Hi all,
I’m curious to know if you have some recommendations of papers that you think did a good methodological work with biomarker research.
Of course there is Frank’s great tutorial, but was curious to see actual published papers that you think used sound methods. Or also any further great resources like Frank’s.
Also really liked checking this discussion on biomarker research that evolved here in DataMethods.
I’m currently working in a field with limited/none therapies and pharma does basically all trial efforts, so a good deal of academia is focused on biomarker research with observational studies. It’s quite frustrating seeing ROC-curve/categorization domination, and it’s sometimes hard to convince others that there are better ways to quantify the added value of a continuous biomarker, especially when I’m still not very confident about these alternative ways myself.
If you have some suggestions on both diagnostic and prognostic studies it’d be great.
Thanks,
Thanks for referencing our earlier discussion - it has since made it into a paper, but I can’t say it was easy to convince reviewers (or co-authors) to use some of the methods we explored in the conversation.
As my head is thoroughly in the cardiovascular space, I can see to a certain degree why categorisation dominates - the clinician wants an easy to understand model, that helps in daily practice. The concept of continuous biomarkers has not been implemented in most risk assessments/pathways with any great success. The exception would probably be Rick Body’s T-MACS probability calculation of a patient suffering an acute coronary syndrome. Then there’s Abbott’s MI3 score using hs-cTnI, age and gender (I think) with heavy weighting towards the biomarker. Again, not something I’ve seen frequently employed in clinical practice.
Yeah, I can also super understand why categorization dominates, in the end people want to know if something is useful or not, and areas like cardiovascular really have critical decision-making aspects in which biomarkers can help. I work with dementia biomarkers, and quite often during biomarker discovery and validation, group-wise unadjusted comparisons and ROC curve comparisons are used to assess novel biomarkers, how good they are to provide a cutoff to discriminate those with/without underlying brain pathology etc. I mean, if the biomarker is super good, these analyses will show it one way or the other, but as more and more biomarkers become available, further subject-of-matter info and methods can come in handy (like Frank explores in his tutorial).
It was a great discussion, is this the paper that came out of it?
Exactly, part of the analysis was driven by that conversation.
How is brain pathology defined in your studies - imaging based such as MRI findings? Or cognitive function scores?
Yeah, it ended up super nice, showing the normogram, when dealing with thresholds showing more than just for one, etc.
In dementia studies, especially Alzheimer’s, underlying brain pathology was traditionally determined with neuropath at autopsy, and now PET and cerebrospinal fluid - and even plasma - biomarkers are being used. Then it’s sort of a loop of validating one against the available hierarchically superior one (e.g. PET vs autopsy; CSF vs autopsy; plasma vs PET/CSF and so on).
I mean, of course there are plenty of alternatives. Even a single test-based ROC can be replaced and better interpreted by inspecting the biomarker’s functional form in a logistic regression model using a spline term if the case is really to keep just with a Y ~ x1 interpretation.
But just to stress the point, I’m curious to see if people have more good general examples of biomarker studies that were more carefully thought about, like yours, beyond just “how not to” papers or a specific modeling doubt.