Hacker News new | past | comments | ask | show | jobs | submit login

Pretty great article. This characteristic that many patients have multiple disorders is an interesting point. Reminded me of my real world PCA applications which were never satisfying. In such a case you'd take the attributes of answers to questions in a population, and then get principle components (PCs) to these vectors of examples. Who says that each example in a population has to be dominated by 1 principle component? Many people have the view that PCs are just whatever linear combinations of attributes that explain your data, and I've seen many people try to name them in ways that encapsulate the attributes that each PC selects for, but it always seems like an awkward exercise after you get past the first few PCs.

And of course, I'm not literally saying this SCID is the attributes to the DSM "PCA," or any other subspace method. I don't work with subspace methods often, so if anyone wants to fill me in on what is done in the medical world (or other high stakes domains) in practice, much appreciated.

In cases of segmenting customers and then targeting each one differently, I have found that subspace methods work well for raising typical startup metrics. But the interpretability was not cut and dry in my experience, something that obviously is needed when you work with individuals like Doctors, not populations like many companies.

My guess is that the DSM segments patients to different treatments though. If this is the case then it does't really matter what the name of the constellation is as long as after they send the patient there, everything is hunky dory.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: