14 Supporting active patient self-care
Abstract: We are currently confronted with a trend of increased pressure on health care, with associated increasing financial costs, due to an aging society and the expected increase in the prevalence of disability and chronic disease. Finding measures for cost reduction, without sacrificing quality of care, is a significant healthcare challenge. Computing technology offers promising solutions in this respect. In this article, we review contributions made by mobile computing technology in supporting the care process. In particular, we consider mobile decision support with a view to enable patient self-management, transferring part of the clinical care from healthcare professionals to the patient. Mobile computing can play an important role in giving patients an active, decisive role in managing their own disease.
14.1 Introduction
Current challenges in health care include cost reduction without sacrificing quality of care, and improving the quality of life. A reduction in the number of face-to-face consultations in outpatient clinics and the prevention of costly hospitalization through early risk-detection are typical examples of cost reduction indicators. At the same time, patients’ involvement in making decisions about their health and the management of their diseases has become more and more important. Measures taken to realize this have been referred to as patient empowerment. A particularly important step to achieve this is adapting decision-making to the individual’s characteristics, usually referred to as personalization.
In this chapter, we consider the computing aspects of turning a mobile computing and communication platform into a personalized smart care-assistant. The architecture and the typical set-up of user interaction modules of such a smartphone-based monitoring system is presented in Figure 1. Such care assistants allow part of the clinical decision-making process to be moved from the clinic to anywhere where the patient resides: at home; at work; or on holiday. In particular, we consider how such a care assistant can be constructed and what the important characteristics are in order to support active decision-making by patients. Although we focus on self-management of one disease in this chapter, the aim is to deliver optimal health care, offering support to both patients and physicians. In the current state of affairs there is little to no computing research addressing the problem of managing multiple diseases at the same time. Mobile computing is expected to offer an important contribution to cost reduction, quality improvement and larger patient participation in health care.[1], [2]
Figure 1: Modules and user interactions in the smart care-assistant’s framework.
14.2 The current situation: needs, gaps and challenges
14.2.1 Health care at distance: a brief history
Figure 2 offers a visual summary of the development of the field. Starting as early as 1905 and leading up to the present, we see an increase in patient involvement in tandem with technological advances.
From the mid-1990s on, eHealth emerged as a promising field for better and more efficient healthcare delivery using web-enabled services. In comparison to telehealth, eHealth is a broader term that encompasses health services, information, education and research. Subsequently, many eHealth systems and tools were developed, including commercial ones such as the Bosch Health Buddy and Intel Health, to monitor a patient’s condition, usually in the home environment, and enable earlier diagnosis and more effective treatment.
The availability of modern mobile computing and communication technology offers new avenues to move part of the medical decision-making tasks, such as diagnosis, selection of appropriate treatment and prognosis from the doctor to the patient at any place and any time. This technology has given rise to the field of mobile health or mHealth, a term coined in 2004.[4] Currently, the most commonly used mHealth technologies include smartphones, wireless tablet computers, wearable wireless biosensors and disease-monitoring devices.
Figure 2: Timeline development of health care at a distance. *Numbers as reported in [3].
A large number of mobile health applications are already available, focusing primarily on healthy-lifestyle tracking, or on the remote monitoring of a few basic vital signs and symptoms in combination with the transmission of data to the clinician for interpretation. These systems simply collect data and provide no, or only very limited, user feedback. They lack the clinical interpretation capabilities of a medical doctor, who takes into account the patient’s characteristics, history and environment, including the nature of the underlying disease. Within the context of disease management, this implies that the patient is passive and mostly dependent on the caregiver, while the clinicians may be overwhelmed by real-time incoming patient data.
14.2.2 Clinical decision-making
Part of clinical decision-making concerns the process of finding a diagnosis. A clinician has to decide, based on the patient’s history and current clinical observations, on the set of possible explanations of the presented symptoms and signs. The relevant decisions are whether to perform tests, and if so, which tests will distinguish the diagnoses. Based on the results, a further choice has to be made on what treatment is appropriate, taking into account disease prognosis. A challenging aspect of these processes is dealing with uncertainty.
