Can Brain MRIs Diagnose and Classify Migraine?

WASHINGTON, DC—Accurate classifiers for the diagnosis and subclassification of migraine have been developed using brain MRI data, reported Todd J. Schwedt, MD, at the 57th Annual Meeting of the American Headache Society. “We’ve built multivariate models of brain cortical thickness, cortical surface areas, and regional volumes that are pretty accurate for classifying individual people with migraine as having either chronic migraine versus episodic migraine,” said Dr. Schwedt. “We have also built pretty accurate classifiers for differentiating individual people as having chronic migraine versus being a healthy control. However, we didn’t have as much success with classifying episodic migraine versus healthy controls.”

In presenting his research, Dr. Schwedt had one important clarification. “I want to be very clear that I am not suggesting that we move towards a time where we should be using brain MRI to make the diagnosis of migraine,” he said. “Obviously that would be wasteful of medical resources. That’s not what we’re suggesting here.” But his team has developed objective imaging classifiers that may help to optimize clinical diagnostic criteria, much of which is currently consensus-generated and symptom-based, and may have other clinical utility, for example in a situation where it is difficult to make the right diagnosis in real time or distinguish between headache types with similar presentations.

An Unmet Need
“It would be great if we had biomarkers that would predict natural migraine progression,” said Dr. Schwedt, who is a Professor of Neurology at the Mayo Clinic in Scottsdale, Arizona. If a biomarker could predict who is likely to transition into a more severe state or who is more likely to revert to a less severe state, that information could inform treatment decisions. “It would be great to have biomarkers that could predict treatment responses within individuals. That would be very welcomed by our patients. It also would be great to have biomarkers that could give us early signals of likely treatment responses. That would really reduce the time it takes to develop new therapies and reduce the expense associated with that.” And finally, Dr. Schwedt said, it would be great to have biomarkers that could diagnose and subclassify headache disorders objectively.

To tackle this problem, Dr. Schwedt and his research colleagues used structural imaging data to build classifiers for migraine. They used measures of cortical surface area and cortical thickness and measures of volume from different regions of the brain to develop accurate classifiers that can identify, on an individual patient level, whether a brain MRI belongs to someone who has chronic migraine or episodic migraine, or to a healthy control. The researchers also sought to test the current threshold of 15 headache days per month for differentiating chronic from episodic migraine.

Building Migraine Models
Dr. Schwedt and research colleagues from the Mayo Clinic Arizona and Arizona State University enrolled adult participants between the ages of 18 and 65. Participants had either episodic or chronic migraine according to ICHD2 criteria. They also enrolled healthy controls. Participants were excluded if they were using migraine prophylactic medications or opiates, met criteria for medication overuse, had any acute or chronic pain conditions other than migraine, or if their brain MRI was abnormal according to typical clinical definitions.

Scans were done at two institutions, “the reason being that I started this study while I was at Washington University in St. Louis and then continued it at Mayo Clinic in Arizona,” explained Dr. Schwedt. Both scanners were Siemens 3-T scanners. “It is important to understand that nearly equal proportions of migraineurs and healthy controls were imaged on each of the two scanners. This really reduces the potential for scanner bias,” Dr. Schwedt added.

After collecting the MRIs, the researchers made the structural measurements with FreeSurfer, which is freely available software. “It allowed us to get 68 regional measurements of cortical thickness, 68 regional measurements of cortical surface area, as well as volume measurements in 68 different regions.” Overall, there were 204 structural measurements per individual. “Because we were looking at a lot of measurements for each participant, we decided to use a technique of dimension reduction called principal component analysis. We built principal components that accounted for 85% of the variability in each of the three structural measures—area, thickness, and volume,” Dr. Schwedt said. To build the migraine models, the researchers added principal components to the overall classifier until adding another one wouldn’t improve the accuracy by greater than 1%.

Next, the researchers performed tenfold cross validation and assessed classification accuracy within each of 10 runs, with data from 90% of participants randomly selected for classifier development and data from the remaining 10% of participants used to test classification performance. The average classification accuracy of those 10 runs was considered the overall accuracy.

Dr. Schwedt and colleagues enrolled 120 subjects into their study—66 migraineurs and 54 healthy controls. Subjects and controls were well matched for age and sex. There were statistical differences in Beck Depression Inventory scores, although the average score of 4.1 among migraineurs was still within normal limits. The migraineurs averaged nine headache days per month, had had 16 years of migraine, and had significant disability from migraine, with an average MIDAS score of 20.5.

Testing the Models
The first classifier that was built was intended to differentiate migraineurs from healthy controls. “What we found was that overall, the average accuracy of the classifier was 68%. If we looked at an individual brain MRI, based on the structural measures, we could tell with 68% accuracy whether or not that MRI belonged to somebody who had migraine versus someone who was a healthy control,” reported Dr. Schwedt. But 68% accuracy was a little disappointing. The researchers theorized that heterogeneity within the migraine group might explain their results. So they separated out episodic and chronic migraine.

The next classifiers were an attempt to differentiate episodic migraine from healthy controls. “Unfortunately, our accuracy did not improve. It was 67%.”

But when they looked at chronic migraine, the accuracy shot up. With this classifier, the researchers could look at an individual brain MRI and tell whether it belonged to someone with chronic migraine versus a healthy control with 86% accuracy.

Next, the researchers built classifiers to differentiate chronic migraineurs from episodic migraineurs. “Again, we had pretty high accuracy—84%.”

Then the researchers probed the headache frequency threshold. They used different numbers of headache frequency to divide the migraineurs into two subsets, high frequency and low frequency. The thresholds ranged from five headache days per month to 15 headache days per month. For each of those thresholds, the researchers built classifiers based upon brain structure to see which cut-point yielded the most accurate classification, based upon brain structure. “And what we found was that, in fact, 15 headache days per month was the best headache frequency threshold for differentiating the migraineurs into two subgroups. And that’s consistent with our current use of 15 headache days per month to divide migraineurs into episodic versus chronic migraine,” Dr. Schwedt said. “We now have some objective data based upon brain structure that actually supports the use of the 15 headache days per month cut-point.”

Brain Regions Involved
The brain regions that most frequently contributed to the classifiers include the temporal pole, superior temporal lobe, anterior cingulate cortex, entorhinal cortex, medial orbital frontal gyrus, and the pars triangularis. “Each of these regions participates in pain processing, but they participate in different ways,” Dr. Schwedt noted. Some are more important for emotional response to pain, some play a more cognitive role in pain processing, and others are more important for multisensory integration. “So it is kind of a complex picture,” he said. “And this is really what we should expect.”

—Glenn S. Williams

WASHINGTON, DC—Accurate classifiers for the diagnosis and subclassification of migraine have been developed using brain MRI data, reported Todd J. Schwedt, MD, at the 57th Annual Meeting of the American Headache Society. “We’ve built multivariate models of brain cortical thickness, cortical surface areas, and regional volumes that are pretty accurate for classifying individual people with migraine as having either chronic migraine versus episodic migraine,” said Dr. Schwedt. “We have also built pretty accurate classifiers for differentiating individual people as having chronic migraine versus being a healthy control. However, we didn’t have as much success with classifying episodic migraine versus healthy controls.”

In presenting his research, Dr. Schwedt had one important clarification. “I want to be very clear that I am not suggesting that we move towards a time where we should be using brain MRI to make the diagnosis of migraine,” he said. “Obviously that would be wasteful of medical resources. That’s not what we’re suggesting here.” But his team has developed objective imaging classifiers that may help to optimize clinical diagnostic criteria, much of which is currently consensus-generated and symptom-based, and may have other clinical utility, for example in a situation where it is difficult to make the right diagnosis in real time or distinguish between headache types with similar presentations.

An Unmet Need
“It would be great if we had biomarkers that would predict natural migraine progression,” said Dr. Schwedt, who is a Professor of Neurology at the Mayo Clinic in Scottsdale, Arizona. If a biomarker could predict who is likely to transition into a more severe state or who is more likely to revert to a less severe state, that information could inform treatment decisions. “It would be great to have biomarkers that could predict treatment responses within individuals. That would be very welcomed by our patients. It also would be great to have biomarkers that could give us early signals of likely treatment responses. That would really reduce the time it takes to develop new therapies and reduce the expense associated with that.” And finally, Dr. Schwedt said, it would be great to have biomarkers that could diagnose and subclassify headache disorders objectively.

To tackle this problem, Dr. Schwedt and his research colleagues used structural imaging data to build classifiers for migraine. They used measures of cortical surface area and cortical thickness and measures of volume from different regions of the brain to develop accurate classifiers that can identify, on an individual patient level, whether a brain MRI belongs to someone who has chronic migraine or episodic migraine, or to a healthy control. The researchers also sought to test the current threshold of 15 headache days per month for differentiating chronic from episodic migraine.

Building Migraine Models
Dr. Schwedt and research colleagues from the Mayo Clinic Arizona and Arizona State University enrolled adult participants between the ages of 18 and 65. Participants had either episodic or chronic migraine according to ICHD2 criteria. They also enrolled healthy controls. Participants were excluded if they were using migraine prophylactic medications or opiates, met criteria for medication overuse, had any acute or chronic pain conditions other than migraine, or if their brain MRI was abnormal according to typical clinical definitions.

Scans were done at two institutions, “the reason being that I started this study while I was at Washington University in St. Louis and then continued it at Mayo Clinic in Arizona,” explained Dr. Schwedt. Both scanners were Siemens 3-T scanners. “It is important to understand that nearly equal proportions of migraineurs and healthy controls were imaged on each of the two scanners. This really reduces the potential for scanner bias,” Dr. Schwedt added.

After collecting the MRIs, the researchers made the structural measurements with FreeSurfer, which is freely available software. “It allowed us to get 68 regional measurements of cortical thickness, 68 regional measurements of cortical surface area, as well as volume measurements in 68 different regions.” Overall, there were 204 structural measurements per individual. “Because we were looking at a lot of measurements for each participant, we decided to use a technique of dimension reduction called principal component analysis. We built principal components that accounted for 85% of the variability in each of the three structural measures—area, thickness, and volume,” Dr. Schwedt said. To build the migraine models, the researchers added principal components to the overall classifier until adding another one wouldn’t improve the accuracy by greater than 1%.

Next, the researchers performed tenfold cross validation and assessed classification accuracy within each of 10 runs, with data from 90% of participants randomly selected for classifier development and data from the remaining 10% of participants used to test classification performance. The average classification accuracy of those 10 runs was considered the overall accuracy.

Dr. Schwedt and colleagues enrolled 120 subjects into their study—66 migraineurs and 54 healthy controls. Subjects and controls were well matched for age and sex. There were statistical differences in Beck Depression Inventory scores, although the average score of 4.1 among migraineurs was still within normal limits. The migraineurs averaged nine headache days per month, had had 16 years of migraine, and had significant disability from migraine, with an average MIDAS score of 20.5.

Testing the Models
The first classifier that was built was intended to differentiate migraineurs from healthy controls. “What we found was that overall, the average accuracy of the classifier was 68%. If we looked at an individual brain MRI, based on the structural measures, we could tell with 68% accuracy whether or not that MRI belonged to somebody who had migraine versus someone who was a healthy control,” reported Dr. Schwedt. But 68% accuracy was a little disappointing. The researchers theorized that heterogeneity within the migraine group might explain their results. So they separated out episodic and chronic migraine.

The next classifiers were an attempt to differentiate episodic migraine from healthy controls. “Unfortunately, our accuracy did not improve. It was 67%.”

But when they looked at chronic migraine, the accuracy shot up. With this classifier, the researchers could look at an individual brain MRI and tell whether it belonged to someone with chronic migraine versus a healthy control with 86% accuracy.

Next, the researchers built classifiers to differentiate chronic migraineurs from episodic migraineurs. “Again, we had pretty high accuracy—84%.”

Then the researchers probed the headache frequency threshold. They used different numbers of headache frequency to divide the migraineurs into two subsets, high frequency and low frequency. The thresholds ranged from five headache days per month to 15 headache days per month. For each of those thresholds, the researchers built classifiers based upon brain structure to see which cut-point yielded the most accurate classification, based upon brain structure. “And what we found was that, in fact, 15 headache days per month was the best headache frequency threshold for differentiating the migraineurs into two subgroups. And that’s consistent with our current use of 15 headache days per month to divide migraineurs into episodic versus chronic migraine,” Dr. Schwedt said. “We now have some objective data based upon brain structure that actually supports the use of the 15 headache days per month cut-point.”

Brain Regions Involved
The brain regions that most frequently contributed to the classifiers include the temporal pole, superior temporal lobe, anterior cingulate cortex, entorhinal cortex, medial orbital frontal gyrus, and the pars triangularis. “Each of these regions participates in pain processing, but they participate in different ways,” Dr. Schwedt noted. Some are more important for emotional response to pain, some play a more cognitive role in pain processing, and others are more important for multisensory integration. “So it is kind of a complex picture,” he said. “And this is really what we should expect.”

—Glenn S. Williams

WASHINGTON, DC—Accurate classifiers for the diagnosis and subclassification of migraine have been developed using brain MRI data, reported Todd J. Schwedt, MD, at the 57th Annual Meeting of the American Headache Society. “We’ve built multivariate models of brain cortical thickness, cortical surface areas, and regional volumes that are pretty accurate for classifying individual people with migraine as having either chronic migraine versus episodic migraine,” said Dr. Schwedt. “We have also built pretty accurate classifiers for differentiating individual people as having chronic migraine versus being a healthy control. However, we didn’t have as much success with classifying episodic migraine versus healthy controls.”

In presenting his research, Dr. Schwedt had one important clarification. “I want to be very clear that I am not suggesting that we move towards a time where we should be using brain MRI to make the diagnosis of migraine,” he said. “Obviously that would be wasteful of medical resources. That’s not what we’re suggesting here.” But his team has developed objective imaging classifiers that may help to optimize clinical diagnostic criteria, much of which is currently consensus-generated and symptom-based, and may have other clinical utility, for example in a situation where it is difficult to make the right diagnosis in real time or distinguish between headache types with similar presentations.

An Unmet Need
“It would be great if we had biomarkers that would predict natural migraine progression,” said Dr. Schwedt, who is a Professor of Neurology at the Mayo Clinic in Scottsdale, Arizona. If a biomarker could predict who is likely to transition into a more severe state or who is more likely to revert to a less severe state, that information could inform treatment decisions. “It would be great to have biomarkers that could predict treatment responses within individuals. That would be very welcomed by our patients. It also would be great to have biomarkers that could give us early signals of likely treatment responses. That would really reduce the time it takes to develop new therapies and reduce the expense associated with that.” And finally, Dr. Schwedt said, it would be great to have biomarkers that could diagnose and subclassify headache disorders objectively.

To tackle this problem, Dr. Schwedt and his research colleagues used structural imaging data to build classifiers for migraine. They used measures of cortical surface area and cortical thickness and measures of volume from different regions of the brain to develop accurate classifiers that can identify, on an individual patient level, whether a brain MRI belongs to someone who has chronic migraine or episodic migraine, or to a healthy control. The researchers also sought to test the current threshold of 15 headache days per month for differentiating chronic from episodic migraine.

Building Migraine Models
Dr. Schwedt and research colleagues from the Mayo Clinic Arizona and Arizona State University enrolled adult participants between the ages of 18 and 65. Participants had either episodic or chronic migraine according to ICHD2 criteria. They also enrolled healthy controls. Participants were excluded if they were using migraine prophylactic medications or opiates, met criteria for medication overuse, had any acute or chronic pain conditions other than migraine, or if their brain MRI was abnormal according to typical clinical definitions.

