Around 40 billion United States dollars per year is being spent on biomedical research in the hope of producing insights that will enable improved future healthcare. The core premise is that an improved understanding of biology and disease will inevitably lead to better medicines and offer increasingly targeted diagnosis and treatments based upon resolving patients into smaller and smaller similar groups (“stratified medicine”). Ultimately, this will make it possible to precisely tailor healthcare for each single individual (“personalized medicine”). Sadly, however, at least in the last decade or two, clinical medicine has not obviously benefited to the degree one might have anticipated, given the progress that has been achieved in basic and “translational” research. Many and various issues can be identified that feed into this lack of real-world impact, but one aspect that is increasingly discussed is the lack of effective connectivity between research and healthcare in terms of information and communication technologies (ICTs) and knowledge management and exploitation (bioinformatics and medical informatics), as emphasized in Figure 1.
Motivated by the need to address this problem, and as part of its mission to help “harness the Web to capture and unify genetic information that fundamentally impacts on a person's health and disease processes,” the GEN2PHEN project (http://www.gen2phen.org) has been working with a large number of international experts to explore the nature of the troublesome gap between research and healthcare informatics systems. Specifically, over the past 2 years, GEN2PHEN has facilitated online, e-mail, and telephone discussions of this topic between a large range of stakeholders, culminating in three physical meetings during 2011. The most recent meeting was an intense 2-day workshop in Brussels, where all the threads of the 2-year discussion were pulled together and critically appraised by experts from many relevant disciplines. A good understanding of many core issues emerged from this gathering, along with specific suggestions for priority actions needed to move the field forward.
Workshop delegates recognized that a research-to-healthcare ICT gap exists on several levels. First, research activities are undertaken by academics (particularly “bioinformaticians”) that thrive in a world of rapidly changing, open source, and often incompletely developed (and therefore necessarily “buggy”) software, whereas in healthcare, the ICT systems are typically purchased from commercial vendors (built and deployed by “medical informaticians”) and they need to be highly robust and validated. Second, the data of interest to researchers are, more often than not, large in scale, highly complex, and full of uncertainty (i.e., involving intermingled true, false, large, and small effects), providing an extensive and rich resource that can be mined for clues, correlations, and new hypotheses to guide future experimentation. In contrast, healthcare providers seek to work with unambiguous and straightforward information in conjunction with reliable decision-making principles so that clinical medicine can be provided consistently and optimally. Clearly then, research and healthcare are operating in quite different realms, and few people will have a deep understanding and experience of both, or the kind of training and thought processes that can easily switch from one world to the other.
Given the existence of such a wide and fundamental chasm between research and healthcare informatics, how might this problem be overcome? Numerous previous projects have set out to extend the scope of research informatics to make it more relevant to healthcare (often called “biomedical informatics”), whereas other initiatives have tried to make patient data more available for research use (e.g., biobanking). But such efforts cannot eliminate the basic problem that these two fields differ quite fundamentally, and are widely separated. Instead, we argue here that the space between the two domains can be bridged only by developing an additional, distinct, and currently poorly funded set of activities that might be called “knowledge engineering for health.” Knowledge engineering itself was first defined in 1983 as “an engineering discipline that involves integrating knowledge into computer systems in order to solve complex problems normally requiring a high level of human expertise” [Feigenbaum and McCorduck, 1983]. The role of engineering in tackling the research–healthcare divide is crucial, as this discipline involves methodologies and expertise that are very different to those of research or medicine. Only engineering is suited to the task of constructing the bridges necessary for effectively spanning the divide between research and healthcare knowledge systems. More specifically, knowledge engineering for health can and should now be used to “integrate and interpret information for individualized healthcare”—a principle, which has now spawned an “I4Health” network of 200 or more groups and consortia (http://www.i4health.eu) intent on promoting, developing, and realizing this goal.
The I4Health Concept
The outline I4Health concept, along with many remaining questions, challenges, and opportunities are summarized in Figure 2. The basic notion is that research and healthcare need to be effectively interconnected in terms of data and knowledge flow and use, in order to dramatically improve the effectiveness of clinical medicine. In particular, this coupling can only realistically be achieved by training and involving a new breed of experts concerned with knowledge engineering for health. By forming this engineering bridge between the other two realms, knowledge and utility will be able to pass productively in both directions. Research data and knowledge will be processed to identify clinically relevant advances, and these will be distilled, repurposed, and delivered into healthcare use. Conversely, experiences, procedures, and outcomes from the practice of medicine will be made suitably available for exploitation by the research community to aid in the understanding of biology and disease. Critically, this system can then begin to self-optimize in that real-world outcomes for patients treated by decisions that were based on latest research knowledge will be fed back to the research world, thereby enabling the latest state of knowledge to be refined and improved for use in treatment decisions for subsequent patients.
How such a system might be assembled and operate in practice, what exactly it might aim to achieve over what time frame, what types of information and standards will be required, how computer modeling can best contribute, and what problems or advantages might emerge for healthcare were just some of the many questions addressed at the Brussels workshop. The most critical perspectives emerging from that workshop shall now be described.
“I-Health 2011” Workshop
A 2-day workshop entitled “I-Health 2011” brought together over 50 interdisciplinary experts and stakeholders from 21 countries to discuss how research and healthcare ICT and knowledge management systems could be effectively integrated to enable improved and eventually fully personalized medicine. The workshop took place at the Thon Hotel Brussels City Centre, Belgium, on 3–4 October 2011, and it was jointly organized by the GEN2PHEN project and the INBIOMEDvision project (http://www.inbiomedvision.eu) with additional financial support and representation from the UK-India Education and Research Initiative. The agenda built upon insights gleaned from a 2-year analysis of the field by GEN2PHEN, with session topics and selected introductory speakers being chosen by the organizers (Anthony Brookes, Debasis Dash, and Ferran Sanz), guided by a highly experienced scientific committee (Søren Brunak, Norbert Graf, and Heinz Lemke) with additional background input (Iain Buchan and Carlos Díaz). All sessions began with one or two 10-min summary talks whose remit was to set the scene, followed by a few preplanned oral statements by experts of 3 min each, leading on to 1–2 hr moderated open discussions involving all attendees. The first day focused on identifying key types of research and healthcare data and their value, plus the informatics capabilities and strategies required for future optimal use of this information. The second day explored the necessary technologies and methods required to put these capabilities into practice.
Main Workshop Conclusions
A number of general conclusions came out of the workshop. One of the most apparent was that universal personalized medicine for most disorders is probably a long way off, and in parallel with this several delegates stated that there had been rather too much hyperbole regarding the personalized medicine in recent times. Of course, some examples of personalized medicine do exist, and more will emerge, but group-based “stratified” medicine (including the use of companion diagnostics) is more likely to impact healthcare in the near term future. All such progress will depend upon the actioning of things such as the I4Health concept. In this anticipated “I4Health future,” the expectation is not that medical doctors will be replaced by intelligent computers, but rather that since optimal medical practice will depend upon the integration of even more facts and factors than it does now, the proposed knowledge engineering for health systems will be needed by the clinicians to help them accomplish the necessary massive knowledge selection and integration tasks. Furthermore, it was widely recognized by the delegates that the devised systems will have to be sufficiently effective to convince doctors that they are worth using, otherwise no one will push for their purchase or installation in healthcare facilities. Particularly exciting was the widely held view that the system will also bring patients much more into the center of their own healthcare by having them add to and access information on themselves and the diseases that concern them, as well as by giving them a more direct role in influencing decision making about their treatments and healthcare policies in general.
Concerning the research-to-healthcare axis, it was noted that activities claiming to be “translational” often failed to actually produce any healthcare benefits or applications. Many such projects target the production of new “biomarkers” but these then fail to be used outside laboratory settings. There are more than 150,000 publications documenting thousands of claimed biomarkers, but fewer than 100 have been validated for routine clinical practice [Poste, 2011]. This failure may be due, at least in part, to lack of knowledge transmission across the research–healthcare gap. But beyond just clever listing and sharing of biomarker information, all clinically relevant knowledge from research needs to be identified, filtered, aggregated, integrated with patient-specific data, and provided in suitable formats for easy comprehension by healthcare providers and patients, and implementation within clinical decision support systems. For this to work well, and to do so in real or semi-real time, it cannot be fully automated. Instead, considerable human effort must also be built into the system, to provide system oversight, expert judgement, and guided data analysis to regulate what data and knowledge items pass through the pipeline. In addition, eventual use of the information must be dovetailed optimally into clinical workflows and hospital IT systems that seek to increasingly connect different aspects of the patient care cycle. Undoubtedly, this all places electronic healthcare records (EHRs) at the heart of the proposed system as part of the mechanism for knowledge entry, aggregation, and access, ideally operating on a global basis.
