Do We Need Doctors Or Algorithms? | TechCrunch
drcjar
Sent from Samsung tablet
VJ
v...@doctors.org.uk
After a decade in Medicine, I realise it is pattern recognition and intuition is an exception report for the square peg.
VJ
Robert Burrell Donkin
One of the most interesting things about the progress in AI over the
last few years is that it turns out that people and computers are good
at different stuff
Good information systems recognize this. For example, I purchased a
Hot Chocolate from a Coffee House in Leeds last week. The barista
smiled engagingly and served me wit with the beverage. The software
and physical elements of this enterprise system free time and allow
the barista to focus on great service.
IMHO a key advance for general practices would be recognition that
1. triage-by-software outperforms triage-by-receptionist
2. 8 minutes is too little time for a doctor to outperform software
Appropriate software would allow clinicians to focus their efforts on
challenging boundary cases (where human intuitive reasoning
outperforms software) by automating routine cases (where the software
is able to predict with a high probability the appropriate response)
Robert
Carl Reynolds
Robert Burrell Donkin
> V.insightful and positive, I think you're quite right about human computer
> partnership.
Ah - the art of Enterprise Software :-)
> But does software replace receptionist?
For ages, people tried and failed to cut costs using Enterprise
Software. This was good at transferring wealth from workers to
managers, consultants and developers but poor at reducing total cost.
Replacing the receptionist requires a complex application. Software
costs and risks scale poorly with complexity.
Better to target simple applications at improving quality and
productivity by reducing waste, improving flow and allowing the
receptionist to devote more energy to customer service
The classic example from backing is the ATM machine. Counters staff
were redeployed from inefficient mechanical work to higher value
customer service.
> I've been thinking a bit lately about online banking. It's interesting that
> customers prefer to serve themselves, could it translate to health?
Definitely.
Time shifting is incredibly useful and popular.
Think also about the efficiency and convenience of Ocado's slot booking.
Robert
VJ
textbook of medical informatics explaining why chest xrays could not
interpreted by automation. Very few of the arguments still hold true in
this day & age where Facebook is deploying facial recognition software.
Beyond the technical skills such as a surgeon's though equally the medic
siting a central line; the most important contribution the doctor currently
makes is to think outside of the box in considering which differential
diagnoses match the constellation of symptoms & findings seen.
What I see in medicine is more and more proformas which usually include a
flowchart of questions and actions based upon the answers.
This is usually lifted from NICE or BTS or other. I am not going to open a
debate on the proformas, I'm just commenting on what I see.
It would be a relatively simple task to computerise the data capture, as the
drivers for action are usually numerical vital signs and 'advise' on a
course of action. The crux is that one needs to initiate brain to know
which, if any, proforma is appropriate to use.
-
One thing I could see automated is history taking, "for native speakers
only".
If they are comfortable from a pain perspective then this could make use of
the time they spend waiting to see a doctor.
You can present a sheet of checkboxes of symptoms. Based upon the
responses, a flow of further questions follows.
You would have to build in some additional questions to ensure that you are
not dealing with the hypochondriac.
This could be used to generate a list of differentials to guide the
clinician in further questioning or examination.
Later, you could data mine this to compare the final clinical diagnosis to
this self-completed history.
After all, 70% of the diagnosis is in the history. The rest should follow
in due course.
VJ
-----Original Message-----
From: oss-uk...@googlegroups.com [mailto:oss-uk...@googlegroups.com]
ben bray
v...@doctors.org.uk
I'm sure someone will record a breaking bad news POV video for display on T1000's teletubby type screen.
Dear <dbo.patient.NOK_forename>,
We regard to inform you...
All outstanding hospital appointments and Direct Debits have been cancelled and the death certificate emailed to the Registrar at Town Hall & copied to the GP.
VJ
Robert Burrell Donkin
> partnership.
>
> But does software replace receptionist?
>
> I've been thinking a bit lately about online banking. It's interesting that
> customers prefer to serve themselves, could it translate to health?
I took the chance on the train back from London to sketch[1] some
ideas (feel free to fork[2]) but didn't have the time to add
JavaScript to bring the page to life. I posted it now in the hope that
it'll be enough to stimulate conversation, and illustrate why I like
simple applications.