Clinical decision support systems emerged as early as the 1970s to assist medical doctors with difficult tasks such as diagnosis and treatment selection by using artificial intelligence techniques.[5], [6] The basic knowledge for the development of clinical decision support systems comes from evidence-based clinical guidelines, treatment protocols, research practices and input from clinical experts. Providing assistance in the management of one or two disorders that are limited in nature is now feasible given the current state of the art. However, reliably dealing with a broad range of diseases is still not feasible, due to the inherent complexity of disease interaction.
A care assistant primarily developed for use by patients by definition operates within the confines of a disease that is already diagnosed. The important aspect of chronic disease care is managing the progress of the disease and responding adequately to deteriorating health. Although this does not change the fundamental characteristics of the problem, the domain is reduced to more manageable proportions that can be captured in a computerized decision-support system.
Requirements for smart mHealth systems have been formulated in the literature. [7] However, actual systems embedding clinical decision-support were either implemented on devices with limited computational capabilities and therefore unable to deal with complex clinical problems;[8], [9] or rely on transmitting data to a healthcare professional for interpretation (see e.g. [10] for examples concerning chronic obstructive pulmonary disease (COPD)). Decision aids that include contextualized data interpretation are not in wide use yet, and progressing from specifications and prototypes to practical systems appears to be the current challenge.
14.3 Proposed solutions: personal smart care-assistants
14.3.1 Remote patient data collection
The continuous monitoring and assessment of the individual’s health status requires the collection of biomedical data including symptoms (e.g. fatigue, headache), signs (e.g. blood pressure, temperature) and biosignals (e.g. electrocardiography (ECG), obstetric ultrasonography). Such data collection outside of the hospital and performed by the patient has been facilitated by recent technological advances in sensors and measuring devices that are compact, cheap, easy to use and wirelessly connected. Compact readers for performing stand-alone biochemical tests such as glucose, hemoglobin and urine analysis are already widely used in home-based practice. Recent developments include usage of the built-in camera for automated image analysis of biochemical test results on the mobile device.[11], [12]
Current mobile technology is capable of complex processing of biosignals, including filtering noise, sampling and quantification. This is exemplified by products such as mobile ECG devices allowing real-time heart monitoring and wireless transmission of data to a smartphone or tablet. Furthermore, body area networks – small, intelligent devices attached to or implanted in the body – provide real-time feedback to the user or medical staff via wireless communication.[13]
Despite the wide availability of measurement devices, using them to provide raw data to mobile devices for further analysis sometimes remains a challenge, because the manufacturer often restricts access to the communication protocols or requires the user to get external software for reading the measurements. Currently this poses limitations on the development of flexible and personalized smart care assistants and restricts both patients and doctors in choosing from the sensors and devices available on the market. A further limitation is to ensure that the measurements are reliably taken by the patient, for which particular procedures, skills or training may be required. See also [14] for a review of wireless technology in disease management.
For data collection, one aspect that could be personalized is the nature of alerts. Whereas a user who often interacts with smartphone apps possibly would not mind being interrupted by relevant warnings, such as an alert for medication intake, another user might dislike such behavior. Preferred days and times should therefore be taken into account when presenting reminders concerning measurements, medication intakes or filling out questionnaires.
Depending on the individual’s health status, the rate of data acquisition can also be varied. If the model predicts a low current risk, monitoring can take place on a weekly basis, keeping the time investment at a minimum. When the risk increases, the system check-in can be automatically adjusted and scheduled daily to ensure the possible worsening is detected and acted upon appropriately. Whether this is appropriate requires an assessment of the cost of data acquisition versus the risk of missing a clinically relevant event.
14.3.2 Embedded intelligent models
For patient self-management, decision aids should be embedded within the mobile devices – enabling direct data interpretation – to provide accurate information about options and outcomes, and help patients become involved in decisions concerning their health. Given the cause-effect relationships and the uncertainty inherent in the medical domain, we use a special class of probabilistic graphical models – Bayesian networks – as appropriate technique for building clinical decision models for smart care-assistants. A Bayesian network is an acyclic-directed graph consisting of vertices representing random variables of interest and arcs representing dependencies between variables.[15] Each random variable has a quantitative part, denoting conditional probabilities of the type P (Y |pa(Y)); that is, the probability that Y takes on a specific value given the values of its parent variables pa(Y). Probabilities of interest can be computed from the joint probability of all variables, e.g. the probability of health deterioration given the evidence obtained from monitoring. An important observation is that although the model describes general relations between the variables of interest, all predictions are personalized by entering patient-specific data. Bayesian network models can be built manually using domain knowledge, learned from data or constructed through a combination of both approaches.
An important advantage of Bayesian networks is that they allow modeling both state and temporal processes, which are inherent in the medical domain. In the state model, the functioning of a particular organ X at a specific moment t determines the outcome of laboratory tests and the presence or absence of symptoms. The set of associated symptoms, abnormal signs or conditions are indicators for the dysfunction of one or more body organs and can be used to establish a diagnosis. The relation is not necessarily directly causal, and it is possible that an underlying problem or risk factor such as existing diseases, age and genetics explains the association. Moreover, treatment decisions affect the status of organ functioning, aiming to reduce risks or further deterioration. The schematic representation of these relationships is depicted in Figure 3a, where the arrows indicate the cause-effect direction.
The health monitoring of a patient usually includes observations made at different time points 1, … , T. Taking into account the functioning of X at the previous observation t − 1 allows the medical practitioners to observe relative change in order to establish a diagnosis at t, and if needed, adjust treatment. Furthermore, given the history and the current status, one can attempt to predict the development at the next time point t + 1. This chain of temporal dependencies is depicted in Figure 3b.
Figure 3: Modeling the development of a syndrome. These systems are discussed in section 3.5.
We followed these state and dynamic modeling principles in developing the disease-specific clinical models for COPD and preeclampsia, embedded in the smart care-assistants. [16], [17]
Personalization
For some diseases, patients may be divided in groups sharing particular characteristics. For example, which symptoms are likely may differ between patients, but this information is usually not readily available from patient records. Therefore, it may be useful to adapt the interpretation model to individual patients. Personalization then entails model updating, sometimes also referred to as online learning, using the initial model as a prior and the data gathered from this patient as new data to change the probability parameters. The interpretation model learns which relations are different for this particular patient compared to the group for which the model was constructed. Clearly, model updating is only useful when patients use the system for an extended period of time.
14.3.3 Decision-making support
The success of a smart care-assistant depends for a large part on correct clinical decision-making. This requires a mapping from data and model outcomes to choices. Since we focus on Bayesian networks as prediction models, it appears natural to use influence diagrams (see e.g. [18]), which can be seen as an extension of Bayesian networks with decisions.
An influence diagram is a graph G = <N, A> with N a set of nodes partitioned into chance nodes, decision nodes and value nodes, and A as a set of arcs denoting dependencies. The probabilistic semantics of a network without decisions is the same as for a Bayesian network. The value nodes assign utilities, cost or benefit, to states of the parent variables. Decision nodes can have arrows to all types of nodes, indicating influence on a random variable or on a value node, or precedence on further decisions. Techniques for solving influence diagrams try to find the optimal policy, that is, the decisions such that the expected utility is maximized. One of the advantages of using influence diagrams stems from their graphical nature, which facilitates interpretation. In a clinical context this is an important feature because decisions, especially when made automatically, should be traceable.
