lwjohnst86
commented
6 years ago Ideas on not just implementation but also the project's mission and "end-product".
What's the point?:
- More data is being generated every day, and yet scientific findings are published as articles, in text form. They are not linked in any physical or computational way. There is no way to meaningfully or adequately get a sense of the entire scientific knowledge in an area without having to read through hundreds, maybe thousands of articles.
- We need some way to aggregrate these results... link them together so that they are a whole, a single "physical" (as stored on a computer) object.
- Scientific studies all have pros and cons, not just in study design, but also in the study population.. they are often very different and hard to completely translate to the wider public.
- We need some way to determine how generalizable results are in the wider context,
- Sometimes, researchers want to see, before running a study, how particular variables or biomarkers associate with a disease. Or maybe policy makers or public health officials or citizen scientists want to see how changing one thing will influence health in a quantifiable and real way, not just "increases risk", etc. How much risk? And in an interactive, digital, quantitative way (e.g. through a web interface or running a computer program).
- We need some way to simulate possible scenarios, influenced by real-world data.
What will it look it?:
- The simulation will contain at least two components:
- A "store" of real world data/scientific results to draw from and inform the simulation. This is the aggregate of as much scientific knowledge as possible. This could be the regression models or model results of a study or something similar...
- Infrastructure to use the real-world data to generate random simulations of possible scenarios and what is most likely.
How will it be used?:
- Researchers could use this as a "first pass resource" when exploring or learning about current knowledge in the area. Using the first component, they can get a sense of what is currently known.
- Researchers could use the simulation to identify areas that are under researched or that have not been researched
- Policy makers, public health officials, the media could use these simulations to test how certain things or conditions could influence health and base their decisions on these results
- Researchers could use this as a way explore possible research questions and run pilot studies to see how their results compare to the simulated results.
Implementation:
(Came from discussion with @danielwitte)
- Some possible ways to implement these:
- Traditional MC models, using baseline values that continuously predict possible future values.
- Typical health economics models?
- Trajectory based, as in predicted values will depend not just on baseline values, but also on how variables operate as trajectories (e.g. weight over time).
- Adversarial neural networks. Have a "generator" who continually creates random
"correlation matrices" (or something similar) and a "discriminator" who chooses
which of a mix of real data/results vs the fake data/results. They compete in
a zero-sum game, until the "discriminator" has a harder time guessing real
vs fake. (this one seems the most promising).
- This method could be generalizable to several situations: simulating from child to adult or from a population at any given time point and followed over time
- This will learn from new data/results that come in, continually building a database of "fake worlds" that improve over time.
Write in this issue ideas for implementing and developing this project. This is brainstorming so nothing is final here.