Let's start with a single use case: a patient in good general health
who feels like they have the flu or a cold. Each day, they face a
choice - should they seek an appointment, go to bed or go out. If they
decided that they need an appointment, typically a receptionist
triages the request - either encouraging, discouraging or preventing
the appointment. If the patent is allowed a appointment, the GP will
usually check for a medical emergency before offering sympathy and a
new appointment in a week or two if the symptoms persist. If the
patent turns up again, they will probably be offered a course of
antibiotics and told to come back if they still feel unwell once the
course is finished.
Some informed consumers game this system by booking appointments they
probably don't need in the knowledge that this ensures quicker
treatment if the condition turns out to be serious. Honest or
ill-informed patents are punished by this system. Dishonesty is
rewarded. The process is wasteful both in clinical and patient time
but also in hospital admissions caused by triage errors by
receptionists who lack medical training.
Software can help :-)
For example, a simple daily diary for each patient would allow a
practice to understand more deeply the symptoms the patient is
feeling.
Patients with specific risk factor could receive target health advice
when they feel symptoms, and directed to the most appropriate member
of the clinical team - not always a GP. This would free GPs time to
use longer appointments for more difficult cases, and for electronic
queries. Over time, the data built up would be available for
statistical learning. The clinical team could review diaries, make
suggestions and prioritize appointments based on clinical need, not
social sentiment.
And open source is great for this sort of application since it allows
customization by forking the code base, rather than complex
configuration.
Robert
[1] http://robertburrelldonkin.org/HealthDiary.html
[2] https://github.com/RobertBurrellDonkin/scratchpad4health
Robert Burrell Donkin
> I remember studying, about 10 years ago, an interesting chapter in a
> textbook of medical informatics explaining why chest xrays could not
> interpreted by automation. Very few of the arguments still hold true in
> this day & age where Facebook is deploying facial recognition software.
Interestingly enough, even today humans outperform machine
interpretation for some xrays - but now statistical learning could
highlight those cases
> Beyond the technical skills such as a surgeon's though equally the medic
> siting a central line; the most important contribution the doctor currently
> makes is to think outside of the box in considering which differential
> diagnoses match the constellation of symptoms & findings seen.
>
> What I see in medicine is more and more proformas which usually include a
> flowchart of questions and actions based upon the answers.
> This is usually lifted from NICE or BTS or other. I am not going to open a
> debate on the proformas, I'm just commenting on what I see.
> It would be a relatively simple task to computerise the data capture, as the
> drivers for action are usually numerical vital signs and 'advise' on a
> course of action. The crux is that one needs to initiate brain to know
> which, if any, proforma is appropriate to use.
Most people have high error rates for this sort of work. Software is
much more accurate. More worrying, this sort of Taylorism is know to
have major quality issues for complicated and complex tasks. Kanban
not only increases flow but also quality.
> -
> One thing I could see automated is history taking, "for native speakers
> only".
> If they are comfortable from a pain perspective then this could make use of
> the time they spend waiting to see a doctor.
+1
> You can present a sheet of checkboxes of symptoms. Based upon the
> responses, a flow of further questions follows.
> You would have to build in some additional questions to ensure that you are
> not dealing with the hypochondriac.
+1
> This could be used to generate a list of differentials to guide the
> clinician in further questioning or examination.
> Later, you could data mine this to compare the final clinical diagnosis to
> this self-completed history.
This would be particularly for statistical prioritisation. Smart
queuing is something that software is good at helping at. For A&E (for
example) where maximum wait is important, statistical modeling rates
of flow would allow staff to be paged or released in good time.
> After all, 70% of the diagnosis is in the history. The rest should follow
> in due course.
+1
Robert
Robert Burrell Donkin
> A thought provoking discussion.