Concerning the healthcare-to-research axis, in equal measure to the above, delegates felt the system must be designed to make available to the research community the wealth of data generated by the real-world practice of medicine. This again focuses ones attention on the content and role of EHRs, how these should be organized and coordinated internationally, and the many salient data privacy and consent issues surrounding the use of patient-specific information (see further discussion below). In addition, human experiences, judgement, and insights also need to be somehow fed into the pipeline, to properly contextualize and manage the data flow from healthcare back to research. Currently, biobanks are beginning to fulfill some of these roles, but on a very small scale, and some early stage knowledge engineering projects have begun accessing many millions of EHRs for research purposes [Coloma et al., 2011]. Some are also arguing that (inter)national EHR banks should be established, to facilitate EHR data sharing.
Overall, it was felt that future healthcare needs can only be met in full, and at an affordable cost, by seriously addressing the issues explored at I-Health 2011. Presently, though, there is no coherent and recognized knowledge engineering discipline explicitly working on these challenges. Instead, people are working in isolation and in competition, often without any kind of engineering background, and often not fully aware of the potential of other technologies/domains. Simply undertaking more research or healthcare informatics work will not produce the essential systems as quickly as they are needed, and not negate the need for this nascent knowledge engineering for health discipline. The I4Health interdisciplinary community, therefore, aims to get some momentum into this field, to develop ideas and pilot projects, and to lobby funding agencies and decision makers for the resources needed to develop and apply knowledge engineering to the challenges at hand. Details of some of the issues we will need to consider, as elaborated at the recent workshop, are briefly presented below.
Detailed Workshop Considerations
Personalized and Stratified Medicine
In addition to the different time frames over which personalized and stratified medicines may become commonplace (as discussed above), workshop delegates explored the basic logic behind these two different approaches. Clearly, we need to do more than just consider everyone to have the same disease risk and treatment response profiles, but how do we do this? Discussions considered the difference between “lumpers versus splitters” when approaching this problem. Whether the objective is more individualized diagnosis, prognosis, or treatment, one strategy is to “split” the starting population only when there is evidential basis for doing so. Even one such split would generate a form of stratified medicine. The alternative is to adopt the starting position that we are all unique individuals, and then look for commonalities as reasons to “lump” individuals together. If no relevant commonality is found, then everyone continues to be considered as individuals, implying the need for personalized medicine. But delivering personalized medicine in this latter scenario will be challenging until a relatively complete understanding of the biomolecular basis of our uniqueness is established. In addition, this “lumper–splitter” debate fed through to discussions about computer modeling of disease (see section below).
These discussions then moved on to consider differences between population and individual level interpretation and utility of research findings. Statistical evidence showing that a certain factor correlates with a certain disease phenotype, or treatment response, provides nothing more or less than a correlation. Establishing actual causality on a population level can be difficult. But even if this is achieved and the effect size is sufficiently large to be clinically actionable, it is not appropriate to automatically apply this knowledge to any one individual. This is because the population level data only reveals the average effect, leaving a major question mark over the range of effect sizes in different individuals. Indeed, something that acts as a disease risk factor on a population level could even be protective in some individuals. This uncertainty must be addressed before a population level observation should be passed through the knowledge engineering for health system for use in the clinic. Whether this validation is best done by clinical trials, an analysis of the robustness of the effect in different populations, or some other means, remains an open question.
Many types and sources of data will need to be accessed for the proposed system to operate. Relevant categories of research data would include forms of omics information for complex and rare disease, multiscale and multimodal data integrated via systems biology approaches, time-dependent variables, biomarkers from translational studies, and so on. The challenge will be to access and distil the clinically useful knowledge and utility from these data in a timely fashion, and as part of this a clear distinction will need to be made between discovery grade observations, validated research findings, and results that have actually been used successfully in real-world clinical situations. Related to this is the question of knowing what degree of “trust” to place in any dataset. Some delegates suggested we may need one or more “database credit rating agencies” or at least some approved mechanism for certifying health information as fit for purpose. Such a mechanism could be analogous to the “CrossMark” initiative for identifying publisher-maintained versions of scholarly content (http://www.crossref.org/crossmark/index.html). In all of these cases, expert led data interpretation will be equally, or even more, important than the data themselves.
