User login
Functional brain connectivity patterns are a stable biomarker of attention-deficit/hyperactivity disorder, new research suggests.
By applying a machine-learning approach to brain-imaging data, investigators were able to identify with 99% accuracy the adult study participants who had been diagnosed with ADHD in childhood.
“Even though the symptoms of ADHD may be less apparent in adulthood, the brain-wiring signature seems to be persistent,” study investigator Christopher McNorgan, PhD, of the department of psychology, State University of New York at Buffalo told this news organization.
The findings were published online Dec. 17, 2020, in Frontiers of Psychology.
Deep-learning neural networks
The researchers analyzed archived functional magnetic resonance imaging (fMRI) and behavioral data for 80 adults (mean age, 24 years; 64 male). Of these participants, 55 were diagnosed with ADHD in childhood and 25 were not.
The fMRI data were obtained during a response inhibition task that tested the individual’s ability to not respond automatically; for example, not saying “Simon Says” after someone else makes the comment.
The behavioral data included scores on the Iowa Gambling Task (IGT), which is used to measure impulsivity and risk taking.
“Usually, but not always, people with ADHD make riskier choices on this task,” Dr. McNorgan noted.
The investigators measured the amount of interconnectedness among different brain regions during the response inhibition task, which was repeated four times.
Patterns of interconnectivity were then fed into a deep-learning neural network that learned which patterns belonged to the ADHD group vs. those without ADHD (control group) and which patterns belonged to the high vs. low scorers on the IGT.
Caveats, cautionary notes
“The trained models are then tested on brain patterns they had never seen before, and we found the models would make the correct ADHD diagnosis and could tell apart the high and low scorers on the IGT 99% of the time,” Dr. McNorgan reported.
“The trained classifiers make predictions by calculating probabilities, and the neural networks learned how each of the brain connections contributes towards the final classification probability. We identified the set of brain connections that had the greatest influence on these probability calculations,” he noted.
Because the network classified both ADHD diagnosis and gambling task performance, the researchers were able to distinguish between connections that predicted ADHD when gambling performance was poor, as is typical for patients with ADHD, and those predicting ADHD when gambling performance was uncharacteristically good.
While more work is needed, the findings have potential clinical relevance, Dr. McNorgan said.
“ADHD can be difficult to diagnose reliably. If expense wasn’t an issue, fMRI may be able to help make diagnosis more reliable and objective,” he added.
However, because individuals with ADHD have different behavioral profiles, such as scoring atypically well on the IGT, additional studies using this approach may help identify brain networks “that are more or less active in those with ADHD that show a particular diagnostic trait,” he said.
“This could help inform what treatments might be more effective for those individuals,” Dr. McNorgan said.
Of course, he added, “clinicians’ diagnostic expertise is still required, as I would not base an ADHD diagnosis solely on the results of a single brain scan.”
No cross-validation
Commenting on the findings for this news organization, Vince Calhoun, PhD, neuroscientist and founding director of the Center for Translational Research in Neuroimaging and Data Science, Atlanta, a joint effort between Georgia State, Georgia Tech, and Emory University, noted some study limitations.
One cautionary note is that the investigators “appear to select relevant regions to include in the model based on activation to the task, then computed the predictions using the subset of regions that showed strong activation. The issue is this was done on the same data, so there was no cross-validation of this ‘feature selection’ step,” said Dr. Calhoun, who was not involved with the research. “This is a type of circularity which can lead to inflated accuracies,” he added.
Dr. Calhoun also noted that “multiple ADHD classification studies” have reported accuracies above 90%. In addition, there were only 80 participants in the current dataset.
“That’s relatively small for making strong claims about high accuracies as has been reported elsewhere,” he said.
Dr. McNorgan and Dr. Calhoun have reported no relevant financial relationships.
A version of this article first appeared on Medscape.com.
Functional brain connectivity patterns are a stable biomarker of attention-deficit/hyperactivity disorder, new research suggests.
By applying a machine-learning approach to brain-imaging data, investigators were able to identify with 99% accuracy the adult study participants who had been diagnosed with ADHD in childhood.
“Even though the symptoms of ADHD may be less apparent in adulthood, the brain-wiring signature seems to be persistent,” study investigator Christopher McNorgan, PhD, of the department of psychology, State University of New York at Buffalo told this news organization.
The findings were published online Dec. 17, 2020, in Frontiers of Psychology.
Deep-learning neural networks
The researchers analyzed archived functional magnetic resonance imaging (fMRI) and behavioral data for 80 adults (mean age, 24 years; 64 male). Of these participants, 55 were diagnosed with ADHD in childhood and 25 were not.
The fMRI data were obtained during a response inhibition task that tested the individual’s ability to not respond automatically; for example, not saying “Simon Says” after someone else makes the comment.
The behavioral data included scores on the Iowa Gambling Task (IGT), which is used to measure impulsivity and risk taking.
“Usually, but not always, people with ADHD make riskier choices on this task,” Dr. McNorgan noted.
The investigators measured the amount of interconnectedness among different brain regions during the response inhibition task, which was repeated four times.
Patterns of interconnectivity were then fed into a deep-learning neural network that learned which patterns belonged to the ADHD group vs. those without ADHD (control group) and which patterns belonged to the high vs. low scorers on the IGT.
Caveats, cautionary notes
“The trained models are then tested on brain patterns they had never seen before, and we found the models would make the correct ADHD diagnosis and could tell apart the high and low scorers on the IGT 99% of the time,” Dr. McNorgan reported.
“The trained classifiers make predictions by calculating probabilities, and the neural networks learned how each of the brain connections contributes towards the final classification probability. We identified the set of brain connections that had the greatest influence on these probability calculations,” he noted.
