Datasets are critical when developing and training AI algorithms. Due to several reasons related to GDPR, Ethical approval secondary intended use of data, etc, etc, access to clinical data is becoming extremely difficult. Most of us have been forced to modify workplans due to lack of access to data as initially was planned.
My question is if anyone have seen studies, or reported these difficulties.
There are several large-scale collaborative projects and repositories you can use instead of collecting the data (because it is even better to test it on unseen data from another source). Like STAR*D, RDoC, COMED, IMAGEN etc...
I think that large datasets, internal and external validation (elatic net, embedded variants, etc) will soon become a must for publishing. Without a proper quantitative assessment of validity of a model you can only have unwarranted optimism. Concerning GDPR, you do not need to do it yourself if you use already used data. So , ideal for developing AI algorithms.
There are a couple of experiences with clinical and preclinical data that can be used for AI and ML. I found it as a big problem in my own research from the flowing reasons:
1) Animal studies are usually funded from the grant, public money, but the research results & medical data, in my case ECGs, are kept within the groups producing those data. They often require being coauthors of any theoretical research conducted above their data.
From my point of view, when they do not know, understand and even are unable to perform mathematical analysis of data they cannot be coauthors. When they provide medical discussion, their contribution must be clearly stated.
2) Clinical research is done on state-owned clinics and therefore data must be publicly available. Again there are present to get undeserved credit by those who are not doing mathematical studies.
3) Data must be available as open source because it can protect society from unwanted fiddling with those data and help to discover errors in research.
There are thousands of AI & ML researchers who will greatly appreciate accessibility to those data as they can train on reproducing already achieved understanding and later pursue their own research on same data that will extend our knowledge.
4) I do hear from all sides, as I follow it for years, cries for valuable and useful data by AI, ML, and BigData scientists who really cannot pursue their research without said data.
5) It seems to be that society, governments, and humanity as the whole highly underestimate the power of free access to biomedical data. Once recognized, it will become a law to put every single data related to biomedical research online as open access to enable control of it any time later. I am strongly for a law that will link the responsibility of biomedical researcher towards the reliability of their data.
Dear JIri,
I know that story so well...Once in attempt to re-analyze someone else's data (if they knew they would have already done that) not only that I was expected to do that for free, they also asked me to guarantee the first authorship to the researcher who is still workin on her PhD and has no connection to machine learning whatsoever...Another time I reused the data, I think they even tried to give me some misleading information about that. So, after playing the fool for several times, I prefer to record it myself (until all the data which collection or recording were funded by us, citizens become freely available, sic!) in order to prevent myself being fool for yet another time.
...when talking to 'fiddling' the data. I am writing something and came across meta-analysis published in very highly ranked and expensive journal with all those -expert names on it... to discover that meta-analysis was based on just six studies, half of which was their own previously published data, with the sample of just 147! What data mining?? Even regular statistics on it, c'mon!
What you wrote sounds utopist, because people are making it up on so many levels, like publishing three or four publications based on the same data. Not to mention money paid to the journals (from those lucky ones who can afford to pay 2500 EUR for the publication costs). So, we are not equal- far from it. I do not think this is going to change soon. It is just that become obvious what lies beneath 'science' in many aspects. And many people like their status quo. So, what will force them not to fiddle the data (and go on receiving high salaries for that)?
- Badges
- Science topic