Preferences are encoded in an influence diagram via utilities, which means that the value of possible outcomes should be established. Determining utilities is crucial for correct decision-making, but is not a trivial task, which, in a healthcare context, was noted already by Torrance and Feeny in 1989: “Whatever method is used to elicit utilities […] We cannot overemphasize the importance of care and precision in preparing, testing and using these instruments”.[19] For mHealth, this poses a problem because ideally the possible outcomes are judged by the individual users. Due to the difficult and time-consuming nature of the elicitation process, it does not appear feasible to do so. Instead, decision policies would have to be constructed either from the preferences of a few characteristic patients, or by healthcare professionals. The latter seems reasonable for clinical decisions on treatment, however, decisions on what information to present and in what fashion should be made based on patient preferences. Luckily, the latter preferences are likely to be easier to elicit and to adjust over time, as they do not directly impact health status.
Personalization of various parts of the smart care-assistant influences decision-making. For instance, the decision to register data can be left to the patient. This is useful as it reduces the burden on patients to provide data regularly, but requires a higher degree of self-management. Whether this is a realistic option will depend on the domain and the type of patients that will be using the system. Further, personalizing the models, as discussed above, has a direct impact on automatic decision making by changing the probability parameters.
14.3.4 User interaction
Besides purely clinical decisions, for example about treatment or medication dosage, the system or system developer also has to make decisions about what kind of information to present. A care assistant can help patients with understanding their disease by presenting, in a meaningful way, the data that have been gathered. This implies visualization of history, trends and previous clinical decisions; possibly in different formats – like easily understandable graphics or more detailed data views – depending on the user’s preference. This is a direct way to support patients to make their own informed decisions, which is an important goal of self-management. Similarly, prognostics can be supported by providing the interpretation and predictions from the model.
An important part of supporting decisions is making the right knowledge available, but if the decisions are complicated, a smart care-assistant might provide more direct guidance on the appropriate course of action. This leads to output patterns ranging from only presenting information, as described above, to automated decisions to generate an alert directly to a care provider. Depending on the domain, it may be appropriate to let the system make autonomous decisions, for example in life-threatening situations. However, self-management should be supported in other domains by informing the patient of appropriate actions based on the available data, or by advising to take a particular action, but leaving the final decision to the patient. For example, it is a design choice whether to advise the patient to contact a care provider when a deterioration of health has been predicted, or whether an automatic alert will be generated. A typical example of the kind of advice that could be provided is “Please repeat the measurements in 6 hours”.
The interpretation capabilities of the assistant also require an appropriate interface for displaying and communicating the current health status to the user. It is expected, for example, that most well-educated people would in general be interested in obtaining insight into their health condition, possibly in terms of an explanation of how well the condition is under control. If the system is used long enough, it should also be possible to learn when it is opportune to directly contact a caregiver, or when it should only be advised. This depends for a large part on the severity of the situation and the self-management capability of the patient. The type of feedback that is provided can be customized to obtain a comfortable and safe level of self-management in consultation with both the patient and caregivers.
14.3.5 Application of personal smart care-assistants
We next present the application of the smart care-assistant in the context of two disorders : (i) chronic obstructive pulmonary disease (COPD), a lung disease; and (ii) preeclampsia, a pregnancy-related disorder. Each assistant consists of:
- – a mobile application (app) for (i) automatic and manual collection of patient data, (ii) automated interpretation of the data based on an embedded Bayesian network model and (iii) reporting the patient’s status to the patient and the healthcare team; and
- – a number of sensors and tests to measure the patient’s condition.
In the following we report mainly on the intelligent components of the two care-assistants in terms of the models built and their capabilities obtained from evaluation on patient data and from pilot studies. In particular, we evaluated the abilities of the models for (i) diagnosis – accurate detection of the clinical condition, and (ii) prognosis – predicting the development of the disorder to allow timely intervention. The diagnostic ability of the data interpretation was tested with ROC-analysis using the area under the curve (AUC) as a standard performance measure in clinical practice. In addition, we conducted pilot studies for both disorders to ascertain the technical feasibility of the systems and to obtain early feedback from end users on usability. The general setup of the pilot studies was as follows: patients were provided with a smart care-assistant (smartphone, app and sensors) that was used in a home setting for a period of two to four weeks. Every day at a predefined time, the assistant gave an alert to the patient to take measurements and fill-in a questionnaire. At the end of the evaluation period every user filled in a feedback form and took part in an interview with the case manager of the pilot study. The feedback concerned various usability issues such as ease of use, user-friendliness, technical operation and overall usefulness of the system.