Detailed clinical phenotypes will need to be considered, both on a population level to enable research based upon the practice and outcomes of clinical medicine, and using patient-specific information to guide decision making in stratified and personalized medicine contexts. Such clinical data will probably come from EHRs, at least in part, but remote biosensors and telemedicine will also be useful in supplying the needed content, especially for real-time acquisition over extended durations. One could also imagine patients entering data about themselves online, and groups of patients contributing to wiki-type systems. Biobanks will probably become increasingly important as sources of both general research data and patient-specific information, and simple interfaces and procedures will be key to the success of data input routes involving humans rather than automated computers (e.g., for doctors working with EHRs).
In addition to data, the system will need to access suitable metadata details pertaining to each utilized dataset. Metadata provides important contextual data about the data, but far too often in current data sharing, it is missing or overly superficial. Depending upon the data in question and its intended use, the metadata might specify things such as provenance information, lists and roles of contributors, method descriptions, details of informed consent, use restrictions, data access and sharing limitations, and cross-references to other related datasets. Questions of data quality will be paramount, and this may need to be broken down into considerations such as stability, reliability, context, completeness, consistency, contemporariness, and correctness.
Coordination, Standardization, and Harmonization
If the future is to involve any kind of data integration, resource federation, orchestrated development, and global utility, then the creation and use of suitable standards must be prioritized. This theme, more than any other, kept re-emerging throughout the workshop. Both semantic and syntactic data standards will be needed, with most delegates expressing concern about the latter—arguing that it is far easier to parse and reformat the structure of data than it is to reliably guess the meaning of the data themselves. There was widespread agreement on the question of how the world should go about creating and evolving the necessary standards. Specifically, rather than either a top-down approach (in which a selected team designs and distributes solutions for everyone else to use—which may fail because the community may not be happy with, or “feel controlling ownership” of, the standards) or a bottom-up approach (where everyone freely innovates approaches—which may fail because of incompatibilities between the solutions that emerge), the preferred option would be a hybrid “middle out” model (where several trusted, leading groups are funded to work together and tasked with tapping in to community expertise, to discuss and develop standards).
Regardless of how standards get to be developed, those doing the work will have to contend with a rapidly changing landscape of data sources and data types. Some ideas were put forward as to how to deal with this, not least: emphasize “flexible” data standards that concentrate on fundamental characteristics of data rather than specific details (e.g., the MaGE-TAB exchange format [Rayner et al., 2006], semantic Web triples, the Observ-OM object model [Adamusiak et al., 2011]); stress harmonization as well as standardization (e.g., DataShaper [Fortier et al., 2010], PhenX [Hamilton et al., 2011], and the EU-funded Semantic Health NET and BioMedBridges projects); support the standards with regularly updated tools that enable the facile use of those standards (e.g., intelligent support for EHR data entry, enabling such things as International Classification of Diseases codes to be automatically extracted from free text, then proffered up or entered); generally improve and more widely use text mining strategies (to remove the distinction between coded and free text); and perhaps most important of all, extensively map between ontologies (to allow intelligent reasoning across sources and domains).
The superimposition of standards, especially new ones, onto infrastructures that have grown up by piecemeal development over many years or decades will not be easy, but must eventually happen—even if this means rebuilding some systems from scratch. Such initiatives will need to be very professionally managed, and adequately funded. Salient examples that were highlighted at the workshop included efforts toward national EHR systems in United Kingdom and Romania (which did not succeed) versus the single, comprehensive, multifunctional eHealth platform created and today running very well in Estonia. The outstanding success of the latter could be due to one or more factors, such as the relatively small size of the population it serves, the funding and planning that went in to it, its stepwise expansion from initial goals that were not overly ambitious, and the fact that they designed the system starting with something of a clean sheet.
