Having reviewed a bunch of radiology papers lately with the same flaws - some #ML but most not - here are my top 10 tips! #AcademicChatter #radiology #AI
1/ Start with what is already known, where is the gap in knowledge and why it's important. Some papers are doomed from the start because there is no clear impact of the work. Remember the reviewer may not be super-specialised in your research niche.
2/ I need to understand the experimental design quickly, so a flowchart is really helpful. Put yourself in the shoes of a first-time reader. If the design is easy to follow then everything else fits into place.
3/ Similarly, always use a reporting checklist (@EQUATORNetwork) and attach in the supplement. This will enhance the clarity and quality of your paper. Even better if you pre-register your study.
4/ Get professional statistical advice - for study design and power calculation, as well as the analysis itself. You can't salvage a poorly designed study and the wrong tests will lead to bad inferences.
5/ There is a lot of debate about p-values but don't make claims about "trends" only when they suit your pre-conceived ideas. Effects sizes and confidence intervals are more useful.
6/ Show all the data in your plots whether large or small samples. This is the only way to understand the real distribution of your data and the influence of outliers. github.com/UK-Digital-Hea…
7/ Don't make claims that aren't supported by your data. You don't need to over-sell your findings to get published - in fact the reverse is often true! Do however put the findings in a clinical context and sketch out what the next steps should be.
8/ Show me the images! If it's an imaging paper I want to see lots of images - not just derived data. If you are doing a classification task - show a range of correctly and incorrectly classified examples.
9/ If you are reporting the performance of a model - whether for classification or prediction - try to follow best current practice in that domain. Roll on the new #TRIPOD-ML, #SpiritAI and #ConsortAI guidelines!
10/ I hate the phrase "data available on reasonable request". Whether it's logistic regression or complex DL with multiple dependencies there is no reason not to enable others to reproduce your findings - doi.org/10.24433/CO.92… doi.org/10.24433/CO.85….
