modern healthcare. These devices can enablecontinuous patient monitoring, and, with the help of machine learning,can build predictive models of patient’s health and behaviour. Finally,through their close integration with a user’s lifestyle mobiles can be usedto deliver personalised proactive therapies. In this article, we develop theconcept of anticipatory mobile-based healthcare – anticipatory mobiledigital health – and examine the opportunities and challenges associatedwith its practical realisation.
Keywords:
Anticipatory Mobile Digital Health, Anticipatory MobileComputing, Mobile Sensing, Ubiquitous Computing, Machine Learning
Mobile computing devices, such as smartphones and wearables represent morethan occasionally used tools, and nowadays coexist with their users throughoutthe day. In addition, these devices host an array of sensors, such as a GPS re-ceiver, accelerometer, heart rate sensors, microphones and cameras, to name afew [11]. When data from these sensors are processed through machine learningalgorithms, they can reveal the context in which a device is. The context can In this paper by wearables we refer to smartwatches, smartglasses, e-garments andsimilar clothing and accessory items equipped with computing and sensing capabil-ities. a r X i v : . [ c s . C Y ] A ug Anticipatory Mobile Digital Health include anything from a device’s location to a user’s physical activity, even stresslevels and emotions [25,14]. Therefore, the personalisation and the sensing ca-pabilities of today’s mobiles can provide a close view of a user’s behaviour andwellbeing.Above all, mobile devices are always connected. They represent the most di-rect point of contact for the majority of the world’s population. Mobile phones,for example, provide an opportunity for an intimate, timely communicationunimaginable just twenty years ago. One of the consequences is that mobiledevices are becoming a new channel for the delivery of health and wellbeingtherapies. For instance, digital behaviour change interventions (dBCIs) harnesssmartphones to deliver personally tailored coaching to participants seeking be-havioural change pertaining to smoking cessation, depression or weight loss [12].Communication through a widely used, yet highly personal device ensures thata person can be contacted at all times, which might be crucial in case of suicideprevention interventions. In addition, the smartphone is used for numerous pur-poses, which protects a dBCI participant from stigmatisation that may happenif the device is used exclusively for therapeutic purposes.Besides the inference of the current state of the sensed context, an ever-increasing amount of sensor data, advances in machine learning algorithms, andpowerful computing hardware packed in mobile devices, allow the predictions ofthe future state of the context.
Context predictions have already been shown inthe domains of human mobility [1,29,3], but also population health state [15].Every next generation of mobile devices comes equipped with new sensors, andsoon we may expect galvanic skin response (GSR), heart rate, body tempera-ture oxymetry sensors as standard features . This would open up the ability toaccurately predict the health state of an individual. Anticipatory mobile computing is a novel concept that, just like context pre-diction, relies on mobile sensors to provide information upon which the models ofcontext evolution are built, yet it extends the idea with reasoning and actioningupon such predictions. The concept is inspired by biological systems that oftenuse the past, present and the predicted future state of itself and its environmentto change the state at an instant, so to steer the future state in a preferreddirection [27]. Anticipatory mobile computing has a potential to revolutioniseproactive healthcare. Health and wellbeing problems could be predicted frompersonalised sensor readings, and preventive actions could be taken even beforethe onset of a problem. We term this new paradigm –
Anticipatory Mobile Digi-tal Health , and in this paper we discuss the challenges and opportunities relatedto its practical realisation. First, we examine the key enablers (i.e., mobile andwearable sensors) that provide the contextual data which can be leveraged toinfer the health state of a user. Then, we discuss machine learning techniquesused for building predictive models of the user’s (health) context. We are par-ticularly interested in the models that describe how the context might changeafter an intervention or a therapy. We investigate the challenges related to un- See for example the proposal by Intel: http://iq.intel.co.uk/glimpse-of-the-future-the-healthcare-smartwatch/ .nticipatory Mobile Digital Health 3 obtrusive learning of the impact of an intervention to a person, and the oppor-tunities for highly personalised healthcare. We also take into account individualdifferences among users, and the potential for capturing and including geneticpre-determinants into the system. We continue with the examination of human-computer interaction issues related to the therapy delivery, and conclude with aconsideration of ethical issues in anticipatory mobile digital health. Finally, whilewe have examined the potential for inducing a change in a person’s behaviourthrough anticipatory mobile computing before [20], this paper extends the ideaon the much larger domain of digital healthcare, and elaborates on particularchallenges and opportunities in the area.