The word “modeling” came up frequently at the workshop, but it soon became clear the people were using it to mean a range of different things. Many approaches to modeling were mentioned, variously seeking to recapitulate biology or provide predictive capabilities for use in research and/or clinical settings, targeting all sorts of single and multiple classes of data with varying degrees of integration, and ranging from molecular to anatomical and physiological scales. The common objective of all this modeling, with differing degrees of intended personalization or stratification in mind, was to generate a valid and useful “virtual patient.” At one extreme, progress in “model-guided medicine” was put forward by the computer-assisted radiology and surgery community to show how a carefully chosen set of 10–20 informational components can be integrated by probabilistic graphical modeling to yield useful clinical decision support in contemporary healthcare. This approach was intended to very much reflect the way doctors actually think, and be tailored to specific “domains of discourse” (viewpoints of interest). Their approach provides a means to computationally support stratified medicine today, but as it typically depends upon the predetermination of correlations between factors observed in groups of individuals, it may never alone be suitable for underpinning fully personalized medicine. More ambitiously, pilot work was presented that is to be used as the basis for a major “IT Future of Medicine” FET Flagship EU Funding bid. Here, up to 4,000 components are being incorporated in a mechanistic framework model targeting the molecular scale for predicting optimal cancer therapy, with the suggestion that this approach could be extended in the foreseeable future to all diseases and plug directly into healthcare decision support systems. This would be the ultimate and most suitable approach for providing completely personalized rather than stratified medicine, as it requires no background epidemiology data but instead entails modeling every person and disease situation as a completely unique “N = 1” system. This would then argue for “lumping” of uniquely modeled individuals with equivalent clinical predictions, rather than the “splitting” by probability approach, discussed above. Yet others argued for an emphasis upon macroscale anatomical and physiological evidence in the modeling, using emergent properties of each disease rather than its molecular underpinnings. Of course, hybrid approaches to the above can be imagined—for example, placing molecular modeling in the research realm and then using knowledge engineering for health approaches that leverage probabilistic modeling to integrate the outputs from the latest molecular/mechanistic models with many other sources of research knowledge and patient data to guide clinical decision making (Fig. 3).
Issues such as how to structure modeling advances so they can be used globally, the important need for process modeling of clinical activities, and the challenges of validating rapidly evolving models for clinical utility were touched upon but not discussed in depth.
Creating the Future
The overall sense from the workshop was that we had broken new ground by realizing and emphasizing the need for a new intermediary discipline (knowledge engineering for health) to bridge the gap between research and healthcare, and that we had made a good start in defining many of the factors relevant to developing effective activities in that domain. In time, this should lead to the ultimate delivery of healthcare, as illustrated in Figure 4. But it was equally clear that very much more needs to be done, not least wider and deeper analysis of the issues that we covered in addition to many others. For example, it was argued that increasingly similar computational and analytical services will be needed by researchers and diagnosticians—but what might that common workbench actually comprise? What should be the relative roles of traditional databasing, computational workflows, and the semantic Web, and how modular can we make the approach to tooling in this field? How do we slice up the challenge in terms of its Web presence, and precisely what functionality should each dedicated knowledge gateway/portal aim to provide?
Furthermore, there are myriad issues surrounding social, economic, ethical, legal, and educational dimensions that we did not attempt to cover in our “informatics”-focused workshop, even though many principles from those other realms should probably be “hard coded” into future IT systems. And as part of those societal dimensions, how might the public be suitably educated about the proposed system, and how might they react to the increasing central role of computationally managed knowledge in healthcare. Here, we might get some surprises—for example, individual patients often seem far less bothered about consent and privacy than one would expect from most ethics discussions, and perhaps the real risks of compromising patient privacy are less than some policy makers might fear. However, following the recent announcement from the UK government that their intention is to make patient data increasingly available to drug companies [Nature, 2011], there was a significant negative response from some patient advocacy groups.
A multitude of dimensions remain to be discussed and explored by many stakeholders, and as part of that, some priorities and strategies need global agreement. Nevertheless, we hope that the findings of the workshop reported here, the extensive work that led up to it, and the work that will be done by the I4Health network going forward will provide a useful contribution to this exciting and important field.
Particular recognition must be given to the attendees of the I-Health 2011 workshop whose input was critical in helping the authors generate and refine the principles and concepts elaborated in this report. The I-Health 2011 workshop participants (excluding authors) were as follows: Imad Abugessaisa, Anurag Agrawal, Anni Ahonen-Bishopp, John Apps, David Atlan, Myles Axton, Joel Bacquet, Stephane Ballereau, Celia Boyer, Peter Coveney, Andrew Devereau, Denise Downs, Monica Florea, Robert Free, Gerard Freriks, Juan M García-Gómez, Peter Ghazal, Abel Gonzalez-Perez, Allan Hanbury, Robert Hastings, Martin Hofmann-Apitius, Nick Holden, Mark Hoogendoorn, Maj Hulten, Traian Ionescu, Erkki Leego, Hans Lehrach, Yves Moreau, Henning Mueller, Juha Muilu, Arijit Mukhopadhyay, José Luis Oliveira, Corrado Priami, Michael Rigby, Peter Robinson, Michal Rosen-Zvi, Simona Rossi, Maria Saarela, Jasper Saris, René Schippers, Amnon Shabo, Nour Shublaq, Jonathan Tedds, Mihaela Ulieru, and Mauno Vihinen.