Re: Science of Machine Learning (was Machine Learning , some thoughts)

[Bert Verhees]

On 03-07-18 13:13, Anastasiou A. wrote:

Initially, I thought that it would have been this one


Opinions from yesterday may still be valid today.

Inventions and business models follow up quickly. But the law is behind, 
as law should be: conservative, keeping an eye on human rights.



___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Re: Science of Machine Learning (was Machine Learning , some thoughts)

[Bert Verhees]

On 03-07-18 12:21, Philippe Ameline wrote:

Le 02/07/2018 à 11:31, Bert Verhees a écrit :


On 30-06-18 17:16, Philippe Ameline wrote:

(improperly labeling images or adding images of objects that are not
plants) could probably make the whole app plainly crappy.

Of course Philippe, but that would be vandalism. Most sensible people
don't do that when they stand behind the goal, and a little bit of
dirt, therefor it is Machine Learning, it can filter it out. It is
part of the learning process.

If a culture of data quality is properly installed, then it is possible
to name improper use "vandalism".
In medicine, since such a culture has never existed, we could name it
"don't carisme", "no time for thisisme" or "was never thaughtisme".
Okay, not vandalism but don't-careism. The result is different. The 
first gives wrong data to frustrate the machine learning process, the 
second does not give data, voluntarily or not of good quality.


Good that there are procedures that create good data to learn from, 
these data are recorded anyway.
For example, in medical imaging diagnosis. Often this is very accurate 
and also cheap and fast. This not science fiction. This not new.
Early detection of diseases can reduce cost for healthcare enormously 
and will change the daily practice of healthcare.


Not only to find cancer, but even early detection of alzheimer is being 
worked on or already in use.
Currently, medical images account for 90% of all medical data, according 
to an IBM-report a year ago. This will be much more, and this will 
happen soon.


These machine learning processes do not suffer from don't-careism 
because the images are produced anyway, and have the manual diagnosis to 
learn from also.
Medical imaging is a good candidate for machine learning, not only 
because of the data which are very suitable, but also because of the 
importance, and (I repeat because of your argument) the processing for 
getting data does not require extra effort.


Upload images to a web-service, so hospitals do not have to buy 
expensive machines or employ specialists for this. Just upload the image 
and within 5 seconds, there is an analysis with high accuracy and cheap.

https://lunit.io/
https://www.vuno.co/

Also ultrasound supported by machine-learning/deep learning, “Users can 
reduce taking unnecessary biopsies and doctors-in-training will likely 
have more reliable support in accurately detecting malignant and 
suspicious lesions,” said Professor Han Boo Kyung, a radiologist at 
Samsung Medical Center.

https://www.samsunghealthcare.com/en/products/UltrasoundSystem/RS85/Radiology/benefit

I think it is time for optimism. But there is one thing that can come in 
the way. People might not accept being diagnosed by a machine, although 
this diagnose is more trustable. Also radiologist may fear for their 
employment, but instead of taking radiologists’ jobs, their job will 
change to prediction and guiding treatment. (so says Dr. Bradley 
Erickson from the Mayo Clinic in Rochester, Minnesota)


Bert



___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Re: Science of Machine Learning (was Machine Learning , some thoughts)

[Philippe Ameline]

BTW, is someone aware of this project by Google?
https://ai.googleblog.com/2018/05/deep-learning-for-electronic-health.html


