Machine Learning May Improve Schizophrenia Diagnosis

Liam Davenport

August 04, 2017

A machine learning approach that uses artificial intelligence (AI) to examine millions of brain cortical links can identify schizophrenia cases with a high degree of accuracy and predict symptom severity, new research shows.

In a brain imaging analysis, investigators led by Irina Rish, PhD, IBM T. J. Watson Research Center, Yorktown Heights, New York, found that, through use of alterations in connections between areas such as the thalamus and the primary cortex, schizophrenia patients could be distinguished from control persons with almost 75% accuracy.

In addition, results revealed that the severity of symptoms such as inattentiveness and bizarre behavior could be predicted from brain scans, although the specific connectivity changes were in different areas.

The study was published online May 16 in Schizophrenia.

Towards Reliable Biomarkers

This study, the investigators write, "represents a step towards finding more reliable objective neuroimaging biomarkers for diagnosing schizophrenia, which have higher reproducibility and generalization accuracy compared to the potential 'biomarkers' reported in association studies," which are typically derived from univariate statistical tests.

Noting that objective measurements have been emphasized as part of the recently proposed Research Domain Criteria (RDoC) initiative from the National Institutes of Mental Health, the researchers believe that their sparse multivariate regression approach to identifying predictors of symptom of severity "is likely to be extremely important in the move towards incorporating the RDoC approach to schizophrenia."

Schizophrenia is typically considered a disease of disrupted brain connectivity. However, the identification of pathologic neuroimaging patterns that are specific to schizophrenia and are related to symptom severity is a challenge. Such patterns could serve as biomarkers for the condition.

To investigate further, the researchers analyzed data from the Function Biomedical Informatics Research Network on structural and task-based whole-brain fMRI scans performed in 95 individuals.

They included 46 patients with DSM-IV-defined schizophrenia or schizoaffective disorder and 49 age- and sex-matched healthy control persons from multiple sites, yielding a total of 380 scans. The patients had no major medical illnesses and had no history of head injury or alcohol/substance dependence. The control persons had no history of major neurologic or psychiatric illness.

The researchers used the Functional Magnetic Resonance Imaging of the Brain Software Library to analyze a total of 26,949 brain voxels, after exclusions, at the voxel and lower-resolution supervoxel level. At 137 time points per sample, this totaled more than 3.6 million variables. The researchers focused on fMRI data collected while participants performed the Auditory Oddball Task.

Using a filter-based approach for identifying features that could discriminate between patients and control persons as well as an analysis of feature stability, the team generated 95 data subsets.

The investigators applied sparse multivariate regression to whole-brain functional connectivity features and leave-one-subject-out cross-validation to evaluate the performance of the model. They found that link-weight features achieved an accuracy of 73.2% in distinguishing between patients and control persons, using 30 features. The accuracy increased to 74.0% using 1024 features.

Specifically, there was a significant increase in connections in schizophrenia patients in comparison with control persons for links connecting the thalamus with the primary cortex, and those connecting the precuneous (Brodmann Area [BA] 7) with the thalamus, BA9, BA44 (inferior frontal gyrus) and the putamen.

High Accuracy, Statistical Significance

In predicting symptom severity, the researchers found that link-weight features predicted scores on the attention, severity of bizarre behavior, and positive formal thought disorder subscales of the Scales for the Assessment of Positive/Negative Symptoms with relatively high accuracy and statistical significance.

For attention, the best model included 129 stable links across 48 nodes in the frontal, temporal, parietal, and occipital lobes, as well as the white matter and cerebellum. For the severity of bizarre behavior subscale, there were nine stable links in areas such as BA6 and BA7.

Scores on the positive formal thought disorder subscale were associated with 11 stable links, primarily spanning BA6 and BA8 on the right and left hemispheres, as well as BA7, the primary motor cortex (BA1), BA22, BA39, and BA40.

"Curiously," the investigators write, "the most predictive links" for distingishing patients from control persons "were not among the most significantly different links between the patient and control groups computed on the full dataset."

Citing the example of thalamocortical connections being involved in classification but not severity prediction, they explain: "The feature selection methods in classification versus regression were different, however, and it would be interesting to use a similar feature selection approach in both tasks before comparing the links involved in classification versus scale prediction."

Speaking to Medscape Medical News, Dr Rish said she believes machine learning could help pin down neuroimaging biomarkers for a range psychiatric conditions and help overcome problems such as the high degree of between-patient variability and overlap between conditions.

"I definitely believe that machine learning and AI can help, especially if you combine it with multisite studies and more larger-scale studies, which is one the critical issues in brain imaging," she said.

"The classical approach was small studies here and there, with different tasks given to different patients. That's why it was very nonhomogeneous, but the recent efforts in the community focus on putting larger datasets together.

"This is a very good direction because we can then use machine learning and AI to learn more reproducible, more stable biomarkers that hopefully will generalize across larger populations," she added.

Dr Rish believes that one of the most important uses of machine learning will not necessarily be in diagnosing patients but in predicting symptom severity.

However, she noted this approach could take up to 10 years to reach the clinic.

"That is much more challenging [but] is something that may, if taken further, help us better define criteria for diagnosis of schizophrenia, not based on existing guidelines like DSM, but based more on objective measurements.

"Maybe we can see how people cluster...and maybe we can somewhat redefine the boundaries for disorders and subtypes of disorders."

The research was supported by IBM (Watson Research Center, US) through the IBM Alberta Center for Advanced Studies, as well as through fellowships from Alberta Innovates Health Solutions and grants from the Natural Sciences and Engineering Research Council of Canada. Internal funding was provided buy Amii (formerly, the Alberta Innovates Center for Machine Learning). The authors have disclosed no relevant financial relationships.

Schizophrenia. Published online May 16, 2017. Full text


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.