The use of wireless and wearable sensors represents a novel and a rapidly evolvingparadigm in healthcare. These sensors have the potential to revolutionise theway of assessing the health of a person. Sensor embedded devices are givento the patients in order to obtain their health related data remotely. Thesedevices do not only help a patient in reducing the number of visits to the clinic,but also offer unprecedented opportunities to the practitioners for diagnosingdiseases and tailoring treatments through continuous real-time sampling of theirpatients’ heath data. Furthermore, some of these devices empower the users withthe ability to self-monitor and curb certain well-being issues on their own.Today’s mobile phones are laden with sensors that are able monitor contextvarious modalities such as physical movement, sound intensity, environment tem-perature and humidity, to name a few. Some previous studies have showed thepotential of mobile phones in providing data that can be used to infer the healthstate of a user [4,5,10,2]. Houston [4] and UbiFit [5] are the early examples ofmobile sensing systems designed to encourage users to increase their physicalactivity. Houston monitors a user’s physical movement by counting the numberof steps taken via an accelerometer that serves as a pedometer. Whereas, UbiFitrelies on the Mobile Sensing Platform (MSP) [13] to monitor varied physical ac-tivities of a user. MSP is capable of inferring physical activities including walking,running, cycling, cardio and strength exercise, and other non-exercise physicalactivities, such as housework. BeWell is a mobile application that continuouslymonitors a user’s physical activity, social interaction and sleeping patterns, andhelps the user manage their wellbeing [10]. Bewell relies on sensors such as ac-celerometer, microphone and GPS, which are embedded in mobile phones. In [2]the authors show that the depressive states of users can be inferred purely fromlocation and mobility data collected via mobile phones. The above examplesdemonstrate the close bond between smartphone sensed data and different as-pects of human health and well-being.A particularly interesting example of mobile healthcare monitoring is givenby
LifeWatch , a smartphone that is equipped with health sensors that con-stantly monitor the user’s vital parameters including ECG, body temperature, Anticipatory Mobile Digital Health blood pressure, blood glucose, heart rate, oxygen saturation, body fat percent-age and stress levels. A user has to perform a specific action in order to takehealth measurements. For example, a user should hold the phone’s thermometeragainst the forehead in order to measure the body temperature, and to takeECG readings, the user should clutch the phone horizontally with a thumb andforefinger placed directly on top of a set of sensors that are placed on the sides ofthe phone. The sensor data are sent to the cloud for the analysis and the resultsare delivered back to the user within a short time interval. Such a phone canprove to be extremely useful in the healthcare domain. However, there is still noproof of the accuracy of its results.Although within their owners’ reach for most of the time, smartphones do notstay in a constant contact physical with the users, and consequently are limitedwith respect to personal data they can provide. More recently, mobile phone com-panies have introduced smartwatches that link with mobile phones and enablethe users to perform actions on the mobiles without actually interacting withthem. These devices open up new possibilities for health data sensing. First, theymaintain continuous physical contact with their users, and second, they host anew set of sensors, usually unavailable on traditional smartphones. In generalthese devices come with the accelerometer, heart rate and body temperature sen-sors. Smartwatches are inspired by the concept of a smart-wristband, a devicethat monitors the health state of a user and presents it in a visual form on thelinked mobile phone. Smart-wristbands enable real-time health state monitoring,and have achieved a considerable commercial success among health-aware pop-ulation (e.g. Jawbone ). Initially these bands were able to report only a user’sphysical activity. However, new sensors, such as body temperature and hearthrate, have been introduced, together with a more sophisticated data analyticsand presentation to the user.Mobile sensing on the phone is for the majority of readings limited by theamount of physical contact the user makes with the phone. Smartwatches andsmart-wristbands ensure that the contact is there, yet are limited to a par-ticular part of the users body – her wrist. Sensor embedded smart-wearablesdesigned to dedicatedly monitor specific health related parameter from a spe-cific part of a user’s body, have appeared recently and promise more reliablesensing. Such smart-wearables could enable healthcare practitioners to obtaintheir patients’ health data continuously and in the natural environment of thepatient. These devices come with a variety of health sensors. Pulse and oxygen inblood sensor, airflow sensor, body temperature sensor, electrocardiogram sensor,glucometer sensor, galvanic skin response sensor (GSR), blood pressure sensor(sphygmomanometer), and electromyography sensor (EMG), are some examplesof the health sensors embedded in the smart-wearables. Some examples of smart-wearables include Epoc Emotiv [6], an EEG headset capable of capturing brainsignals that can be analysed to infer a user’s thoughts, feelings, and emotions.
