There is a
crisis in the British Accident & Emergency (A&E) system at the moment, the causes are are a "perfect storm" of previous decisions, mainly around reduction in funding in the end to end care system overall, and the usual crop of winter bugs, especially ones that over-impact the ever-increasing numbers of elderly. But this week I discovered a possible self-inflicted cause - their own expert
self-diagnosis system on the NHS website will erroneously send people who can safely medicate at home to A&E.
(As I understand it is the same system that the medically unqualified telephone helpers on the 111 system use for diagnosis, with
similar results, it would seem).
TL;DR- I would suggest the NHS expert system is too poor a model of symptoms and then defaults to the lowest risk position - i.e. to send anybody with the slightest probabiliy of something serious to A&E, causing an unwarranted increase in demand for the most expensive part of the formal medical sytem. And that this lesson applies to similar expert systems being touted everywhere by the Technorati as the Next Big Thing.
The idea is laudable - you go online or ring up the 111 helpers and "self provision" your diagnosis, and potentially self treat, thus saving the expensive professional medical system a lot of time, capacity & money. An excellent idea, especially as the "family doctor" system in the UK only really operates for about 10 out of 24 hours at best, and even to access that in a hurry is damn difficult if you are working. So you go to the online help out of hours, or even in hours if you can't get an appointment that day.The flaw is that both the web system and the tele-helper are relying on an Expert Algorithm, that unfortunately isn't.
This is how it (doesn't) work. A member of my family started getting pretty ill with aching limbs, nausea and vomiting, and a blinding headache. If you go on the symptom website with any permutation of these symptoms, the system soon decides that you may have meningitis,
a very nasty condition, and tells you to get to A&E post haste, if not sooner.
However, these symptoms are also the near-identical symptoms for
Norivirus, or Winter-Flu, a nasty bug but one that can be safely self treated at home in the vast majority of cases.
Needless to say, the probability of my family member having Norivirus with those symptoms, in winter, while the bug is occurring at its max, is infinitely more likely than having meninigitis. The expert system does not seem to have this "probabiity" function however, and appears to default straight to a lowest risk stance "this could be meningitis - go to A&E now!"
Even so, this Expert System is probably OK for the single case - but now scale it to a country of 50 million or so people, make it the easiest method of getting advice, scale down on the more skilled alternative (aka "doctors") as a second opinion and and multiply me by all the other people getting hit by the current Norivrus bug going round, and you have very likely generated a large number of unnecessary visits to A&E by the "Worried-not-quite-well" who have had the bejeezus frightened out of them by having an automated system that can't discriminate properly between a dangerous condition and a nasty winter bug, and drops to the lowest risk position of "get thee to A&E".
Also, you'd expect the system to try and do a few clarifying questions, given the huge difference in potential outcomes in a diagnosis of norivirus vs meningitis. There are a few to be fair - far fewer than I'd have expected though - but even then they are odd. For example, one question is "have you flown in from abroad in the last 3 weeks". Well yes, we have, as a matter of fact. But I was then expecting the system to ask where from - we flew in from alpine Europe, not exactly a dangerous disease hothouse. Now I don't know if the lack of the "where did you fly from" question in this "Expert" system was to make it easier to use, or to not offend anyone, but to my mind it was a pretty useless way of separating norivirus from meningitis.
The main issue with the question chain though, was you sat there looking at the screen thinking that none of these options properly described the issue - where was the "other/tell us what really happened/that isn't right" button?
Anyway, our diagnosis was sorted by talking to a real expert system - our doctor - over the telephone the next morning. "Oh yes, there is a lot of it going round right now, here's a prescription coming via email, go to XXX pharmacy and pick up the medicine". Five minutes on the 'phone with a
real expert system = no panicked visit to A&E, no blocking resources for those who need it more, etc etc. But to use this option you (i) have to take the risk that it isn't meningitis on yourself over an anxious night, and (ii) have the self confidence that is probably the right call despite the prevalence of lurid medical "advice" Google gives at the touch of a button. Not a good failsafe resort for a nationally deployed medical expert system.
I would therefore bet a lot of money that a lot of people, especially out of hours, used that website or 111 and, when told to get to A&E PDQ, went there in a panic and helped no little bit in creating the huge crunch in A&E over the last few days. It may have only been a small % uplift, but the way capacity constrained complex systems work is they don't degrade linearly, they degrade in a non linear (aka accelleratingly worse) mode until they collapse.
Now it is something of a belief in the Technorati that these expert systems
are inevitable and will be the saviour of many services/automate many people out of jobs/be the great leap forward for mankind etc etc (choose your Future of Technology belief set) but this experience maps onto something far more prosaic that I have seen repeatedly in the 30 or so years I've been building system models and simulations, to wit that:
- no simulation ever captures the overall complexity of reality
- any error, no matter how small, is magnified massively when a high volume of "events" (the simulation term for, in this case, a user going through the diagnosis) are thrown at it
- those errors then cause huge downstream problems long before they are detected
- fixes are put in, usually with intervention by much higher cost informal side-systems
- all too often, a more compex edifice is created than the one it replaced
- the system winds up costing far more from its errors than any putative savings
In other words, be very careful of how these early-day expert systems are deployed, as their errors could cost a hell of a lot more than the theoretical savings they may generate (unless, as in the case of the banks in 2007/8, you can make the public pay of course).