14.3.5.1 Home monitoring of COPD patients
COPD is a progressive lung disease, currently affecting some 64 million people worldwide. It is one of the major chronic diseases in terms of both morbidity and mortality. The main cause of COPD is exposure to tobacco smoke. COPD is currently not curable, but treatment does reduce the burden considerably.
Exacerbations – episodes of acute deterioration of the patient’s condition, usually caused by an airway infection – have a profound impact on patient well-being and on health-care costs. Patients with frequent exacerbations usually have faster disease progression, which makes exacerbation prevention a particularly relevant goal. Additionally, a faster treatment response to exacerbations appears to lead to better recovery. The state of the respiratory system is observable via symptoms including dyspnea, productive cough and decreased activity due to breathlessness and a number of physiological signs such as the forced expiratory volume in one second (FEV1) and blood oxygen saturation.
Our COPD care-assistant consists of the AERIAL app and two Bluetooth-enabled sensors – a pulse-oximeter and a micro-spirometer. We studied different types of Bayesian network models – constructed by hand in close cooperation with lung physicians and by using machine-learning techniques – for a timely prediction of exacerbations. The model input consisted of self-reported symptoms using a questionnaire on the phone and the measurements from the sensors.
We performed preliminary evaluation studies with both retrospective data and pilot scale prospective data.[16] The model based on an expert opinion yielded an AUC of 0.97 for detecting exacerbation events. This indicates that the model can detect exacerbations as they are happening, which is a useful baseline and required for further model development. In order to predict exacerbations, we constructed temporal models, based on the time series data resulting from monitoring patients with our system. It turns out that it is possible to predict at least part of the exacerbations a day in advance with the best model resulting in an AUC of 0.82.[20] The results from the pilot studies with respect to the usability goal revealed that the patients’ impression of the system after using it is fairly positive. This early feedback from actual COPD patients is important for these kinds of systems because acceptance is often a concern.
The next step involves assigning an advice to the predictions. Relevant decisions for COPD care include rescheduling monitoring because of increased risk, dosage adjustment of bronchodilator drugs and contacting healthcare professionals. A decision analysis will be carried out with family doctors as well as lung specialists to establish what decisions the system should make given a particular clinical situation.
14.3.5.2 Home monitoring of high-risk pregnant women
Preeclampsia is a pregnancy-related syndrome associated with a high blood pressure (> 140/90 mm Hg) and a leakage of protein into urine (so-called proteinuria). It is the most important cause of death among pregnant women and a leading cause of fetal complications such as low birth weight. As a pregnancy-related condition, the only way to cure preeclampsia is to deliver the baby. To allow a timely detection of preeclampsia in the current clinical practice, frequent outpatient visits are required, leading to high pressure on the pregnant woman as well as on the healthcare centers. In addition, pregnant women are relative young with an interest in applying modern technology to the management of their condition. Despite these facts, smart solutions for supporting self-monitoring during pregnancy are lacking. The eMomCare smart care-assistant we developed is the first of its kind.
The care-assistant consists of the eMomCare app, a Bluetooth-enabled blood pressure meter and automated analyzer of urine reagent strips to measure protein-to-creatinine ratio.[12] Given the dynamic nature of a pregnancy, we manually developed a temporal Bayesian network model for preeclampsia in cooperation with gynecologists. The model, embedded in the eMomCare app, includes 13 risk factors and signs such as blood pressure measured at 10 time points (routine check-ups) during the pregnancy to estimate the current risk for preeclampsia and give a prognosis of the risk until the end of the pregnancy. The model has been evaluated with retrospective pregnancy data and showed high accuracy at the week of the diagnosis – AUCs ranged between 0.80 and 0.99 for the weeks 28–38 of the pregnancy. In addition, in 60% of the preeclamptic women, the model was capable of making the prognosis for the syndrome at least 4 weeks before the diagnosis was actually made (with 10% false positive rate). For the remaining patients, this early prediction was not possible mostly due to inaccurate or missing measurements in the data.