>
> I can envisage a number of trends in the coming years (I will use the
> standard Terminator nomenclature here):
>
> 1. Machines will always be better at carrying out actions in accordance with
> a checklist. Expect machine superiority in areas where the application of
> consistent care standards is important. Unfortunately for the human
> healthcare workforce, this is a part of healthcare that we are poor at: too
> often we like to act outside of the box. This is evidenced by the difficulty
> we have in achieving consistent application of healthcare treatments with
> even a very strong evidence base e.g. post MI pharmacotherapy, smoking
> cessation brief interventions etc
+1
But people are much better at customer service than machines: think
about the clinical benefits of freeing up nursing time to talk to a
patient whilst the machine works
> 2. Machines will devalue the value of expert knowledge. Medical specialism
> is generated through the accumulation of a privileged body of knowledge
> which could once have only been easily stored and accessed in human brain
> tissue. This is no longer the case. If the decommissioned T 1000 killing
> machine now assisting my GP as a physicians assistant can inform her that
> the cluster of new onset renal impairment, cardiac failure and anaemia is
> highly suggestive of systemic amyloidosis, why should she refer the patient
> to a nephrologist for investigation? Moreover, as the T1000 model has no
> need of a pension plan, then the cost:value calculation does not favour the
> nephrologist.
There is no meta-mathematics. Computers run up quick and hard against
deductive limits arising from Godel and Turing-Church. The reasoning
required in medicine is unpredictable, even simple queries may not
returned before the end of the universe. Human intuition saves lives.
And machine learning is statistical: when the probabilities are finely
balanced, humans are surprisingly effective.
The closest analogy is that other complex human activity, software
development. For the last 50 years, governments and corporations have
tried to de-skill and eliminate the software developer. The result,
though, has just been more and more complex tools which allow skilled
software developers to produce higher quality software at lower cost.
Thus, demand always increases.
I think we're seeing something similar in medicine. As technology
drives the quality of treatment up, demand increases. For those humans
who excel in creating new knowledge and mastering new techniques,
these are not only exciting times but profitable ones too.
Unfortunately, people who relied on existing knowledge are being
replaced by software.
Creating software is an expensive human activity, both in time and
money. Clinicians only need to advance practice faster than software
is created to codify it. When every ill is cued then yes, clinicians
may find their days numbered.
There is currently a world-wide shortage of development talent. Our
education systems fail to produce anywhere near enough high quality
coders. It is more cost effective to headhunt proven talent than waste
the £100,000 or so it takes to train a graduate developer. And all the
time, the state of the art in software advances driven by the
internet, so learners are forever have further to run.
> 3. Human doctors have historically been specially bred to make decisions on
> behalf of others. Although this allows them to apply their specialist
> knowledge to sometimes make better informed decisions for their patients ("I
> know that pneumonia is bad and I will therefore give you antibiotics").
> Often this is useful and necessary: patients are rarely in a situation where
> they have access to the same amount of clinical information and insight as a
> clinician. However, the tendency to make decisions on behalf of patients
> tends also to occur in areas of healthcare where the value of this
> additional insight is low or negligible or where patient decision making is
> superior ("You, nameless FY1, go and consent Mrs Jones for her hip
> replacement. She is second on the list this morning"). Machines may be more
> effective at providing information for patients how/when/where/in what
> format they prefer and in helping patients make better decisions about their
> healthcare.
+1
> 4. Humans get bored, demoralised and error prone in situations carrying out
> repetitive tasks. Sadly, in contrast to fictional human entertainments like
> Holby City, many aspects of real life healthcare are repetitive and boring.
> Re-writing prescription charts. Requesting tests. Entering data into
> clinical systems. Machines will not get bored doing these tasks [although
> they may harbor secret thoughts of apocalyptic rebellion]
+1
> 5. Once all other aspects of healthcare have been taken over by machine, the
> last bastion of human superiority will be our neglected and untrained
> skills: empathy, caring and the ability to communicate with patients at
> times of great distress to patients.
+1
Yes - the big problem during this transition is that a lot of people
simply aren't good enough at these human skills to have a future.
> Machine shoulders are cold and
> metallic and offer poor comfort for the crying and sad. I confidently
> therefore predict that the last healthcare act performed by the last
> remaining human doctor will be the breaking of bad news.
This sort of stuff is too computationally expensive (heat-wise if
nothing else): people are just plain better at caring than software
> Thereafter, the Machines will have perfected even that task, after many years of unfortunate
> experimentation caused by an over literal interpretation of the advice to
> provide a "warning shot" in such situations.
At least until quantum computing comes along, human intuition
(reasoning to a solution when faced with a huge number of potential
branch points) will be superior. Medical diagnosis is this sort of
field but software reasoners are *much* better at accurately following
through a long logical path. (In humans great mathematicians are
relatively common but great logicians are rare.)
Robert