I was at the University College London three years ago, studying machine learning, and a researcher gave us a presentation about a new machine learning technique she was trying (it was applying SVMs to, err, I forget the medical imaging technique. Maybe EKGs). It turned out that she could reliably predict depression in a significant percentage of patients, and severe depression in 100% of patient, or some other similar, extremely high, percentage.
This blew me away, but I only now realized why. It was because I knew depression was symptomatically defined, and she had probably found a way to define it physically. I thought that the "lesser" cases she couldn't detect weren't actually physically depressed, and that her discovery was amazing. Other students didn't share my enthusiasm, but maybe that was because it wasn't very interesting from a machine learning perspective.
I wonder what happened with that. I can't remember her name now.
Neat find. From the reading the abstract it seems that the key finding wasn't so much predicting depression (it just confirmed a correlation with the Hamilton Scale - in real life it would be easier just to use the Hamilton Scale itself, no need for expensive imaging and the subsequent ML analytics).
The key finding seemed to be predicting whether or not treatment would be successful. This is quite powerful as it would help guide clinicians in figuring out whether or not a specific treatment regimen would be advisable or just a waste of time (and time is critical in treating depression).
This blew me away, but I only now realized why. It was because I knew depression was symptomatically defined, and she had probably found a way to define it physically. I thought that the "lesser" cases she couldn't detect weren't actually physically depressed, and that her discovery was amazing. Other students didn't share my enthusiasm, but maybe that was because it wasn't very interesting from a machine learning perspective.
I wonder what happened with that. I can't remember her name now.
EDIT: I found the paper: http://discovery.ucl.ac.uk/1316862/