Le 03/07/2018 à 12:40, Birger Haarbrandt a écrit :
> Hi Philippe,
>
> I completely agree with your view. This is why data stewardship is
> needed before we can make real use of the data:
> https://en.wikipedia.org/wiki/Data_steward
>
> As we use this approach in HiGHmed, I might be able to report in 2020
> about lessons learned :)
>
> Best,
>
> -- 
> *Birger Haarbrandt, M. Sc.
> Peter L. Reichertz Institut for Medical Informatics (PLRI)
> Technical University Braunschweig and Hannover Medical School
> Software Architect HiGHmed Project *
> Tel: +49 176 640 94 640, Fax: +49 531/391-9502
> birger.haarbra...@plri.de
> www.plri.de
>
>
>
> Am 03.07.2018 um 12:21 schrieb Philippe Ameline:
>> Le 02/07/2018 à 11:31, Bert Verhees a écrit :
>>
>>> On 30-06-18 17:16, Philippe Ameline wrote:
 (improperly labeling images or adding images of objects that are not
 plants) could probably make the whole app plainly crappy.
>>> Of course Philippe, but that would be vandalism. Most sensible people
>>> don't do that when they stand behind the goal, and a little bit of
>>> dirt, therefor it is Machine Learning, it can filter it out. It is
>>> part of the learning process.
>> If a culture of data quality is properly installed, then it is possible
>> to name improper use "vandalism".
>> In medicine, since such a culture has never existed, we could name it
>> "don't carisme", "no time for thisisme" or "was never thaughtisme".
>>
>> My point, and what the paper I previously pointed out explains, is that
>> trying to get something out of machine learning in a domain of poor data
>> quality is a modern kind of magic thinking.
>> It just means that any such project should first organize for data
>> quality as a first step.
>>
>> When considering it in hindsight, it makes sense since machine learning
>> involves statistics and data quality is paramount in this domain.
>>
>>
>> ___
>> openEHR-clinical mailing list
>> openEHR-clinical@lists.openehr.org
>> http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org
>
>
>

___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Re: Science of Machine Learning (was Machine Learning , some thoughts)

[Birger Haarbrandt]

Hi Philippe,

I completely agree with your view. This is why data stewardship is 
needed before we can make real use of the data: 
https://en.wikipedia.org/wiki/Data_steward


As we use this approach in HiGHmed, I might be able to report in 2020 
about lessons learned :)


Best,

--
*Birger Haarbrandt, M. Sc.
Peter L. Reichertz Institut for Medical Informatics (PLRI)
Technical University Braunschweig and Hannover Medical School
Software Architect HiGHmed Project *
Tel: +49 176 640 94 640, Fax: +49 531/391-9502
birger.haarbra...@plri.de
www.plri.de



Am 03.07.2018 um 12:21 schrieb Philippe Ameline:

Le 02/07/2018 à 11:31, Bert Verhees a écrit :


On 30-06-18 17:16, Philippe Ameline wrote:

(improperly labeling images or adding images of objects that are not
plants) could probably make the whole app plainly crappy.

Of course Philippe, but that would be vandalism. Most sensible people
don't do that when they stand behind the goal, and a little bit of
dirt, therefor it is Machine Learning, it can filter it out. It is
part of the learning process.

If a culture of data quality is properly installed, then it is possible
to name improper use "vandalism".
In medicine, since such a culture has never existed, we could name it
"don't carisme", "no time for thisisme" or "was never thaughtisme".

My point, and what the paper I previously pointed out explains, is that
trying to get something out of machine learning in a domain of poor data
quality is a modern kind of magic thinking.
It just means that any such project should first organize for data
quality as a first step.

When considering it in hindsight, it makes sense since machine learning
involves statistics and data quality is paramount in this domain.


___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org




___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Re: Science of Machine Learning (was Machine Learning , some thoughts)

[Philippe Ameline]

Le 02/07/2018 à 11:31, Bert Verhees a écrit :

> On 30-06-18 17:16, Philippe Ameline wrote:
>> (improperly labeling images or adding images of objects that are not
>> plants) could probably make the whole app plainly crappy.
>
> Of course Philippe, but that would be vandalism. Most sensible people
> don't do that when they stand behind the goal, and a little bit of
> dirt, therefor it is Machine Learning, it can filter it out. It is
> part of the learning process.

If a culture of data quality is properly installed, then it is possible
to name improper use "vandalism".
In medicine, since such a culture has never existed, we could name it
"don't carisme", "no time for thisisme" or "was never thaughtisme".

My point, and what the paper I previously pointed out explains, is that
trying to get something out of machine learning in a domain of poor data
quality is a modern kind of magic thinking.
It just means that any such project should first organize for data
quality as a first step.

When considering it in hindsight, it makes sense since machine learning
involves statistics and data quality is paramount in this domain.


___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Re: Science of Machine Learning (was Machine Learning , some thoughts)

[Bert Verhees]

On 30-06-18 17:16, Philippe Ameline wrote:

(improperly labeling images or adding images of objects that are not
plants) could probably make the whole app plainly crappy.


Of course Philippe, but that would be vandalism. Most sensible people 
don't do that when they stand behind the goal, and a little bit of dirt, 
therefor it is Machine Learning, it can filter it out. It is part of the 
learning process.



