This is a guest post from Think GP (Reed Medical Education) – With latest research showing that the majority of Australians seek medical advice from the internet, Dr Simon Cowap wonders if his job is on its way out. Can AI (artificial intelligence) take over doctors jobs?
Up until recently, automation has tended to result in the loss of lower-skilled or blue collar jobs. More highly skilled ‘white collar’ occupations were thought of as safe from being replaced by robots. But now studies suggest that increasing deployment of computers with advanced artificial intelligence (AI) capabilities mean this is no longer the case. A recent Fortune magazine article on white collar jobs robots are already doing included surgeons, anaesthetists and diagnosticians. IBM, the company whose supercomputer Deep Blue beat world chess champion Gary Kasparov all the way back in 1996, is now offering us Watson, a far more powerful computer that they are ‘sending to medical school’. It must be said that Watson has not graduated quite yet, but when it does IBM says it will work like this.
“First, the physician might describe symptoms and other related factors to the system. Watson can then identify the key pieces of information and mine the patient’s data to find relevant facts about family history, current medications and other existing conditions. It combines this information with current findings from tests, and then forms and tests hypotheses by examining a variety of data sources — treatment guidelines, electronic medical record data and doctors’ and nurses’ notes, as well as peer-reviewed research and clinical studies. From here, Watson can provide potential treatment options and its confidence rating for each suggestion.”
Watson’s advantage over mere humans is that it can review a patient’s entire medical record (assuming it has been encoded) and access the entire content of PubMed, Medline and a range of databases and be continually updated.
In the end, I’m not sure how much or how soon systems like Watson will impact on the humble domain of office based general practice. My understanding is it’s mostly being used at this stage for quite specific data-intensive tasks like reviewing imaging and suggesting treatment options at major cancer centres like Sloan Kettering. A physician still needs to take the initial history and perform the examination. The vagueness, variety and inherent uncertainty of general practice presentations probably make them harder for AI to assess than malignancies with clearly defined parameters, treatment protocols and relevant research data.
I can well imagine computerised diagnostic assistants becoming more widespread before I retire – differential diagnosis generators such as Isabel are already quite popular – but I’m fairly confident I won’t be completely replaced in the remainder of my working life. For those just graduating, it may be a different story. Dealing with Dr Google is one thing, dealing with Dr Watson may be quite another!
In the meantime though, the key clinical task of diagnosis remains with us humans. As someone interested in education, I have an interest in how we improve our diagnostic skills. In another role, I’m also involved in assessing medical performance, including diagnostic skills. But while diagnosis is something we all do every day, we don’t spend much time analysing the process.
The model I absorbed, though it was never really spelled out, and is presumably the reason IBM called its computer Watson, is Sherlock Holmes’ method. It’s no surprise Conan Doyle trained as a doctor. We hunt for evidence, in the form of history, symptoms, and examination findings, supplemented by relevant investigations, until we can confidently identify a perpetrator. The brilliant clinician, in my medical school imagination, was the one who could recognise the right clues and link them in the right way. The plodder, like poor old human Watson, would either miss the relevant information; lack the knowledge base to understand its significance or the mental acuity to join the dots.
I’m not sure what guided the IBM programmers, but the essential elements in forming a diagnostic hypothesis do seem to be acquiring relevant information and matching that against a database containing descriptions of different conditions. As far as I can tell from observing myself and other people, there are two common ways we do this, and I think we often use both at the same time.
I suspect an awful lot of diagnoses, especially by experienced practitioners, is pattern recognition. Human beings are pretty good at this, it’s something we find easy and we’re not really even aware of doing it. Two days of myalgia lethargy coryza headache low grade temp =urti. 55 yr old smoker hypertension chest tightness on exertion radiating to jaw relieved by rest = IHD. Some patterns are obvious and emerge straight away. Other patterns are hidden; we might have to take a lot of history and get a lot of other information before they emerge.
Another common tool we use, and that appears a lot in educational material, is the algorithm – a branching tree of possibilities where key history, examination or lab findings direct us down ever narrower branches until we arrive at a specific diagnosis. For instance, in the patient presenting with dizziness we first want to know if it is light-headedness, pre-syncope or vertigo. If vertigo, if it is central or peripheral. If peripheral, if it is pure vertigo or if there are associated auditory symptoms. If pure vertigo, if it is triggered by neck movement – and so on. Eliciting relevant information at each step gives us a narrower range of possibilities. This approach is somewhat more laborious and requires conscious effort. It’s an example of what Daniel Kahnemann (Nobel prize winning psychologist of ‘Thinking Fast & Slow ‘ fame) calls ‘slow thinking’, as opposed to the ‘fast thinking’ of pattern recognition. Thinking back to my first day in casualty, I probably tried to use an algorithm on all my patients – with very little success.
Nowadays, I use a mix of these approaches. If a pattern is evident, that’s great. If not, we might use an algorithmic framework to guide further data collection until a pattern does emerge. Once it does, we tend to stop.
Patterns are great – they’re quick, easy and usually right. But not always. The danger of patterns is that having found one, we stop thinking of alternatives. Sometimes we mistake one pattern for another. We can be too committed to one pattern to notice that it’s changing over time. And sometimes we see patterns when they aren’t there – some GP patients just don’t fit into a neat picture. For all the terabytes of data at his disposal, Watson would crash trying to make sense of them.
There may yet be an X-factor in all this. Watson can crunch through immense amounts of data in the time it takes me to introduce myself. Much slower human clinicians have to get to relevant data very quickly. I think something called clinical intuition, more on which another time, probably kicks in extremely early and helps us focus our history from T0.
It’s a fascinating process. It’s amazing how often we get it right, but we do need to be aware that we don’t always. Keeping the old medical school habit of generating differential diagnoses and being prepared to dump the favourite in the light of new information will generally keep us out of trouble. As always, I’d be keen to hear your thoughts and experiences with diagnosis and its dilemmas. I won’t pass them on to IBM – we don’t want to help them do us out of a job just yet.
This article is a guest piece via Think GP, originaly published here.
Dr Simon Cowap MBBS (Hons), FRACGP
Simon is something of an accidental GP who likes to pretend he’s an artist trapped in a professional’s body. He dropped out of his first degree (arts) and went to London to play bass guitar in a band too musically challenged even for punks. Dropping back in to university, he subsequently also failed to complete a science degree and a Masters of philosophy. His remarkable lack of artistic success has been continued by the non-publication of his several novels. Somewhere along the line he did finish a medical degree. He still harbours dreams of literary success but his family have forbidden him to give up the day job.