Rise of the Machine in General Practice: Improving GP Decision-Making with Machine Learning

150116_main_entrance_big_brain_003_187610_001jpg-1

Rise of the Machine in General Practice: Improving GP Decision-Making with Machine Learning

Imagine a routine consultation with your GP. You tell them your symptoms, they listen, make a diagnosis, and perhaps write you a prescription or send you for tests. But how can you be sure that they’ve got the diagnosis right? Often, many of the symptoms of serious diseases such as cancer can be vague enough to mislead a GP into thinking they’re something far more common and innocuous.

Work by Dr Olga Kostopoulou, Reader in Medical Decision making at Imperial, funded by Cancer Research UK,  has shown that if GPs do not consider cancer as a possibility when the patient first presents, they are less likely to ask about cancer symptoms and less likely to make a cancer diagnosis in the future, sometimes with serious consequences. However, existing computerised decision support systems, which aim to trigger the GP to consider statistically rarer possibilities that might not immediately spring to mind, are not widely used as they do not fit into the flow of a consultation and are likely to be dismissed if the GP is not already thinking of cancer  

Brendan Delaney, Chair in Medical Informatics and Decision Making, is leading a CRUK-funded consortium that aims to transform this situation. The project brings together an interdisciplinary group of clinicians, epidemiologists, informaticians, data scientists, computer scientists, psychologists and experts in human computer interaction and design, across five academic institutions, and in collaboration with three industry partners; the breadth of the collaboration reflects the scope of the work involved.  

For a decision support system to work well it has to be easy to use and feel ‘friendly’ to both GP and patient

For a decision support system to work well it has to be easy to use and feel ‘friendly’ to both GP and patient. The ultimate aim is that the system should be running in real time during the consultation, so that if the patient wishes, they can participate in the diagnostic process. Design of this part of the system will be the domain of the psychologists and computer interaction experts, with iterative input from GPs and patients.

Brendan’s own interests lie in using informatics tools to link both knowledge generation and knowledge utilisation to routine healthcare IT systems. Although doctors currently have a coding system for entering symptoms during a consultation, only about ten percent of what actually takes place is recorded in this way, and the doctor’s own bias as to what they think the diagnosis is affects which information they choose to code   

The challenge is therefore to develop an unbiased coding system which captures far more information as a ‘Learning System’. If done with large numbers of patients, this information can then be used to drive the decision support system— the richer and larger the dataset, the better the system becomes, and the probability of a set of symptoms leading eventually to cancer can be integrated with the patient’s own medical history to create a highly personalised diagnostic tool. Initial information gathering will rely on recruiting hundreds of GPs willing to pump-prime the system, and Brendan is currently working with GP networks to find participants.  

The level of complexity in setting up such a database is considerable, especially as the eventual aim is to interface with any one of several national electronic health record systems. The underlying knowledge base is a little like a dictionary, with definitions held together in a particular way, known in computer science as an ontologyMachine learning based on the data will contribute to the ontology, and an additional computer science challenge will be to link all the systems together so that the central knowledge base will be available in real time via the internet..  

Taken together, the project is a fascinating blend of the highly esoteric and the intensely human: at one end, the computer scientists and informaticians can get their teeth into grappling with the artificial intelligence and data challenges, whilst at the other, doctors in surgeries around the country, and perhaps one day around the world, will acquire a tool with the potential to change the lives of their patients. 

Breaking through the language barrier: Brendan Delaney offers some tips for nurturing collaborations

"Collaboration is always good. You make huge progress when you’re pushed slightly out of your comfort zone—you start to consider things in slightly different ways which really moves your thinking along. Intellectual stimulation is why most people are in academia and you definitely get that if you fully engage with your collaborators. And prosaically, there are a lot of funding streams pushing discipline crossing to get to grips with global health problems, so it makes financial sense to engage. 

Multidisciplinary research is very much the future – bringing applied and basic science together means you can do so much. If you’ve not yet tried it, you need to know that there are people out there who have practical applications for problems that would give you intellectual pleasure to solve in your own field; you just have to go out and find them. And there’s money out there to fund you!"