Machine Learning , some thoughts

Clinical (archive)
1

Today my wife showed me Plantnet.

https://plantnet.org/en/

It recognizes over 6000 plants from showing a flower or a leaf to your phone. It has learned from machine-learning 700.000 pictures, and its knowledge every day grows stronger, because it keeps on learning. And not only the looks of a flower, but if it takes location (biotope) and date in consideration, the certainty of recognizing gets stronger.

Now you can imagine that it must be hard to recognize a plant from a picture, without seeing the dimensions and showed in many possible angles, in sunlight, cloudy or twilight.

I was impressed how good it already was. Very advanced computer-knowledge for free in the hands of the millions.

There is also an app, I did not try it, which recognizes birds from audio. You walk somewhere, hear a bird and want to know what kind of bird that is.

The Berlin Natural History Museum leads a contest of 29 teams using 23 different methods, with more than 82% good identifications for isolated bird recordings, and more than 74% correct identifications for recordings mixing several bird songs.

I often notice there is a trend in thinking that Machine Learning cannot be much help, see how miserable google-translate translates. But then we for get to see how much progress is made in other areas.

Why am I writing this? Just to let you think about it.

I wonder, Is OpenEhr usable for recognizing pattern in diseases over Machine Learning, isn't behind every diagnosis a small cloud of archetypes which forms a pattern? The features of recognizing/learning should not be found in archetypes ID's, although, that can help a lot, but it should also look to datatypes, their semantics and relations.

Isn't OpenEhr better for recognizing pattern then whichever classic storage structure, because the data-structures in OpenEhr are in semantic models, this instead of some weird Codd-structure, which only has technical reasons to exist.

(Classic data stored in classic SQL schema's could be brought over to archetyped structures, to make the base of machine-learning larger.)

I think, when this is developed, we should be able to get to at least two advantages.

1) We don't need CKM anymore, computers can understand archetypes, we don't need to restrict ourselves to a limited number. We can also use archetypes we do not know, and maybe we never know. Even, we wouldn't need archetypes anymore, just as reminder/instruction. But the computer could create the archetypes on the fly, when seeing the kind of data, the relations, the diagnosis.

2) We could use the pattern to recognize healthcare situations, and maybe treat/handle/cure on base of instructions coming from machine learning.

Some thoughts when walking with my wife through the wonderful dunes, and its special vegetation. Maybe I must write a blog about it.

Have a nice day.

Bert

2

82% of correct recognition rate is a desaster in healthcare.
74% is even worse.

My evidence based feeling is that we still will need to sort it out manually for some years to come.

Hope this helps,
Stefan

Stefan Sauermann

Program Director
Biomedical Engineering Sciences (Master) ->
Medical Engineering & eHealth (Master) in September 2018!

University of Applied Sciences Technikum Wien
Hoechstaedtplatz 6, 1200 Vienna, Austria
P: +43 1 333 40 77 - 988
M: +43 664 6192555
E: stefan.sauermann@technikum-wien.at
I: www.technikum-wien.at/mme
I: www.technikum-wien.at/bhse
I: healthy-interoperability.at
fb: www.facebook.com/uastwMME
portfolio: https://mahara-mr.technikum-wien.at/user/sauermann

3

92% would be a disaster in healthcare … I am slightly more optimistic: I suspect that the key bit of research is to create machines, e.g. for interpreting images, that can accurately distinguish between 3 of image: don’t know; not sure (error rate likely to be too high); and sure (e.g. less than 0.2% error rate or similar). Such machines would throw images in the first two groups to humans, and would do the work on the ‘sure’ group. The key is to be able to recognise ambiguity or the lack of it. Doing this properly might require more than one kind of AI. And of course, AI image interpreters would not need to work with displayed bit maps, but would work with computable 3-D and 4-D matrices. - thomas

4

Not in visual classification of dermatological health concerns.

Or areas of radiological diagnostics.

Karsten Hilbert

5

Providing health and care is part science and for a large part an art.
Meaning that humans are needed.

Artificial Intelligence is a nice scientific hyped topic and nothing more.

That is not to say that AI might play a role and can be of use.
It needs to be properly designed, engineered and not hacked together.
It is certain that AI applications in healthcare must be treated as Medical Devices.

For it function properly we need to be able to document healthcare topics including the full context/epistemology.
Present OpenEHR/13606 and terminology developments form a good foundation.
But are not sufficient.
What is lacking are well researched and designed shared patterns that capture the full context, epistemology.
CIMI is trying to do that.
CIMI, part of HL7, is possibly diverging and getting under the influence of practical thinking as it is adjusting ist gols to encompass FHIR.

GF

Gerard Freriks
+31 620347088
gfrer@luna.nl

Kattensingel 20
2801 CA Gouda
the Netherlands

6

Dear Bert and all

I wonder, Is OpenEhr usable for recognizing pattern in diseases over
Machine Learning, isn't behind every diagnosis a small cloud of
archetypes which forms a pattern? The features of recognizing/learning
should not be found in archetypes ID's, although, that can help a lot,
but it should also look to datatypes, their semantics and relations.

openEHR and the dual modelling approach, model the data which reflect the
evolution of a person's health condition.