The pilot study we conducted with the eMomCare system included 6 pregnant women from Maastricht University Medical Center. The overall feedback obtained was positive in terms of relatively easy system operation (smartphone, app and sensors). Some technical problems and design issues encountered during the pilot have been used to improve the system. The participants (completely) agreed that the eMomCare system has added a value to the usual care, and they are willing to use the system very often, showing the potential of the system for use in practice.
14.4 Discussion
Given the framework for smart care-assistants and the first results from applying care-assistants in practice, what remains is to examine their overall clinical impact and to make a comparison with existing solutions. Our experience with pilot studies shows that it is challenging to obtain reliable and objective outcomes in a home environment. A careful design, organization and control throughout the study are critical for the successful evaluation of mHealth systems. Training programs (e-Learning) for patients and caregivers are also expected to alleviate certain problems concerning, for example, the technical operation of the system. Such programs can facilitate the mind-shift towards sharing responsibilities between the patient and the caregiver during the care process.
Acceptance of smart care-assistants depends on the extent of integration within existing healthcare structures. It is important to consider at an early stage of development how the system will be used: patient initiative (complete self-management) or prescribed and monitored by a physician (shared-care). In the latter case, integration of the smart care-assistant and healthcare information systems is essential, allowing for communication between the patient and caregiver in the context of the patient-care process.
Establishing the public’s trust in smart care-assistants for personalized health-care management will be key for ensuring the public’s acceptance of this new medical technology. Current concerns relate to the privacy and security of collecting and transmitting patient data. These are valid concerns that should be addressed, for example, by using electronic signatures, authentication and multi-application smart card solutions; access to healthcare data should be under control of the patient in negotiation with caregivers. Furthermore, a smart care-assistant can, in principle, function without access to remote data, and so whether or not the patient’s data are transmitted can be controlled by the patient.
Given the current capabilities of the smart care-assistants, we expect that supporting people’s decisions about their health will stimulate a more pro-active attitude toward health concerns, thus increasing patient involvement in the care process and reducing the burden on the healthcare system. This can help detect health deterioration at early stages, thus allowing timely treatment interventions and preventing hospitalizations. As a result, costs and work pressure within healthcare centers are to be reduced while maintaining or even improving health outcomes. Furthermore, smart care-assistants can play an educational role. Through better information provision to patients, the system can assist in teaching them to cope with their diseases.
Compared to existing systems, improvements consist of the use of clinically sensible data interpretation models and the improved availability of support. Many existing systems do not offer autonomous operation, relying on interpretation by care providers, or are purely technology-driven and do not provide clinically validated advice. This even appears to be the case in recent systems, e.g. [21] on using HIT for diabetes management where mobile technology is only used as an additional communication channel to facilitate self-management support. Furthermore, personalization appears to be an important aspect of data interpretation that is gaining attention in the literature. For instance in [22], where a personalized system for lifestyle advice is described.
14.5 Conclusion
In conclusion, the rapid penetration of mobile devices in our daily lives and the deployment of mHealth systems is expected to provide greater access to health care to larger segments of the population (see expected number of mHealth users in 2017 in Figure 2). This will be beneficial for the healthcare system as a whole and for its ability to provide qualitative, safe and cost-effective care through improved early risk assessment, with appropriate referral and consultation among providers of basic and specialty care.
Acknowledgement
This work has been partially supported by the Dutch organizations STW, ZonMw and STITPRO.
