An 'AI' that can diagnose schizophrenia from a brain scan – here's how it works (or doesn't)
Simple model reaches 75% accuracy
Analysis Scientists have had a crack at using simple machine-learning software to make psychiatry a little more objective.
Why, you ask. Well, rather than rely on a professional opinion from a human expert, as we've done so for years, why not ask a computer for a cold logical diagnosis?
One benefit of using code as opposed to a psychiatrist is that its results should be consistent across all patients, whereas your mental health assessment may differ from doctor to doctor. On the other hand, it's claimed this particular piece of software has a 74 per cent success rate when diagnosing people with schizophrenia – meaning it misses more than a quarter of cases – so don't quit medical school just yet.
A team from the Universities of Alberta, Calgary, and Memphis, together with folks from IBM, built the aforementioned program – a computational model that predicts schizophrenia in humans with 74 per cent accuracy – and published their findings in Nature partner journal Schizophrenia in May.
Fast forward to July, and Big Blue, fast and nimble as ever, boasted how AI can help predict schizophrenia. Well, we decided to take a closer look.
Mental disorders like schizophrenia are complex. The cause and the relationship between biology and behavior are still largely unknown. Diagnoses rely on psychiatric tests and results are often subjective, said Russell Greiner, a coauthor of the paper and a professor at the Alberta Machine Intelligence Institute (AMII) at the University of Alberta.
In an attempt to study schizophrenia more objectively, the researchers collected a small dataset of MRI brain scans from 95 test subjects. Forty-six of those folks were diagnosed with schizophrenia by doctors, and the other 49 were healthy. While being scanned, the subjects were asked to perform a simple task that would test their grey matter and thus light up their scans, revealing the organization of the neurons.
To ensure fair results, participants were required to perform the same task during the scan: when they heard an “oddball” tone – a sound that is the odd one out from all the other sounds played – they had to press a button.
The scans thus provided the researchers with a model of the brain for each person. Each model was then broken down into 27,000 voxels, with each voxel representing a small individual three-dimensional space within the brain.
Hotspots ... Arrows point to clusters of brain activity identified in scans (Source)
A linear support vector machine algorithm analyzed the strength of the interconnections between brain regions by looking at what was happening in each voxel. This simple machine-learning code picked out the brain patterns that identified which physiological features are closely associated with schizophrenia. For example, it learned that a strong connection between the thalamus and primary motor cortex areas of the brain was a good predictor of the mental disorder.
This information was then used to train a sparse multivariate regression – a type of classifier algorithm – to determine if someone had schizophrenia or not.
“Medical data is hard to obtain and it’s a small dataset to work with,” Mina Gheiratmand, coauthor of the paper and postdoctoral research fellow at AMII, told The Register.
The researchers carried out their experiments using 94 subjects for training data and left one out for testing. They performed the same process 95 times, each time leaving a different scan out for testing to obtain an average of 74 per cent accuracy across all test runs. In other words, the team picked out 94 of the 95 subjects, told the software which of those brains were diagnosed with schizophrenia, then told it to work out for itself how to identify them, and finally tested its newly learned ability with the remaining subject. Then repeat over and over with different test subjects.
It’s not the best result, but it’s the first step in showing that mental disorders may be able to be diagnosed computationally in the future, Greiner told The Register:
“It’s still very early days, and there are a lot of challenges before something like this can ever be used in clinical settings. We need to try this on different datasets and understand schizophrenia better.”
Guillermo Cecchi, a researcher working in IBM’s Computational Psychiatry and Neuroimaging groups, said: “Schizophrenia is a challenging disease because it cannot be attributed to a single mechanism or a specific area. This is what we try to address by looking at the brain as a network, but both our theories and our experimental methods are limited.
“Moreover, like many other diseases, it is not a ‘single’ instance of the disease, but appears on a spectrum, like is the case for instance with cancer. That means we need a much better characterization of its symptoms, which are multi-dimensional as opposed to binary ... This is what we try to address by inferring the clinical scales from imaging, but this also requires improvements on the process of clinical evaluation itself.” ®