MyoLink is another wearable that can continuously monitor the user’s musclesand heart. It can capture muscle energy output, which in turn can be used to jawbone.com nticipatory Mobile Digital Health 5 quantify the user’s fatigue, endurance and recovery level. Also, it can be placedon the chest to continuously track the heart rate of the user. ViSi mobile , wornon a wrist, measures blood pressure, haemoglobin level, heart rate, respirationrate, and skin temperature. The device is highly portable and enables the userto monitor their health at anytime and anywhere.The next step in wearable computing is the one in which devices becomecompletely stealth, and as in the Weiser’s vision of pervasive computing, com-pletely integrated with people’s lives [32]. Shrinking the size of smart-wearablesis push in that direction, for example reducing the size of a device from some-thing obtrusive to a small adaptive device that the user can wear on their bodiesand forget about it. BioStamp [23] is a device composed of small and flexibleelectronic circuits that stick directly to the skin like a temporary tattoo andmonitors the user’s health. It is a stretchable sensor capable of measuring bodytemperature, monitoring exposure to ultraviolet light, and checking pulse andblood-oxygen levels. The company envisions future versions of BioStamp ableto monitor changes in blood pressure, analyse sweat, and obtain signals fromthe user’s brain and heart in order to use them in electroencephalograms andelectrocardiograms [23].These wearable sensors enable the continuous measurement of health metricsand deliver treatment to the patients on time. Yet, the difficulty of continuousmonitoring is not the only problem in modern healthcare. Recent studies haveshown that around 50% of the prescribed drugs are never taken [18,19], andthus, prescribed therapies fail to improve the health of the patients [26]. In or-der to address this problem, Hafezi et al. [7] proposed Helius , a novel sensorfor detecting the ingestion of a pharmaceutical tablet or a capsule. The systemis basically an integrated-circuit micro-sensor developed for daily ingestion bypatients, and as such allows real-time measurement of medication ingestion andadherence patterns. Moreover, Helius enables practitioners to measure the cor-relation between drug ingestion and patients health parameters, e.g. physicalactivity, heart rate, sleep quality, and blood pressure, all of which can be sensedby mobile sensors.The ecosystem of devices supporting health sensing is already substantial andconstantly increasing. Soon, healthcare practitioners will have a remote multi-faceted view of a patient’s health in real time. The key enabler is the unobtru-siveness of these smart sensing devices. Furthermore, issues such as the accuracyof measurements, accountability for mistakes and the security of a user’s pri-vacy, need to be thoroughly addressed before these devices can penetrate intothe official medical practice.In this paper we discuss the novel concept of anticipatory mobile digitalhealth, outlining the challenges and opportunities in this promising field. Al-though smart health sensing devices are still in their infancy, we believe that wewill witness a rapid evolution of this research area in the coming years. Anticipatory Mobile Digital Health
Anticipation, for living systems, is the ability to reason upon past, present andpredicted future information. Such, a systems Rosen described as “a systemcontaining a predictive model of itself and/or its environment, which allows itto change state at an instant in accord with the model’s predictions pertainingto a later instant” [27], thus indicated that there is an internal predictive modelthat an anticipatory system builds and maintains. The concept of an antici-patory computing system envisions a digital implementation of such a model,and automated actioning based on the model’s predictions. Yet, an anticipatorycomputing system is of interest only if the anticipation carries a value for theend-user.We argue that modern mobile computing devices fulfil the necessary prereq-uisites for anticipatory computing. First, thanks to built-in sensors and person-alised usage these devices can gather the information about theirs, and indirectlythe user’s state, and the state of the environment; second, their computing ca-pabilities allow devices to build predictive models of the evolution of the state;finally, the bond between a device and its end-user is so tight that automatedsuggestions (based on the anticipation) a device might convey to a user, arelikely to influence the user’s actioning. After all, people already look into theirsmartphones when they need to navigate in a new environment or choose arestaurant. To clarify the concept of anticipation on mobile devices (termed
An-ticipatory Mobile Computing ), in Figure 1 we sketch a system that senses thecontext and builds a model of the environment evolution, which gives it theoriginal predicted future. The system then evaluates the possible outcome of itsactions on the future. An action that leads to the preferred modified future isrealised through the feedback loop that involves interaction of the system withthe user.