Inferences from the data (irrespectively of whether they are stored in an openEHR enabled
server or not) would return information about the health condition itself.

Inferences from the data models (descriptions of the data) would return information
about the modeller's understanding of the health condition and anything else that stems from
the requirements for cataloguing it.

As a similar example, consider what else does ICD-10 encodes along with the rest of the "...Classification of Disease":
https://www.icd10data.com/ICD10CM/Codes/V00-Y99/V30-V39

We usually express research questions in sets of codes, sometimes with branches to make sure that we pin point
exactly the information about the condition and not anything else that might be encoded along.

Isn't OpenEhr better for recognizing pattern then whichever classic
storage structure, because the data-structures in OpenEhr are in
semantic models, this instead of some weird Codd-structure, which only
has technical reasons to exist.

Yes and no. It is more a question of whether the EHR does encode all information that is dependent on the class.

1) We don't need CKM anymore, computers can understand archetypes, we
don't need to restrict ourselves to a limited number. We can also use
archetypes we do not know, and maybe we never know. Even, we wouldn't
need archetypes anymore, just as reminder/instruction. But the
computer could create the archetypes on the fly, when seeing the kind
of data, the relations, the diagnosis.

The last bit where the "computer" composes the archetype is indeed incredibly interesting and in the current climate
it could be something like "AI Assisted Archetype Composition"......I doubt we can go 100% automated there.

Do words encode everything in a language exactly?

Logic expressions / deductions are "stale". Given the axioms and the operations, you can work out the complete universe.

This is partially true for **some** of the stuff that we do but for others, there comes a point where the "Concept Set" gets
re-distributed or enriched. New things, new ideas, new conceptions are coming in. The "computer" needs a way to encode
this process in order to express it and this will happen as soon as "we" understand how that happens too :slight_smile:

2) We could use the pattern to recognize healthcare situations, and
maybe treat/handle/cure on base of instructions coming from machine
learning.

The time scales for doing this would be enormous. We can probably work out a lower limit by looking at the lifecycle of archetypes
in the current CKM.

BTW, I am not shooting down the proposals / ideas, there is definitely fertile ground for the use of openEHR along the lines you suggest here.

All the best
Athanasios Anastasiou

7

It much depends. In typical care "92%" (of what ?) can be an
extremely brilliant result far beyond anything available
today.

Say, correct ECG interpretation for NSTEMI in the presence of
thoracic pain in an unselected primary care setting.

Karsten

8

I agree, especially on GP-level, I checked with my wife, she is GP, as you (Gerard) know. I asked her if the context/epistemology in a EHR is sufficient for machine-learning. It is not, she sufficient, and that will never be. GP's have other things to do then carefully record all datapoints that describe a disease.
Even when using archetyped-systems this does not change.

Allthough, there are some patient-conditions which are very typical for a disease, mostly this is not the case.
For example, many infection-diseases have fever as a symptom, and one person gets pain in his back, and the other has headache as a result of fever and other inconveniences coming with infection disease.

So, the GP cannot do much with machine learning, the best source of knowledge is his experience, and if he cannot solve with that, he should ask someone else, or send the patient to the hospital to a specialist.

But there, machine learning can do things in some specialties.

Anyway, thanks for your reply
Bert

9

Not really Stefan, but thanks for trying.

10

Allthough, there are some patient-conditions which are very typical for a
disease, mostly this is not the case.
For example, many infection-diseases have fever as a symptom, and one person
gets pain in his back, and the other has headache as a result of fever and
other inconveniences coming with infection disease.

So, the GP cannot do much with machine learning, the best source of
knowledge is his experience,

Experience is, at most, an equal source to evidence. It
becomes "better" over time.

and if he cannot solve with that, he should ask
someone else, or send the patient to the hospital to a specialist.

Nonetheless, an algorithm _can_ scan records in the
background looking for telltale constellations indicating the
"I am sure" group (which we sometimes DO miss) and highlight
those to the GP.

Karsten

11

So we have always worked with 82% or 92% or 74% recognition, and we never called that a disaster. We called that healthcare, and it is only a few years ago, that is was like that, and in many cases it is still like that. I now know they are using machine learning for checking x-ray’s for cancer, and it is a fact that these machine-learning algorithms are much better and much more accurate then humans are. The find micro-metastasis of only a few cells, and they are of course manually checked. So that is a real improvement. Now, thanks to machine learning the rates of detection are 99%. It is not only much better, but also much cheaper. How much do you think it costs of a radiologist is staring at pictures? Thanks for your reply and confirming this. Bert

12

Thanks, for your answer, I agree with you and others, and already wrote that, that an EHR will not be good enough for machine learning.

I was too optimistic and to much impressed by some results of machine learning. It will do very good things in healthcare, but only on very specific cases.

But while writing this

What would be good, however, an improvement. I suggested to my wife (a GP), and she agreed (partly)

Classic EHR software only has few datapoints on a screen, and many particularities come into free text, and if the GP is really motivated, maybe he finds some ICPC code.

Archetypes do not really change this practice. A GP is a busy person.

