 |
 |
One more thought/question in regards to correcting for multiple IC maps declared significant from one subject Melodic rs-fMRI run: perhaps, I should divide the -thresholded IC-maps by the number of IC thought to be significant - in this way, FWE is taken into account? Does this make sense? Thank you, Varina From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of Anderson M. Winkler Sent: Friday, April 27, 2012 4:01 AM To: [log in to unmask] Subject: Re: [FSL] Randomise after Melodic Hi Varina, Sorry to beat a discussion so long... I'll try to comment (again, sorry!!) and others may add more... I am reporting the first case of a 4 year old evaluated by rs-fMRI-Melodic, which directly influenced the surgical approach of epileptogenic foci removal. Hmm..., maybe these questions could have been sorted out before the surgery. Certainly the decision was based also on all usual clinical information (history, EEG, failure of anticonvulsants and so on). I'm looking forward to see the report published describing a successful treatment, and note that most of the imaging tools we discuss in research are for research only. Clinical use are at the sole discretion of the doctors. Reporting in the medical literature that "False positive weight was set equivalent to False negative weight" is not heard of, and will be heavily critized, I doubt I can get that published. Folks are expecting p values with confidence intervals to get a feeling if they trust the findings or not. Is there a way to back track to a p value with confidence intervals from the thresholded IC maps? With respect to the paper, I think you'd be more well guarded by describing and citing the probabilistic ICA method, even and specially in a medical journal, instead of coming up with a classical p-value or confidence interval. Consider this: a p-value indicates the probability of observing the tested effect when no effect actually exists, right? Let's look into what happens in ICA: The spatial map for each component tells the weight that the respective component's time-course has on the time course observed for that voxel before the ICA decomposition took place. The z-maps are these weights normalised by the noise level. The noise is what remains after the components have been extracted. In the parts of the brain where the time-course for a given component is not present, the map looks like gaussian noise (or t-distributed noise), whereas where the time-course is present, the weight tends to be substantially different than zero. One might say then that we could take the z-maps and compute p-values, as in classical testing. Would these p-values be correct? Would they have the interpretation as above of being a "probability of observing a similar effect when no effect actually exists"? Nope! Because these maps don't result of either an observation or an experiment which null hypothesis is of "no effect". Instead, they are sought for by the ICA algorithm. In fact, in all maps "something" is present, otherwise the map would have not been identified by the method. This "something" can be simply structured noise, can be artifacts, and can be something very useful and interesting with a clear biological interpretation. It's different than classical testing, where if there is nothing, then most likely nothing will be found. And if something is found, then it is so unlikely, that the p-value is small and we say then it's a significant effect. And why is the mixture-modelling appropriate here? The mixture modelling does not assume simply that "there is nothing". Instead it begins from the principle that "there is something", and the idea is to discriminate the "something" (the signal in some parts of the brain) from the "nothing" (the background noise). How to convince reviewers and readers of the journal then that the statistical method is valid? If you describe correctly what you did and cite the appropriate papers, you should be well guarded in this aspect. These methods have been published in peer-reviewed journals and have been cited 500+ times. Moreover, as I mentioned in a previous message, the method of thresholding isn't conceptually different than balancing sensitivity and specificity (here across voxels), and people should be familiar with these concepts from the lab tests we use in clinical practice. If it helps, maybe these two papers could be cited to support the methods: - http://www.ncbi.nlm.nih.gov/pubmed/9882086: This is about mixture modelling applied to model-based fMRI - http://www.ncbi.nlm.nih.gov/pubmed/14964560: This is specifically about the Probabilistic ICA that uses mixture-modelling for inference. All the best! Anderson