Because the network classified both ADHD diagnosis and gambling task performance, the researchers were able to distinguish between connections that predicted ADHD when gambling performance was poor, as is typical for patients with ADHD, and those predicting ADHD when gambling performance was uncharacteristically good.
While more work is needed, the findings have potential clinical relevance, Dr. McNorgan said.
“ADHD can be difficult to diagnose reliably. If expense wasn’t an issue, fMRI may be able to help make diagnosis more reliable and objective,” he added.
However, because individuals with ADHD have different behavioral profiles, such as scoring atypically well on the IGT, additional studies using this approach may help identify brain networks “that are more or less active in those with ADHD that show a particular diagnostic trait,” he said.
“This could help inform what treatments might be more effective for those individuals,” Dr. McNorgan said.
Of course, he added, “clinicians’ diagnostic expertise is still required, as I would not base an ADHD diagnosis solely on the results of a single brain scan.”
No cross-validation
Commenting on the findings for this news organization, Vince Calhoun, PhD, neuroscientist and founding director of the Center for Translational Research in Neuroimaging and Data Science, Atlanta, a joint effort between Georgia State, Georgia Tech, and Emory University, noted some study limitations.
One cautionary note is that the investigators “appear to select relevant regions to include in the model based on activation to the task, then computed the predictions using the subset of regions that showed strong activation. The issue is this was done on the same data, so there was no cross-validation of this ‘feature selection’ step,” said Dr. Calhoun, who was not involved with the research. “This is a type of circularity which can lead to inflated accuracies,” he added.
Dr. Calhoun also noted that “multiple ADHD classification studies” have reported accuracies above 90%. In addition, there were only 80 participants in the current dataset.
“That’s relatively small for making strong claims about high accuracies as has been reported elsewhere,” he said.
Dr. McNorgan and Dr. Calhoun have reported no relevant financial relationships.
A version of this article first appeared on Medscape.com.
Functional brain connectivity patterns are a stable biomarker of attention-deficit/hyperactivity disorder, new research suggests.
By applying a machine-learning approach to brain-imaging data, investigators were able to identify with 99% accuracy the adult study participants who had been diagnosed with ADHD in childhood.
“Even though the symptoms of ADHD may be less apparent in adulthood, the brain-wiring signature seems to be persistent,” study investigator Christopher McNorgan, PhD, of the department of psychology, State University of New York at Buffalo told this news organization.
The findings were published online Dec. 17, 2020, in Frontiers of Psychology.
Deep-learning neural networks
The researchers analyzed archived functional magnetic resonance imaging (fMRI) and behavioral data for 80 adults (mean age, 24 years; 64 male). Of these participants, 55 were diagnosed with ADHD in childhood and 25 were not.
The fMRI data were obtained during a response inhibition task that tested the individual’s ability to not respond automatically; for example, not saying “Simon Says” after someone else makes the comment.
The behavioral data included scores on the Iowa Gambling Task (IGT), which is used to measure impulsivity and risk taking.
“Usually, but not always, people with ADHD make riskier choices on this task,” Dr. McNorgan noted.
The investigators measured the amount of interconnectedness among different brain regions during the response inhibition task, which was repeated four times.
Patterns of interconnectivity were then fed into a deep-learning neural network that learned which patterns belonged to the ADHD group vs. those without ADHD (control group) and which patterns belonged to the high vs. low scorers on the IGT.
Caveats, cautionary notes
“The trained models are then tested on brain patterns they had never seen before, and we found the models would make the correct ADHD diagnosis and could tell apart the high and low scorers on the IGT 99% of the time,” Dr. McNorgan reported.
“The trained classifiers make predictions by calculating probabilities, and the neural networks learned how each of the brain connections contributes towards the final classification probability. We identified the set of brain connections that had the greatest influence on these probability calculations,” he noted.
Because the network classified both ADHD diagnosis and gambling task performance, the researchers were able to distinguish between connections that predicted ADHD when gambling performance was poor, as is typical for patients with ADHD, and those predicting ADHD when gambling performance was uncharacteristically good.
While more work is needed, the findings have potential clinical relevance, Dr. McNorgan said.
“ADHD can be difficult to diagnose reliably. If expense wasn’t an issue, fMRI may be able to help make diagnosis more reliable and objective,” he added.
However, because individuals with ADHD have different behavioral profiles, such as scoring atypically well on the IGT, additional studies using this approach may help identify brain networks “that are more or less active in those with ADHD that show a particular diagnostic trait,” he said.
“This could help inform what treatments might be more effective for those individuals,” Dr. McNorgan said.
Of course, he added, “clinicians’ diagnostic expertise is still required, as I would not base an ADHD diagnosis solely on the results of a single brain scan.”
No cross-validation
Commenting on the findings for this news organization, Vince Calhoun, PhD, neuroscientist and founding director of the Center for Translational Research in Neuroimaging and Data Science, Atlanta, a joint effort between Georgia State, Georgia Tech, and Emory University, noted some study limitations.
One cautionary note is that the investigators “appear to select relevant regions to include in the model based on activation to the task, then computed the predictions using the subset of regions that showed strong activation. The issue is this was done on the same data, so there was no cross-validation of this ‘feature selection’ step,” said Dr. Calhoun, who was not involved with the research. “This is a type of circularity which can lead to inflated accuracies,” he added.
Dr. Calhoun also noted that “multiple ADHD classification studies” have reported accuracies above 90%. In addition, there were only 80 participants in the current dataset.
“That’s relatively small for making strong claims about high accuracies as has been reported elsewhere,” he said.
Dr. McNorgan and Dr. Calhoun have reported no relevant financial relationships.
A version of this article first appeared on Medscape.com.