By Dr Fergus Fairmichael.
Dr O’Hanlon is EMIS Group’s Chief Medical Officer. He started with EMIS in 2006 as Clinical Design Director and was responsible for the clinical architecture of EMIS’s flagship product EMIS Web. Dr O’Hanlon trained at Cambridge and St Thomas’ Hospital, becoming a GP principal in 1994 and became Medical Director of LifeGard (Health Smartcards) between 1999 and 2000. He is also a Director of QResearch (a joint research collaboration between EMIS and the University of Nottingham).
Suitability of routine data for research purposes
Routinely collected data can be used for research purposes in certain cases, depending on the use case. In order to apply academic research to clinical practice, there needs to be an underlying understanding of how clinical practice works. If this understanding is not present then it is unlikely that people will change their behaviour.
EMIS has tried to take real world data and work with that, rather than creating projects where specific data need to be collected. For example, there has been a lot of hard work to keep QRISK (a cardiovascular risk scoring system) relevant and based on what is recorded in day to day general practice as opposed to asking clinicians to record new data to develop a different risk score, people won’t do it, they are busy people with only 10 minute consultations during which time they are more focussed on giving high quality clinical care than they are on feeding a high quality clinical research network. We have seen that when the data reflects day-to-day clinical practice, the project is more successful. The infrastructure to collect more bespoke data is there but clinicians would need to be incentivised and would require much closer collaboration with academics, who in turn would have to work very hard to help clinicians use the tools.
It is possible to modify practice to support research but this needs to be done in the right way. There are 2 ways to achieve this
1. Taking what is currently done and applying it
- •This approach does not necessarily meet the academic rigour that is often sought, this may be a draw back for academic researchers as their papers would not hit the top notch journals which would look for greater control of confounding factors in observational studies, which can be virtually impossible in every day practice
2. More interventional studies with increased structure
- •This is more difficult to recruit big numbers unless there are the right incentives, e.g. financial support for the additional data collection costs, unless it is going to support patient care or patient processes in their organisations
- •This approach requires significant collaboration between the academic research team and the clinical team to ensure tools are used correctly
It is important not to belittle the challenges of information governance. Sharing data across networks can become very intricate and complex due to concerns about privacy. In particular 3rd party references shared within free text can be an area of contention as these 3rd parties have not given consent. There are attempts to pseudo-anonymise this information and some systems have ways of masking this information from export but it is not highly used or standardised across different systems. A lot of Caldicott guardians don’t feel free text is appropriate.
Free text in GP systems often doesn’t contain a lot of the academic data because most of this is already highly codified. Data is often well structured in primary care for a combination of reasons:
- •Comes in through the pipeline, e.g. lab work,
- •Prescribed medicines are highly structured already
- •The use of templates that are clinically coded
There is a lot of structure in GP records that you don’t see elsewhere in the NHS. Primary care records are world leading, for example we can transfer data between different systems, and this can retain its context after transfer, we can file something from a different system and it will look the same as one created originally on that system. We are in a fortunate position to have the combination of primary care infrastructure and IT infrastructure that works well for both the patients and for the doctors.
Natural language processing
People may be looking for an excuse not to go through a structured record approach, and this may have resistance as it may create additional work. NLP could superficially offer a pathway towards this, in particular it may have a role for inbound information that does not have structure behind it. An example where this would be significantly useful is in mental health where, with the exception of the ICD10 diagnosis code, the record is mainly free text and un-coded data. In areas of medicine where dictaphones and secretaries rule there is a lot of potential for NLP where elements could be structured.
However, language can be very complicated and NLP has not solved all of these puzzles yet. There is a feeling that we are not in a position for it to be used safely for decision support at present. It may work in certain models for research purposes but it remains very challenging to work safely.
There is a risk balance that needs to be met, with the understanding that the system could potentially lead to harm and doing what can be done to prevent this. We do not want to become the first hole in the Swiss cheese model of system failure.