Fig. 1.
Anticipatory mobile systems predict context evolution and the impact thatcurrent actions can have on the predicted context. The feedback loop consisting of amobile and a human enables the system to affect the future.nticipatory Mobile Digital Health 7
The opportunity to infer the health and well-being state of an individual withthe help of mobile sensing, together with the perspective of anticipatory mo-bile computing, pave the way for preventive healthcare through anticipatorymobile healthcare systems. We sketch the main ideas behind such a system inFigure 2. Physiological (e.g. heart rate, GSR) and conventional mobile sensors(e.g. GPS, accelerometer) provide training data for machine learning models ofthe context (e.g. a user’s depression level) and its evolution. The models predictthe future state of the context, termed the original future, and the state afteran intervention or a therapy, termed the modified future. Based on the predic-tions, a therapy with the most preferred outcome is selected and conveyed tothe user. Finally, different users may react differently to the same therapy, andclose sensor-based patient monitoring, together with a-priori inputs, such as auser’s genetic background, are used to custom tailor the therapies.
Fig. 2.
Anticipatory mobile systems predict context evolution and the impact theiractions can have on the predicted context. The feedback loop consisting of a mobileand a human enables the system to affect the future.
A practical realisation of an anticipatory mobile digital health system requiresthat the following building blocks are present: – Mobile sensing.
The role of this block is to manage which of a number ofavailable mobile sensors are sampled, and how often. Mobile devices’ sensorswere originally envisioned as occasionally used features, and their frequentsampling can quickly deplete a device’s battery. At the same time, importantevents may be missed if sampling is too coarse. – Therapy and prevention toolbox.
This block contains definitions of pos-sible therapies and prevention strategies that can be delivered to the user.Although in future we envision further automatisation of this module, fornow, we feel that a professional therapist’s expertise should be harnessed tolimit the number of possible therapies, and oversee their deployment. – Machine learning core.
Anticipatory mobile digital health employs ma-chine learning for two separate aspects of health state evolution modelling:
Anticipatory Mobile Digital Health context evolution model and therapy/prevention-effect model . The formerconnects sensor data with higher-level context, and provides a predictivemodel of how the context might evolve. The latter provides a picture of howdifferent therapies might affect a user’s health state. We discuss these modelsin detail in the next section. – User interaction interface.
The success of an anticipatory mobile digitalhealthcare system is limited by the user’s compliance with the providedtherapy and prevention strategy. The look, feel and the behaviour of themobile application that delivers the therapy or prevention strategy to a useris crucial in this step. in the following section, we also discuss the challengesin designing a successful user interaction interface.
Numerous challenges obstruct the path towards implementations of anticipatorymobile digital healthcare systems. Rooted in mobile sensing, anticipatory mobiledigital health faces challenges such as resource, primarily energy, inefficiency ofcontinuous sensing, and the difficulty of reliable context modelling. Yet, thesechallenges are common for a larger field of mobile sensing, and a thorough dis-cussion on these issues is available elsewhere [11,22,9]. Instead, here we focus onaspects that are unique to anticipatory mobile healthcare. The use of machinelearning algorithms to model and predict user behaviour and the effect of a ther-apy or a prevention strategy on the future health state of a specific user is themain challenge. The value of machine learning models, for instance, increaseswith the amount of available training data for her. Second, the mobile moni-tors the user, and may suggest therapies, yet, it is the user herself that decideswhether to take the therapy or to follow certain preventive measures or not.Besides machine learning, future anticipatory mobile digital health developersshould pay a special attention to the human-computer interaction issues in thisfield, and try to answer – what is the best way to convey an advice/therapy toa user, so that the compliance with the proposed therapy or prevention strat-egy is the highest? Finally, the area of ethics, responsibility and entity roles inanticipatory mobile digital health remains an uncharted territory. In the rest ofthe section we discuss each of the challenges individually, and provide positionalguidelines for overcoming the challenges.