What could help is modularity. A GP should be able to add datapoints to his screen. For example, beside all the normal things, the GP sees that there are red eyes, but how can he make this available to the system in a way that it can be found back?

What about micro-archetypes which describe only one datapoint? And the GP should be able to invoke them by voice. He says "red eyes" and magic happens, there is a datapoint on the screen which offers a possibility to click on a checkbox. Eventually a choice, A bit red, medium red, very red.

This kind of software does not have to be something for the far future, but can be available already now.

Also thanks to machine learning, a limited form of NLP (natural language expression (machine learning helping with NLP) can be used, and that was my idea of generating archetypes, last Saturday. A computer could, in some cases of simple datapoints, also even generate micro-archetypes for them, and with templates or container-archetypes, generate evaluation-archetypes

Maybe, when it is so easy to create datapoints, and store them, maybe then machine learning in diagnostic can come closer, also in some cases for a GP, or machine learning can do suggestion: look to the tongue of the patient, but the fact remains, a good GP needs experience for diagnotics.

Bert

13

This approach much reminds me of what Philippe (sp?)
described of his fils guides. Instances of "micro achetypes"
would be generated on the fly while typing/speaking.

As for GP land, care is also rarely about "a diagnosis", but
rather about "a person".

Karsten

14

The doctor mumbles to his screen while the patient tells it story, or the doctor does something and mumbles. He wouldn't even have to change his behaviour.

Looks promising that others think so too.

Bert

15

You may be interested in this paper (from my Tech Trends):
http://philippe.ameline.free.fr/techtreads/additionalMaterial/Boyd_MagicOfBigDataAndArtificialIntelligence.pdf

A friend of mine recently published a paper, after studying a group of GPs located in the South of France. He found out that the diagnosis is not reported in observations in more than one encounter out of two.

Another point is that many medical documents don’t get external feedback and can be of very low quality. In France, patients leave the hospital with an hospitalization report. My mother in law was admitted in an hospital for 3 days several months ago, and her “outpatient report” didn’t mention any of the things she had discussed with doctors… and her weight was supposed to be 20 kg less than her actual one.

Wrong patient, copy-pasted document not properly actuated?

Since nobody dare having this kind of document corrected, it remains forever wrong in hospital’s records. Successfully using machine learning demands a prior culture of data quality and information awareness.

Best,

Philippe

16

Have anyone tried AQL adapter to pandas(python data analysis package
for machine learning and statistics)?

Shinji

17

Hello Bert and all

I am a little bit "worried" with "micro-archetypes" the way you describe them.

I think that what you are probably referring to is "Disease Specific Templates", which I really hope is what we are all working towards :slight_smile:

So, archetypes do indeed describe one conceptual quantity, or aspect of a person's healthcare and then a template describes a multidimensional "Point"
which characterises the patient journey within the disease.

Consider for example "Total Brain Volume". You can use it to track cognitive decline in AD (https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(04)15441-X/abstract)
and this gives you one "explanatory variable". But, still, there are patients whose brain volume is abnormal (for their age) and they still perform well in other tests, so you need more
"data points" (a richer template) around the phenomenon to understand it better.

I think that what you are describing is something like "An automated approach to constructing disease specific 'Minimal Clinical Datasets'".

Once you have this minimal dataset discovered, THEN you could compose the template or automatically create the archetypes.

And yes, this CAN be done today, definitely.

All the best
Athanasios Anastasiou

18

That's because it rarely actually matters.

And it is rarely actually specifiable.

There's a four-stage grouping of "diagnostic certainty" in
German GP research lore:

  'A': _('A: Sign'),
  'B': _('B: Cluster of signs'),
  'C': _('C: Syndromic diagnosis'),
  'D': _('D: Scientific diagnosis')

Most encounters or even episodes of care don't
reach C, let alone D.

Karsten

19

Dear Philippe, I read your document later.

I have to disagree with the word "prior".

It makes it sound like, is has gone wrong long time ago, and there is nothing what we can do.

Big data for machine learning can be build very quick, we have millions of people in healthcare every day.

Imagine a GP making a picture of an eye, or a part of skin, and gets within a second a good explanation about what is there to see.

It is cheap. If many GP's agree to use an app for classifying viewable symptoms, the supporting big database will grow fast.

I also have to agree with Karsten, it is not only a disease which needs to be cured, but it is a person having that disease. So, age, weight, gender, ethnicity, profession, social status, country, those are all factors which limit the search area in which the machine learning database must find what it sees.

20

There is an understandable mindset which aspires to work with a standard-set of archetypes, which are many times reviewed, and which have a review-status and a kind of guaranteed quality. There is a risque of bias in this, for example: That datapoint is not practical, a GP will never record that, or it is not significant. Those predefined archetypes are always a filter on what can occur. But they have also an advantage because they are build on common sense, on what is desirable in healthcare to know. So mostly they cover what is to say about a disease, but it is always knowledge from the past, and always in common sense, so it is quite conservative.

I wonder if besides that approach an approach of archetypes growing in the wild could be of use. They could be used beside the predefined archetypes.

So we don't need to worry, we throw nothing away. We are adding, not replacing.

Bert