References
Force ET. e-Health: Redesigning health in Europe for 2020 [Internet]; 2012. [cited 10 December 2014] Available from: http://www.epractice.eu/en/library/5362646
Blackburn M, Pitts J, Walden G, Bilbray B, Burgess M, Gingrey P. US Congres Letter to FDA and FCC on mHealth; 2012. http://blackburn.house.gov/uploadedfiles/letter_from_congress_to_fda_and_fcc_-_3apr2012.pdf
Research2guidance. Mobile Health Market Report 2013–2017: The commercialization of mHealth applications [Internet]; 2013;(3). [cited 10 Dec 2014]. Available from: http://mhealtheconomics.com/mhealth-developer-economics-report
Istepanian RSH, Jovanov E, Zhang YT. M-Health: Beyond seamless mobility for global wireless healthcare connectivity. IEEE Trans Information Technology in Biomedicine 2004;8(4):405–412.
Buchanan BG, Shortliffe EH. Rule-based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Reading: Addison-Wesley; 1984.
Miller RA, Jr HEP, Myers JD. INTERNIST-1: An Experimental Computer-Based Diagnostic Consultant for General Internal Medicine. New England Journal of Medicine 1982;307(8):468–476.
Kumar S, Nilsen W, Pavel M, Srivastava M. Mobile health: revolutionizing healthcare through transdisciplinary research. IEEE Computer 2013;46(1):28–35.
Rubin MA, Bateman K, Donnelly S, Stoddard GJ, Stevenson K, Gardner RM, et al. Use of a Personal Digital Assistant for Managing Antibiotic Prescribing for Outpatient Respiratory Tract Infections in Rural Communities. J Am Med Inform Assoc 2006;13:627–634.
Lee NJ, Bakken S. Development of a prototype personal digital assistant-decision support system for the management of adult obesity. Int J Med Inform 2007;76:S281–S292.
McLean S, Nurmatov U, Liu J, Pagliari C, Car J, Sheikh A. Telehealthcare for chronic obstructive pulmonary disease. Cochrane Database of Systematic Reviews 2011;(7).
Mudanyali O, Dimitrov S, Sikora U, Padmanabhan S, Navruz I, Ozcan A. Integrated rapiddi-agnostic-test reader platform on a cellphone. Lab on Chip 2012;12(15):2678–2686.
Velikova M, Lucas PJF, Smeets R, van Scheltinga JT. Fully-automated interpretation of biochemical tests for decision support by smartphones; 25th IEEE International Symposium on Computer-Based Medical Systems (CMBS 2012), Rome, Italy, 20–22 June, 2012, pp. 320–325.
Chen M, Gonzalez S, Vasilakos A, Cao H, Leung V. Body Area Networks: a survey. Mobile Netw Appl 2011;16:171–193.
Clifford G, Clifton D. Wireless technology in disease management and medicine. Annual Review of Medicine 2012;63:479–492.
Pearl J. Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan Kaufmann; 1988.
van der Heijden M, Lucas PJF, Lijnse B, Heijdra Y, Schermer T. An autonomous mobile system for the management of COPD. Journal of Biomedical Informatics 2013;46:458–469.
Velikova M, van Scheltinga JT, Lucas PJF, Spaanderman M. Exploiting causal functional relationships in Bayesian network modelling for personalised healthcare. International Journal of Approximate Reasoning 2014;55:59–73.
Shachter R. Evaluating Influence Diagrams. Operations Research 1986;34(6):871–882.
Torrance G, Feeny D. Utilities and quality-adjusted life years. International Journal of Technology Assessment in Health Care 1989;5:559–575.
van der Heijden M, Velikova M, Lucas PJF. Learning Bayesian networks for clinical time series analysis. Journal of Biomedical Informatics 2014; 48:94–105.
Nundy S, Lu C, Hogan P, Mishra A, Peek M. Using Patient-Generated Health Data From Mobile Technologies for Diabetes Self-Management Support. Journal of Diabetes Science and Technology 2014;8(1):74–82.
Hu B, Naseer A, Fukuda K. A personalised lifestyle advisory system. In: Healthcom 2012: IEEE 14th International Conference on e-Health Networking, Applications and Services; 2012.