Anticipatory mobile digital health, as stated in the previous section, employsmachine learning for two separate aspects of health state evolution modelling: context evolution model and therapy/prevention-effect model . First, a model ofa user current and predicted future health state is needed. In this model, arelationship between mobile sensor data and high-level health state is built. Themodel can be direct, if certain values of physiological sensor readings indicatea certain health state, or indirect, if sensor readings reveal contextual aspects nticipatory Mobile Digital Health 9 that can be connected to a health state of an individual – for instance, GPSreadings can reveal user mobility, which in turn can hint a user’s depressionstate [2]. In the next step, inference models are extended to provide predictionsof the future state of the health state, either directly, or indirectly through thepredictions of the future context. Forecasting user’s next location is an activearea of research, with substantial achievements [1,29,3]. For many other aspectsof the user’s behaviour and health state, reliable predictive models still do notexist, and even the possibility of them being built remains an open question.The second major machine learning model in anticipatory mobile digitalhealth is the model of the impact of a possible therapy or prevention strategyon the predicted future health state of a user. There are two non-exclusive waysto construct such a model: one is to harness the existing expertise in healthcareto map available therapies to health state transitions. For example, we couldmap antidepressants to a transition from depressive states to a healthy state.However, these rules are not suitable for preventive healthcare. Anticipatorymobile digital health operates on predictions, and consequently therapies shouldaim to prevention . In addition, although mobile devices remain highly personal,and the sensor data uncovers fine-grained individual health state information,these general rules limit the ability of the system to deliver personalised health-care. An alternative approach is to build a therapy/prevention-effect model bymonitoring the evolution of a user state after a proactive therapy or preventionstrategy is delivered. By comparing the original predicted state with the ac-tual state recorded some time after the therapy (or prevention strategy), we canidentify the relationship between the therapy (or prevention strategy) and thefuture health state change. Built this way, a model reveals successful proactivetherapies, which is difficult to achieve in the traditional practice. Moreover, whatworks for one patient may not work for another – these models are highly person-alised, and can reveal therapies that are useful for a particular kind of a persononly. Still, we argue that these models should not be built from the scratch – theavailable therapies that could be automatically suggested to a particular patientin a particular situation should be determined by the rules stemming from theexisting medical expertise.
Learning with a user.
Automated tool-effect modelling in anticipatorymobile digital health requires that a therapy (or prevention strategy) is inducedto a user so that its effects can be observed. This outcome is then used to trainand refine the model.
Reinforcement learning where an agent uses a tool inthe intervention environment (which for example can be represented through aMarkov decision process) is a natural way to model the problem [30]. In everystep, a certain tool is selected, used, and the observed change in the health stateelicits a reward that reflects how positive the change is.
Measuring health state.
Thus, there is a need for a suitable metric for measuring the health state change . Here we need to evaluate the effect of aproactive therapy or prevention strategy, basically compare the original predictedhealth state and the modified predicted health state . We argue that the comparisonmetric has to be domain dependent. For example, if an anti-stress therapy is evaluated, the difference between the predicted skin conductivity and heart ratevalues without an intervention, and the actual values after the intervention,is a reasonable measure of stress level change [8]. However, system designersshould have in mind that the metric has to be both suitable for machine learningalgorithms as well as relevant from the healthcare point of view.
Learning without interfering.
Reinforcement learning uncovers the map-ping between therapies and health state changes. Delivering a previously unusedtherapy or prevention strategy refines the model, as we learn more about howthe user reacts to this tool. From the practical point of view, however, we face adilemma: use a tool that is known to result in a positive health change outcome,or experiment with an unused tool that might yield an even better outcome.In reinforcement learning this dilemma is known as exploration vs. exploitationtrade-off . Strategies for solving the dilemma in an anticipatory mobile digitalhealth setting should be aware of the possible irreversible negative consequencesof a wrong therapy or prevention strategy. Preferably, the system should learn asmuch as possible without explicit delivery of therapies to a user. Such a learn-ing concept is called latent learning . It is a form of learning where a subjectis immersed into an unknown environment or a situation without any rewardsor punishments associated to them [31]. Latent learning has been demonstratedin living beings who form a cognitive map of the environment solely becausethey are immersed into the environment, and later use the same map in decisionmaking. We argue that mobile computing devices, through multimodal sensing,can harness latent learning to build a model of the user reaction with respectto certain actions or environmental changes that correspond to ones targeted bythe therapies. This is particularly relevant for therapies that are not based onmedications, such as behavioural change interventions [20]. For example, sup-pose a depression prevention system can provide the user with the suggestionto go out for a dinner with friends. We can get an a priori knowledge of howthis suggestion would affect the user, for example if on a separate occasion wedetect that the participant went out for a dinner with friends, and we gaugethe depression levels, estimated through mobility and physical activity metrics,before and after the dinner. Defining how the expected action – going out withfriends – should manifest from the point of view of sensors – e.g., a number ofBluetooth contacts detected, location, time of the day – is one of the prerequi-sites for practical latent learning. Again, interdisciplinary efforts are crucial toensure that the detected reaction corresponds to the one that should be elicitedby the tool.
