The paper that caught my eye today is a paper discussing the application of machine learning to the diagnosis of Schizophrenia by looking at electroencephalogram (EEG) data. The article title is Auditory prediction errors as individual biomarkers of schizophrenia, and is written by a group of mainly Australian researchers, including two colleagues of mine, Natasha Matthews and Pat Michie.
Since it's first descriptions by physicians and psychiatrists in the late 1800's and the coining of the term schizophrenia in 1908 by Eugen Bleuler the precise definition has been in flux. The current (DSM-5) guidelines for diagnosis specify at least 6 months of ongoing psychopathology, including some (but not necessarily all) of the following aspects:
As these symptoms can vary in degree and aspect, and the fact that a full diagnosis requires assessment over a long time period by a highly qualified individual then the cost of service provision is very expensive. This, combined with the struggle to get people suffering from psychotic symptoms or lack of motivations from seeing said professional means the ability to diagnose (or at least assist diagnosis) with an automated tool would be of great value, freeing up the time for professionals to focus on interventions and getting treatment earlier.
The diagnosis of schizophrenia is based on self report measures, and while previous studies suggest the validity of schizophrenia diagnoses is approximately 90%, the fact that the judgement is necessarily subjective in the end means the discovery of a reliable biomarker to help inform these decisions would be of great value. There have been several biomarker alternatives proposed for schizophrenia over the years focussing on genetic, MRI and EEG measures, and one of the more robust results within the schizophrenia population is the reduction in mismatch negativity (MMN) EEG waveforms, which is the brains response to a surprising (deviant) stimulus after the build up of a prediction based on a regular or common (standard) stimulus. The predominant form of paradigm that elicits this response is an auditory task (an oddball task)
The goals of the study were to develop a predictive model that aids in the diagnosis of schizophrenia, assessing the performance of different oddball tasks, classification algorithms and analysis strategies.
A little over 20 schizophrenia patients and matched controls were recruited to participate in the study, with patients diagnoses being verified by way of the Diagnostic Interview for Psychosis and controls were verified to be free of significant psychopathology using the same tool.
Three different oddball (MMN eliciting) paradigms were run for all participants. In one the deviant (oddball) sounds were longer in duration, the second had deviants with a short gap in the middle of the sound and the final task had deviants where the sound presented to one ear was slightly behind the other ear (making the sound appear to be coming from the left or right). Each of these tasks have historically shown strong results in schizophrenia populations.
The extracted EEG features and clinical measures that were acquired were then fed into a number of machine learning algorithms to test for binary classification of subjects between the Schizophrenia/Control groups as well as investigating quantifiable relationships between neuronal responses and the severity of symptoms.
The classification of subjects into the schizophrenia/control groups was able to be performed with slightly over 80% accuracy - yielding better results than similar approaches of machine learning to classification of schizophrenia in the past - perhaps due to the fact that the whole spatio-temporal EEG signal was input into the classifiers, as opposed to just certain extracted features. There was a higher accuracy of 90% reported on a previous study, however they used fMRI data for there classification, which is an expensive diagnostic tool.
However whilst the results are promising, one fact that is raised is that schizophrenia is not a binary condition (otherwise healthy controls may have mild schizotypal symptoms, and in general symptoms vary a lot in terms of expression and severity), therefore the training of a binary classifier raises some concerns.
Whilst this study provides some incentive to do more research, especially in a much larger dataset of people who have just been diagnosed (before long term treatment), it does not provide a definitive solution to the problem of diagnosis in schizophrenia. Another aspect I would be wary of is that MMN reduction is not limited to people with Schizophrenia, so extensive research including these conditions would probably be appropriate to tease out any schizophrenia specific features (if such exist).