Ultimately trust in the system is crucial. If the system is providing inappropriate or poor quality alerts then people will begin to ignore them, and this would subsequently lead to the system becoming unsafe. The credibility of the system is also commercially important as it would turn people away from the product.
At present the model that works well for general practice is to codify what can be codified and keep free text as it is.
Due to the nature of primary care there is often a lack of opportunity to record outcome measurement. Often patients will see their GP and if treatment is successful they do not return. For longer term conditions, there is more opportunity to capture the longer outcomes as patients will likely return periodically for review, however these conditions often lack an end point at which outcomes would traditionally be recorded.
For this area to develop there is a need to think about how the system could better capture or track outcomes. These may be better collected by the patient themselves rather than burdening primary care. There would be no purpose in patients attending their GP merely to inform them that they are feeling well. GPs would not have the time or resources for that. The best place to capture outcomes would be from the patient.
There has been a lot of work into wearable health technologies, applications, Personal Health Records etc. however this field remains in its infancy. The technology is improving and some may argue it is in place to enable the recording of outcomes, but applying it to healthcare is more difficult.
Currently, patients are comfortable browsing for information but we are yet to go through the change piece that other industries have seen, for example banking where people have really engaged significantly with technology and online services. People may currently feel comfortable booking appointments or ordering repeat medications, but they are not yet ready to use online services for sharing individual information.
There is a feeling that this will become increasingly accepted. However there are concerns regarding the cost implications of providing such services. The cost of creating and maintaining online health personas may be quite significant, particularly when taking into consideration the necessary privacy requirements. There is an expectation that services such as these may help to save money in the long run but there is not much evidence to support this yet and it is a large investment and undertaking. There are a lot of start-up organisations that are doing some exciting things in this field, however at present they do not have the scalability and the data they collect ends up in an unconnected silo.
The issue of interoperability is multifaceted. On one hand it could be considered that EHR suppliers have not made this particularly easy, but on the other hand there needs to be the consideration that this is not their primary task and that they are not there to supply data for free. There are significant costs incurred from providing data sharing from host systems. It is also technically very difficult to do with scale, whilst ensuring that availability of service is not compromised. The main purpose of the EHR remains the provision of an uninterrupted clinical records service.
There are further issues surrounding the data itself. Researchers do not often get the data they expect because they do not understand what primary care does. They need to understand the model of consultation and model of data entry in primary care in order to understand what you get out. Data out is a direct reflection of what is put in.
The biggest issue is that the data does not belong to the EHR providers. In this respect they merely act as a data processor. The data controller in this instance is the GP practice or the health trust. The data controller is key as it is their responsibility that this data is held and used responsibly. They have a duty to data subjects, which is the patient. It is clear that nobody owns the data.
Over the next five years we will see some change in this area, as EU guidance on information governance comes into effect and will make the patient far more in control of their own data.
Another approach is that of ‘data safe havens’ such as that provided by UCL. These provide a service where data can be stored and analysed within the security of one system. In many respects the question of interoperability should be focussed around these safe havens, how they are set up, how they apply information governance and how well they understand the data provided by clinicians.
The role of decision support
Decision directive tools do not work, but decision supportive tools would be more likely to be successful and more acceptable to physicians. Doctors don’t go to university to be told by a computer what to do. The history of electronic records is littered with decision support tools, but the question this could be compared to is “would you use you satnav to tell you how to drive to work every day?” which is the equivalent of “if you see someone with conjunctivitis, do you need guidance to give fucilthalmic?”. Experience has told us that physicians do not want the paperclip popping up giving this information. What is needed is something with more understanding, that can help guide more complicated cases, for example patients with complex comorbidities with social problems. These are the biggest burden on resources and the source of more frequent mistakes as physicians may not have considered all of the complexities. These should not just be solely decision support but should help structure a care plan that involves the MDT, which would lead to better clinical outcomes.