Current therapies are often created as “one size fits all”, yet in many casesindividuals react differently. For example, antidepressants are ineffective in 38%of the population, while cancer drugs work for only one quarter of the patientpopulation [24]. Personalised therapies promise to revolutionise healthcare, byavoiding the traditional trial-and-error therapy prescription, minimising adverse nticipatory Mobile Digital Health 11 drug reactions, revealing additional or alternative uses for medicines and drugcandidates [16], and curbing the overall cost of healthcare [24].Anticipatory mobile digital health is poised to bring personalised healthcarecloser to mainstream practice. Not only can mobile sensing provide a glimpse intoindividual behavioural patterns, identifying risky lifestyles, but therapy/prevention-effect machine learning models can also take into account a patient’s genetics inorder to individualise the therapy or prevention strategy. Investigation of whichgenes impact the occurrence and reaction to a treatment of a certain diseaseis a very active area of research. The potential for healthcare improvement isimmense, having in mind that with some conditions, such as melanoma tumors,the majority of cases are driven by certain person-specific genetic mutations,and could be targeted by specific drugs [24]. The relationship is not one way,and anticipatory mobile digital health could also help with pharmacogenomics,the study of how genes affect a person’s response to drugs. Identifying com-mon pieces of genetic background in populations who reacted to an anticipatorytherapy or prevention strategy in the same way would help find the relation-ship between genes and health treatments. Finally, the inclusion of the geneticbackground in the common medical practice is not far from reality – in 2014 ahuman genome sequencing for less than USD $1000 became available.
Despite the automation that anticipatory mobile digital health brings, in theend, it is up to a user to comply with the given therapy or prevention strategy.This is particularly important for behavioural change intervention therapies, thatare delivered in cases where the health state is directly influenced by patient’sbehaviour. Consequently, the communication between the system and the patienthas to be seamless. Users are an important part of the system, and their inclusionrequires an appropriate interface between the participant and the system. Asnoted by Russell et al. [28], a system that autonomously brings decisions andevolves over the course of its lifetime needs to be transparent to the user. Throughthe user interface such a system must be understandable by the user and capableof review, revision, and alteration. In addition, the content should be framed toemphasise that the tool can help, yet it is fundamental to avoid to harass andpatronise the participant.The timing of a therapy or a prevention strategy is also important for itssuccessful delivery. This is particularly true for automated therapies delivered viaa mobile device. An inappropriately timed intervention that comes, for instance,when a patient is in a meeting, or riding a bicycle, may lead to annoyance, or maybe completely overlooked by the patient. Mobile sensing helps with identifyingopportune moments to deliver therapies. The context in which a user is, such asher location, physical activity and engagement in a task, to an extent determinesher interruptibility [21,17]. Machine learning and mobile sensing is harnessed formonitoring a user’s reaction to an interruption arriving when the user is in acertain context, and from there on a model of personal interruptibility is built.Querying the model with a momentarily value of a user’s context returns the estimated interruptibility at the moment. While practical implementations of theabove models already exist [21], in future we envision predictive models of userinterruptibility. Finally, we highlight that opportune moments denote those timeat which a patient is likely to quickly acknowledge/read the content of a deliveredmessage. Identifying moments at which the delivered information will have thehighest medical impact is even more important, yet due to the difficulty of gettingthe training data (we would need to deliver the same therapy or preventionstrategy at different times to the same user) identifying such moments remainsvery challenging.
