Ten challenges

It’s difficult to frame healthcare knowledge in computable format because the process of clinical decision making is so complicated and precise (and yet sometimes messy or tentative). Every step must be broken down, with each term clearly defined to avoid confusion and mistakes. A clinical informatician explains: “Take for example, guidelines for treating rheumatoid arthritis. There is no such thing as a patient who simply has rheumatoid arthritis. There are probably four or five sub-types of patients. 

“What if the patient has had a bone marrow transplant? Do you mean patients with rheumatoid arthritis who’ve already been treated for 10 years with methotrexate? Teenagers with rheumatoid arthritis need to be treated differently to someone who is over 60 with impaired renal function. If you are going to provide decision support, based on guidelines, you must be able to map each pathway, covering each sub-type. Each pathway should define the right drug, the right dosage. Should it be injected into the joint? Is a tablet best?”

“We’ve got all the technical power in computing to develop highly sophisticated decision support. However, it will only work well if the meaning and implications of differences around conditions and disease are precise expressed and standardised.”

Traditional healthcare guidance is typically couched in language with rather vague wording that clinicians are able to interpret. Computer-driven decision support needs knowledge to be presented in a much more precise and unambiguous form. A guidance expert asks: “Often the knowledge can justify that clear instruction. But whose job is it to go that extra step and develop a more directive pathway? Does that belong to organisations such as NICE or somewhere else?”

“A classic example of defining language more clearly is ‘severity’. Guidelines will say ‘depending on severity’, try this drug or that drug. But severe is an idiosyncratic term that people define in slightly different ways. It can be made more precise on a sliding scale – for example, if x, y and z apply, then, in that case, severity equals 1.” Where medical knowledge does allow more precision, then it could be made instructional, converted into algorithms and thence into computer-driven decision support. Of course, not everything can be made algorithmic, so sometimes advice can only state; ‘Try this drug and, if doesn’t work, then try this drug’.”

Much content in electronic patient records (EPRs) is still written in free text. The challenge here is to structure and code this material so that decision support can interpret the patient record and set out pathways for an individual patient by selectively combining clinical knowledge with the EPR data.

Understandably, many clinicians feel burdened by typing in patient data. However, if it is incomplete or inaccurate, then the decision-support based on incomplete data will be unhelpful. Additionally, some clinicians may not respond carefully enough to decision-support prompts. For example, in dropdown menus, the top term – say eczema – is typically chosen more often than a lower one like, say, psoriasis. Other clinicians may reject decision-support as questioning their expertise. However, a doctor explains: “We should think of decision support as a form of continuing professional development. It is almost impossible to keep up with all the new research in any given clinical field. Yet it’s vital to have better access to it, because what we learned in medical school may now be frowned upon now or apply only in certain cases.”

Computable knowledge is the raw material from which decision-support is created. It needs to be in a standardised format. Accessibility demands that the knowledge is held in common in an open e-library that is well-indexed.

Recommendations using open standards from research, regulators and guidance-creators are the raw material of decision support, but they need to be regularly updated as the science advances. So, for example, recommendations for treating COVID-19 have changed frequently. Computer-driven decision support is an excellent way to bring these changes into clinical practice rapidly. However, it requires a commitment and capacity to frequently update both diagnostic and care pathways as well as their digital equivalents. The computable healthcare library in the Cloud will be a busy, labour-intensive place.

Building these decision-support systems means melding clinical and informatics expertise into relevant, workable, accessible advice. An informatics academic explains: “How do you create multidisciplinary teams that involve clinicians who must assess the evidence base to develop recommendations and determine which parts of the recommendation are computable and which are not? They need to come up with a sufficiently precise definition of a recommendation so that it can be made computable. All of this need to be a highly collaborative process between clinicians and knowledge engineers. Otherwise, there is a wall between them: the clinicians throw something over the wall to the engineers who then make all sorts of assumptions about what it means and what it doesn't mean, which is where these things always go wrong.”

Another expert in this field states: “There is a real need for a multidisciplinary approach where participants trust the others’ experience and expertise. Digital practitioners and informaticians should work alongside clinicians and public health advisers. They must communicate to understand the opportunities, the language barriers between them, the political drivers, and what people see as important. I've not yet seen a true multidisciplinary conversation where there's a full shared understanding of this space.”

Computable biomedical knowledge is still in its infancy. Should it be done in-house or outsourced? Should the task be left to the private sector? A digital healthcare expert explains: “Big players in the digital space such as Facebook and Microsoft, are all looking at healthcare. But, even if Google, Apple or Microsoft take this on, the concept will only work if there is standardisation around the data and care pathways. They could build a brilliant algorithm that works for a group of patients, but it won't work safely and to a high quality unless all of the terms in their pathway have been used correctly.”

Decision-support is like a dance between the clinician and the support tool. Each must stay in step with the other or the clinician is wrong-footed, rather than guided. The process is a series of stages, beginning with what’s known about the patient, and moving on to what tests to do, evaluating results and later suggesting possible diagnoses and treatments. An informatics expert explains: “Decision support must try to capture clinical decision-making patterns. By following the logic of real decision-making, the support speeds up and facilitates each step along the way with relevant knowledge and advice.”

Oversight by a trusted government agency will be required to guarantee safety and quality, regardless of who performs the tasks of formatting computable knowledge and developing new decision support systems. A healthcare informatician suggests: “We could consider something like the model used for the Medicines and Healthcare products Regulatory Agency.” The MHRA has licenced “Notified Bodies” that do the testing and certification of medical devices. We could have something like that to quality assure decision support tools.” Adherence with decision support advice also needs to be monitored anonymously at organisational or national level so that we can change guidance based on understanding of both expected and unexpected variations.

Health systems, like all organisations, struggle to lead change because they are restricted by their commitment to business as usual. There is no single stakeholder within government with overall responsibility for leading and implementing this complex initiative. Government leadership is required to set out a strategy that defines the goals and the roles of the many different parts of healthcare required to create and implement the strategy.

Planning and delivery will not only be inter-departmental (bringing together, for example, those responsible for standards and guidance setting), but also those charged with healthcare digital strategy, regulation, capital infrastructure and clinical delivery in primary, community, social and secondary care. It will also be multi-disciplinary, requiring the collaboration of, for example, informaticians and clinicians.

Health is a devolved policy in the UK. So, for example, NICE provides national guidance and advice to improve health and social care in England and Wales. But the Scottish Intercollegiate Guidelines Network plays this role north of the border. Some conditions do not have published guidance from an authorised UK body so practitioners rely upon American or European guidelines. We will need to clarify the rules for how Britain’s various guidance standards should be used for decision support. Ultimately, a global framework in also required, within which jurisdictional systems can operate.