Diagnosis by machine

One of my scientific interests is the use of EEG or fMRI for diagnosis of neurological disorders in individual patients. So far, most studies (including my own) have been able to find statistically significant differences between groups of patients or between patients and healthy subjects. But, as not everyone realizes, these results unfortunately do not always imply that the measure under scrutiny provides a valuable diagnostic tool. For this to be true, the measure needs to have high sensitivity and high specificity, as well. In other words, if the measure does not link abnormal values to patients and normal values to healthy subjects in most of the cases, it is of no clinical value. Now, for applications in psychiatry an interesting result has been obtained by Khodayari-Rostamabad and co-authors from McMaster University in Hamilton (CA), which was published in Current Biology. They show that machine learning algorithms can be used for an automated diagnostic procedure, employing a selection procedure of the most relevant features derived (statistically) from an individual’s EEG. This is a very powerful technique, which only works however, if the training dataset is large enough (in their case they used a training dataset of 207 subjects, entailing three psychiatric disorders and data from healthy subjects, yielding 85% correct diagnoses after training and cross-validation experiments). I would love to do such a study employing EMG-fMRI in patients with different Parkinsonisms or tremors. All I have to do is find the funding …