Privacy issues in mobile sensing emerged soon after the proliferation of smart-phones started about a decade ago. Misuse and leaking of information that canbe collected by a mobile device, such as a user’s location, collocation with otherpeople, physical activity of a user, may deter people from trusting mobile appli-cation. Trust is a key component for the success of anticipatory mobile digitalhealth applications, and every care should be taken that personal informationdoes not leak. Ensuring that sensor data do not leave the device at which theywere collected is one way to minimise the risk. However, this complicates theconstruction of joint machine learning models discussed earlier.The responsibility chain in the domain of anticipatory mobile digital healthis yet to be defined. Unsuccessful therapies can have serious consequences. Itis unclear who is to blame if a delivered therapy or prevention strategy doesnot improve the health state of a patient, or even worse, endangers the person’slife. A therapist who designed the therapy, a software architect who devised theunderlying machine learning components, and the patient herself, all play a rolein the process.
Personalised and proactive healthcare brings undisputed benefits in terms oftherapy (or prevention strategy) efficiency and cost effectiveness of the health-care system. Mobile devices have a potential to become both our most vigilantobservers, and closest advisors. Anticipatory mobile digital health harnesses thesensing capabilities of mobiles to learn about the user health state and predictits evolution, so that proactive therapies tackling predicted health issues are de-liver to the user in advance. With the help of machine learning that takes intoaccount rich sensor data and a user’s genetic background, anticipatory mobiledigital health applications can tailor personalised therapies. Yet, in addition,the concept can be used to learn more about how therapies affect different de-mographics, users who behave in a certain way, or have a particular geneticbackground. Generalising from a larger pool of users and therapies can identifygroups for which a therapy (or prevention strategy) is successful, basically un-covering new facts about drugs. Finally, we believe anticipatory mobile health nticipatory Mobile Digital Health 13 applications warrant a discussion on their inclusion into the health insuranceframeworks.
Acknowledgements
This work was supported by the EPSRC grants “UBhave: ubiquitous and so-cial computing for positive behaviour change” (EP/I032673/1) and “Trajectoriesof Depression: Investigating the Correlation between Human Mobility Patternsand Mental Health Problems by means of Smartphones” (EP/L006340/1).
References
1. D. Ashbrook and T. Starner. Using GPS to Learn Significant Locations and PredictMovement Across Multiple Users.
Journal of Personal and Ubiquitous Computing ,7(5):275–286, October 2003.2. L. Canzian and M. Musolesi. Trajectories of Depression: Unobtrusive Monitor-ing of Depressive States by means of Smartphone Mobility Traces Analysis. In
UbiComp’15 , Osaka, Japan, September 2015.3. Y. Chon, E. Talipov, H. Shin, and H. Cha. SmartDC: Mobility Prediction-basedAdaptive Duty Cycling for Everyday Location Monitoring.
IEEE Transactions onMobile Computing , 13:1, 2013.4. S. Consolvo, K. Everitt, I. Smith, and J. A. Landay. Design requirements fortechnologies that encourage physical activity. In
CHI’06 , Quebec, Canada, April2006.5. S. Consolvo, D. W. Mcdonald, T. Toscos, M. Y. Chen, J. Froehlich, B. Harrison,P. Klasnja, A. Lamarca, L. Legr, R. Libby, I. Smith, and J. A. Landay. ActivitySensing in the Wild: a Field Trial of UbiFit Garden. In
CHI’08 , Florence, Italy,April 2008.6. T. Fraga, M. Pichiliani, and D. Louro. Experimental art with brain controlledinterface. In
Universal Access in Human-Computer Interaction. Design Methods,Tools, and Interaction Techniques for eInclusion , pages 642–651. Springer, 2013.7. H. Hafezi, T. L. Robertson, G. D. Moon, K. AuYeung, M. J. Zdeblick, and G. M.Savage. An Ingestible Sensor for Measuring Medication Adherence.
IEEE Trans-actions on Biomedical Engineering , 62(1):99–109, 2015.8. J. Healey and R. W. Picard. Detecting stress during real-world driving tasks usingphysiological sensors.
IEEE Transactions on Intelligent Transportation Systems ,6(2):156–166, 2005.9. P. Klasnja and W. Pratt. Healthcare in the Pocket: Mapping the Space of Mobile-Phone Health Interventions.
Journal of Biomedical Informatics , 45(1):184–198,2012.10. N. D. Lane, T. Choudhury, A. Campbell, M. Mohammod, M. Lin, X. Yang, A. Do-ryab, H. Lu, S. Ali, and E. Berke. BeWell: A Smartphone Application to Monitor,Model and Promote Wellbeing. In
Pervasive Health’11 , Dublin, Ireland, May 2011.11. N. D. Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury, and A. T. Campbell. ASurvey of Mobile Phone Sensing.
IEEE Communications Magazine , 48(9):140–150,2010.12. N. Lathia, V. Pejovic, K. Rachuri, C. Mascolo, M. Musolesi, and P. J. Rentfrow.Smartphones for Large-Scale Behaviour Change Intervention.
IEEE PervasiveComputing , 12(3), 2013.4 Anticipatory Mobile Digital Health13. J. Lester, T. Choudhury, and G. Borriello. A practical approach to recognizingphysical activities. In
Pervasive Computing , pages 1–16. Springer, 2006.14. H. Lu, G. T. C. Mashfiqui Rabbi, D. Frauendorfer, M. S. Mast, A. T. Camp-bell, D. Gatica-Perez, and T. Choudhury. StressSense: Detecting Stress in Uncon-strained Acoustic Environments using Smartphones. In
UbiComp’12 , Pittsburgh,PA, USA, September 2012.15. A. Madan, M. Cebrian, D. Lazer, and A. Pentland. Social Sensing for Epidemiolog-ical Behavior Change. In
UbiComp’10 , Copenhagen, Denmark, September 2010.16. L. Mancinelli, M. Cronin, and W. Sade. Pharmacogenomics: the promise of per-sonalized medicine.
AAPS Pharmsci , 2(1):29–41, 2000.17. A. Mehrotra, M. Musolesi, R. Hendley, and V. Pejovic. Designing Content-drivenIntelligent Notification Mechanisms for Mobile Applications. In
UbiComp’15 , Os-aka, Japan, September 2015.18. K. Nasseh, S. G. Frazee, J. Visaria, A. Vlahiotis, and Y. Tian. Cost of medicationnonadherence associated with diabetes, hypertension, and dyslipidemia.
AmericanJournal of Pharmacy Benefits , 4(2):e41–e47, 2012.19. L. Osterberg and T. Blaschke. Adherence to medication.
New England Journal ofMedicine , 353(5):487–497, 2005.20. V. Pejovic and M. Musolesi. Anticipatory Mobile Computing for Behaviour ChangeInterventions. In
Workshop on Mobile Systems for Computational Social Science(MCSS’14) , Seattle, WA, USA, September 2014.21. V. Pejovic and M. Musolesi. InterruptMe: Designing Intelligent Prompting Mech-anisms for Pervasive Applications. In
UbiComp’14 , Seattle, WA, USA, September2014.22. V. Pejovic and M. Musolesi. Anticipatory Mobile Computing: A Survey of theState of the Art and Research Challenges.
ACM Computing Surveys (CSUR) ,47(3), 2015.23. T. S. Perry. Giving your body a” check engine” light.
Spectrum, IEEE , 52(6):34–84,2015.24. Personalised Medicine Coalition. The Case for Personalized Medicine, 2014.25. K. K. Rachuri, M. Musolesi, C. Mascolo, P. J. Rentfrow, C. Longworth, andA. Aucinas. EmotionSense: A Mobile Phones based Adaptive Platform for Ex-perimental Social Psychology Research. In
UbiComp’10 , Copenhagen, Denmark,September 2010.26. J. N. Rasmussen, A. Chong, and D. A. Alter. Relationship between adherence toevidence-based pharmacotherapy and long-term mortality after acute myocardialinfarction.
Jama , 297(2):177–186, 2007.27. R. Rosen.
Anticipatory Systems . Pergamon Press, Oxford, UK, 1985.28. D. M. Russell, P. P. Maglio, R. Dordick, and C. Neti. Dealing with Ghosts: Manag-ing the User Experience of Autonomic Computing.
IBM Systems Journal , 42:177–188, 2003.29. S. Scellato, M. Musolesi, C. Mascolo, V. Latora, and A. T. Campbell. Nextplace:A spatio-temporal prediction framework for pervasive systems. In
Pervasive’11 ,San Francisco, CA, USA, June 2011.30. R. S. Sutton and A. G. Barto.
Reinforcement Learning: An Introduction . MITPress, Cambridge, MA, USA, 1998.31. C. Tavris and C. Wade.
Psychology in Perspective . Longman, New York, NY,USA, 1997.32. M. Weiser. The Computer for the 21st Century.