Scans were done at two institutions, “the reason being that I started this study while I was at Washington University in St. Louis and then continued it at Mayo Clinic in Arizona,” explained Dr. Schwedt. Both scanners were Siemens 3-T scanners. “It is important to understand that nearly equal proportions of migraineurs and healthy controls were imaged on each of the two scanners. This really reduces the potential for scanner bias,” Dr. Schwedt added.

After collecting the MRIs, the researchers made the structural measurements with FreeSurfer, which is freely available software. “It allowed us to get 68 regional measurements of cortical thickness, 68 regional measurements of cortical surface area, as well as volume measurements in 68 different regions.” Overall, there were 204 structural measurements per individual. “Because we were looking at a lot of measurements for each participant, we decided to use a technique of dimension reduction called principal component analysis. We built principal components that accounted for 85% of the variability in each of the three structural measures—area, thickness, and volume,” Dr. Schwedt said. To build the migraine models, the researchers added principal components to the overall classifier until adding another one wouldn’t improve the accuracy by greater than 1%.

Next, the researchers performed tenfold cross validation and assessed classification accuracy within each of 10 runs, with data from 90% of participants randomly selected for classifier development and data from the remaining 10% of participants used to test classification performance. The average classification accuracy of those 10 runs was considered the overall accuracy.

Dr. Schwedt and colleagues enrolled 120 subjects into their study—66 migraineurs and 54 healthy controls. Subjects and controls were well matched for age and sex. There were statistical differences in Beck Depression Inventory scores, although the average score of 4.1 among migraineurs was still within normal limits. The migraineurs averaged nine headache days per month, had had 16 years of migraine, and had significant disability from migraine, with an average MIDAS score of 20.5.

Testing the Models
The first classifier that was built was intended to differentiate migraineurs from healthy controls. “What we found was that overall, the average accuracy of the classifier was 68%. If we looked at an individual brain MRI, based on the structural measures, we could tell with 68% accuracy whether or not that MRI belonged to somebody who had migraine versus someone who was a healthy control,” reported Dr. Schwedt. But 68% accuracy was a little disappointing. The researchers theorized that heterogeneity within the migraine group might explain their results. So they separated out episodic and chronic migraine.

The next classifiers were an attempt to differentiate episodic migraine from healthy controls. “Unfortunately, our accuracy did not improve. It was 67%.”

But when they looked at chronic migraine, the accuracy shot up. With this classifier, the researchers could look at an individual brain MRI and tell whether it belonged to someone with chronic migraine versus a healthy control with 86% accuracy.

Next, the researchers built classifiers to differentiate chronic migraineurs from episodic migraineurs. “Again, we had pretty high accuracy—84%.”

Then the researchers probed the headache frequency threshold. They used different numbers of headache frequency to divide the migraineurs into two subsets, high frequency and low frequency. The thresholds ranged from five headache days per month to 15 headache days per month. For each of those thresholds, the researchers built classifiers based upon brain structure to see which cut-point yielded the most accurate classification, based upon brain structure. “And what we found was that, in fact, 15 headache days per month was the best headache frequency threshold for differentiating the migraineurs into two subgroups. And that’s consistent with our current use of 15 headache days per month to divide migraineurs into episodic versus chronic migraine,” Dr. Schwedt said. “We now have some objective data based upon brain structure that actually supports the use of the 15 headache days per month cut-point.”

Brain Regions Involved
The brain regions that most frequently contributed to the classifiers include the temporal pole, superior temporal lobe, anterior cingulate cortex, entorhinal cortex, medial orbital frontal gyrus, and the pars triangularis. “Each of these regions participates in pain processing, but they participate in different ways,” Dr. Schwedt noted. Some are more important for emotional response to pain, some play a more cognitive role in pain processing, and others are more important for multisensory integration. “So it is kind of a complex picture,” he said. “And this is really what we should expect.”

—Glenn S. Williams