___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Re: Science of Machine Learning (was Machine Learning , some thoughts)

[GF]

Data of perfect quality means, in my opinion, data and their complete context.
A diagnosis by a nurse is not the same as one by a patiente, or strting intern, 
or one MD with 20m years experience.
Just mentioning one example.



Gerard   Freriks
+31 620347088
  gf...@luna.nl

Kattensingel  20
2801 CA Gouda
the Netherlands

> On 30 Jun 2018, at 17:16, Philippe Ameline  wrote:
> 
> Le 27/06/2018 à 22:26, Bert Verhees a écrit :
> 
>> On 27-06-18 16:43, Philippe Ameline wrote:
>>> 1) you can find a bunch of practitioners that agree on working extra
>>> hours to comment a big bunch of images, or
>> 
>> Did I tell you about the plant-app? I believe I did. 700.000 pictures
>> are reviewed, often by volunteers.
>> 
>> The app recognizes 16000 plants. Important is how you do it, and that
>> it does not cost effort by the volunteers, for example in relation to
>> what they do anyway.
>> 
>> https://plantnet.org/ 
>> 
>> It is a French product.
> 
> Dear Bert,
> 
> The plant-app was the subject of your initial post.
> 
> The math in support of deep learning are being studied. To make it
> short, it remains somewhat mysterious since such classification
> algorithms "should not work", but actually, they do ;-)
> 
> From an article I just read, such NP complete algorithms are similar to
> finding a needle in a hay stack and shouldn't provide valuable
> answers... unless the conditions (large enough needle, correctly ordered
> stack) make the problem handy.
> 
> To sum it up, data quality (signal over noise ratio) is paramount. In
> the plant-app you mentioned, adding a certain level of fuzziness
> (improperly labeling images or adding images of objects that are not
> plants) could probably make the whole app plainly crappy.
> 
> Just to say that building a deep learning system starts from making
> certain that the data it will be fed with are of proper quality. This is
> usually not the case in medicine, largely because IT is considered a
> back office concept and there is seldom the kind of feedback loop that
> could lead to having errors fixed.
> 
> My point is that you can perfectly (but with considerable efforts)
> organize a trained network of practitioners to feed a "data lake" in
> order to train a neural network... but will probably be disappointed if
> you try to process existing information.
> 
> Best,
> 
> Philippe
> 



signature.asc
Description: Message signed with OpenPGP
___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org

Science of Machine Learning (was Machine Learning , some thoughts)

[Philippe Ameline]

Le 27/06/2018 à 22:26, Bert Verhees a écrit :

> On 27-06-18 16:43, Philippe Ameline wrote:
>> 1) you can find a bunch of practitioners that agree on working extra
>> hours to comment a big bunch of images, or
>
> Did I tell you about the plant-app? I believe I did. 700.000 pictures
> are reviewed, often by volunteers.
>
> The app recognizes 16000 plants. Important is how you do it, and that
> it does not cost effort by the volunteers, for example in relation to
> what they do anyway.
>
> https://plantnet.org/
>
> It is a French product.

Dear Bert,

The plant-app was the subject of your initial post.

The math in support of deep learning are being studied. To make it
short, it remains somewhat mysterious since such classification
algorithms "should not work", but actually, they do ;-)

From an article I just read, such NP complete algorithms are similar to
finding a needle in a hay stack and shouldn't provide valuable
answers... unless the conditions (large enough needle, correctly ordered
stack) make the problem handy.

To sum it up, data quality (signal over noise ratio) is paramount. In
the plant-app you mentioned, adding a certain level of fuzziness
(improperly labeling images or adding images of objects that are not
plants) could probably make the whole app plainly crappy.

Just to say that building a deep learning system starts from making
certain that the data it will be fed with are of proper quality. This is
usually not the case in medicine, largely because IT is considered a
back office concept and there is seldom the kind of feedback loop that
could lead to having errors fixed.

My point is that you can perfectly (but with considerable efforts)
organize a trained network of practitioners to feed a "data lake" in
order to train a neural network... but will probably be disappointed if
you try to process existing information.

Best,

Philippe


___
openEHR-clinical mailing list
openEHR-clinical@lists.openehr.org
http://lists.openehr.org/mailman/listinfo/openehr-clinical_lists.openehr.org