https://m.medicalxpress.com/news/2018-10-machine-uncovers-dementia-subtypes-implication.html Just thought this was interesting in relation to how subtypes are being found in other illness.
This is potentially interesting for us as people do not have problems with the idea of dementia (a symptom based diagnosis) even though there are a wide range of underlying different biomedical deficits and causes. So why does this biomedical heterogeneity in relation to people with ME cause problems in recognising the existence of it as a significant clinically relevant syndrome. [Where are the doctors and psychologists/psychiatrists trying to rename dementia as chronic confusion syndrome? How is it not seen as a problem when doctors diagnose dementias such as Alzheimer's on the basis of symptoms rather than brain scans, but the lack of clinical biomarkers for ME is seen as problematic?]
Having quickly skimmed the paper I don't understand what they are doing. They are using an unsupervised ML technique but combining this with disease models and taking into account the temporal stages of the disease. So it does look like something interesting. Features appear to be from MRI scans. A simple unsupervised technique for subtyping is hard to use for subtyping because it will find different clusters but that could be noise or have many other factors that effect the data (disease progression, life-style, reporting issues etc). This means it is very hard to understand whether any clusters are just a feature of the data or other factors or truly relate to subtypes. This appears to be where using disease models will help (but I didn't get a feel for how they were doing this). They also tried to use genetic factors as a model of ground truth in trying to verify the model - which is interesting but I may worry it could become a self fulfilling prophecy.
The machine learning is just a fancy way of finding patterns right? So, it's only going to detect patterns in the data that are already in the data? I ask, because I read something about this approach in policing, and all it seemed to do was amplify the bias that was in the data from the start. (If the police data is biased, and targets are then chosen from the data, and policies built around the data, all you did was create a more efficient bias mechanism)
