First of all, the basic stuff:
Get somebody who’s never seen your study before to read through the abstract - to ensure fundamental information isn't missing.
(And no, you won't notice yourself, because you’re too concerned with whether you can squeeze in another AUC value ...)
Get somebody who’s never seen your study before to read through the abstract - to ensure fundamental information isn't missing.
(And no, you won't notice yourself, because you’re too concerned with whether you can squeeze in another AUC value ...)
If you are submitting to a radiotherapy conference, maybe make clear what the relevance is for radiotherapy? Several image analysis / radiomics / AI abstracts were probably technically excellent, but I scored them low due to lack of radiotherapy relevance
For ESTRO, you are allowed figures. Use them! Really! Why would you write a whole abstract about dose-response relationships, radiomic features, image processing, etc, without illustrating even a small part of this?
For the nitty-gritty stuff:
If you report a prognostic model, make sure to actually define your outcome measure. Yes, some of you didn't ...
And if you are reporting on a late endpoint - survival, tumour control, late tox, etc - I want to know your follow-up (time)
If you report a prognostic model, make sure to actually define your outcome measure. Yes, some of you didn't ...
And if you are reporting on a late endpoint - survival, tumour control, late tox, etc - I want to know your follow-up (time)
State your statistical methods. Some of you forgot to note which regression model you used ... for a prognostic modelling abstract, I kind of need to know this!
Also: That you 'corrected for clinical factors' just make me wonder if you even know what clinical factors are
Also: That you 'corrected for clinical factors' just make me wonder if you even know what clinical factors are
“We extracted imaging biomarkers / radiomic features / etc” is not a sufficient methods description for a radiomics abstract. Especially if you don’t tell us which volume you extracted features from. By now, I expect radiomics studies to conform to IBSI
Further on standardization and guidelines:
It takes >35 characters to state that you followed the TRIPOD guidelines.
(You all followed the TRIPOD guidelines, right? Right?? All the previous deficiencies notwithstanding??)
tripod-statement.org
It takes >35 characters to state that you followed the TRIPOD guidelines.
(You all followed the TRIPOD guidelines, right? Right?? All the previous deficiencies notwithstanding??)
tripod-statement.org
Some of you also need reminding that
'univariate' -> 'multivariate'
still isn't a robust way to select your predictors.
And neither is forward / backward / stepwise variable selection.
Please.
towardsdatascience.com/stopping-stepw…
sciencedirect.com/science/articl…
'univariate' -> 'multivariate'
still isn't a robust way to select your predictors.
And neither is forward / backward / stepwise variable selection.
Please.
towardsdatascience.com/stopping-stepw…
sciencedirect.com/science/articl…
However, to end on a positive note:
1) Some excellent model performance reporting this year (not just AUC, yay!)
2) And multiple studies with external validation cohorts (more yay!)
So all in all, we're getting better, as a field 😀
1) Some excellent model performance reporting this year (not just AUC, yay!)
2) And multiple studies with external validation cohorts (more yay!)
So all in all, we're getting better, as a field 😀
And I'm now definitely looking forward to seeing you all - and your posters and presentations - in Vienna in April! 
