Whitaker lab meeting with Ed Palmer (6th Feb)

[LouiseABowler]

Kirstie has invited Ed Palmer, one of the new SSI fellows, to her lab meeting on Thursday 6th February. Ed is doing some work on synthetic data so Kirstie has said that the QUiPP team is welcome to come along if we're interested.

Lab meeting runs 3-4pm (with the option to stay later), room ~TBC~ David Blackwell.

Ed's profile is here - not much about his synthetic data work, we heard about that via Sarah from one of the fellows' meetings.

If you're planning to come, can you please put a 😄 emoji on this message so I can make sure we get a big enough room? If there are lots of us, it might be better to set up a separate meeting with Ed earlier that day rather than overfill the usual lab meeting.

[LouiseABowler]

The meeting will be held in David Blackwell.

[LouiseABowler]

Ed's bio and project plans:

I’m a clinician scientist in the final year of my PhD at UCL. My clinical background is in intensive care and anaesthesia, where a large part of my work involves treating those with sepsis (a life threatening infection). My research utilises data from the electronic health record that have been collected as part of the normal care of a patient. I repurpose these data to investigate patterns in patient physiology during sepsis, with the hopes of identifying particular patient subtypes who might respond more successfully to different treatments.

A side interest has been to promote scientific reproducibility within my community, and to help foster a culture where it is ok to share code and make mistakes (you would be surprised how far behind medicine is in this regard). Data sharing is very limited in medicine, due to concerns of confidentiality. Simulated data is a good solution, but security standards and best practices need to be developed. I would love to see a future where it is standard practice to share realistic simulated data, and all study code alongside a manuscript.

[LouiseABowler]

My notes from the meeting:

• Ed is involved in the collection of a large dataset from patients in intensive care
  • Camila: This is the same one that we're applying for access to
  • 250,000,000 data points
  • Data held in the UCL safe haven
• Statistical disclosure control
  • This approach is used for less sensitive, aggregated data
  • Types of columns:
  • Direct identifiers, e.g. name
  • Key variables (not unique, but could be a key in combination), e.g. age and time of admission
  • Sensitive variables e.g. HIV status
  • Non-sensitive e.g. heart rate
  • Typical approach is to remove directly identifying variables, and to "soften" keys (k-anonymity)
  • See also L-diversity for measure of how identifiable the data could be
  • However, so much public data is available that key variable paradigm aren't necessarily valid
• Samples of this data (I think 1/1000) are released
  • Some context is lost
  • Lose linkages between records, so no readmissions data etc.
• Types of synth
  • Things we've looked into: MICE, GANs etc.
  • Augmented synthesis: synthesise only the identifying variables
• Ed's SSI Fellowship:
  • Hold a "synth-a-thon" to bring together people who hold data, people who want synthetic data and people who have ideas about the synthesis process
  • Create new data resources
  • Set community standards
  • Synthesise and share more data (links to the Turing Way)

From follow-up discussion:

• A "refresh" of the UCL safe haven is happening at the moment. Ed would like to get involved in a trial of the new system; there's potential for us to link in there.
• Release of synthetic data models and potential for combining synthetic datasets derived from the same original dataset are hard problems for privacy.
• Data generator - is it safe to release?

And some more notes from others (mainly Kirstie!) in the lab's HackMD document:

• Definition of intensive care: keeping the organs alive while the body heals itself
  • iron lung --> manual ventilator (medical students) --> automated monitoring
  • Lots of data!

• The NIHR Health Informatics Collaborative (HIC) brings together all NHS trusts with NIHR Biomedical Research Centres (BRCs) and their partner trust to make NHS clinical data more readily available to NHS trusts, researchers, industry and the wider NHS community.

  • from https://hic.nihr.ac.uk
  • doi: 10.1016/j.ijmedinf.2018.01.006

• Can't release data and so therefore:
  • can't verify work that has been published
  • barrier of reusing the data for new analyses and greater innovation
• The big secret: anonymisation relies on subjective assessments
  • What is a key variable (erm....everything)
  • What is sensitive (one person's signal is another person's noise)
• Questions:
  • When you say "community standards" what do you mean by that?
  • A: Who gets to answer the question about "what is sensitive"? So community standards is trying to build consensus across a broad coalition of medical professionals and experts in privacy preserving analyses.
  • Could you have a fine grained consent?
  • A: It would be a bit tough... intensive care is exempt from legal requirements for consent as the patients are so unwell when that data is collected. Might be easier in other areas within the medical community.
  • Are there GDPR restrictions?
  • A: Not really - it affected documentation, but not really the actions that are taken around the data.

Ed seems keen to meet the rest of the QUiPP team in a follow-up meeting, so let's try to arrange that sometime soon.

[martintoreilly]

Thanks @LouiseABowler - sounds like it was a super interesting meeting. Let's definitely catch up with Ed as a project team. We could also potentially support Ed's "synth-a-thon" by offering to pair up on hosting a synthetic data workshop. It's something we've thought about doing and in terms of funding we could potentially swap this for one of our planned away days. What do people think?

[gmingas]

Thanks @LouiseABowler - sounds like it was a super interesting meeting. Let's definitely catch up with Ed as a project team. We could also potentially support Ed's "synth-a-thon" by offering to pair up on hosting a synthetic data workshop. It's something we've thought about doing and in terms of funding we could potentially swap this for one of our planned away days. What do people think?

I think joining/hosting the synth-a-thon makes sense. Happy to help with this.

[LouiseABowler]

I've sent Ed a message asking if he'd like to join us for one of our Monday meetings, or another time if that would be inconvenient.

[ots22]

Notes in the comment above now in 'meeting-notes' Issue #136 for further discussion on this dataset

Closing