User login
Personalized treatment for depression may soon become a reality, thanks to an artificial intelligence (AI) algorithm that accurately predicts antidepressant efficacy in specific patients.
A landmark study of more than 300 patients with major depressive disorder (MDD) showed that a latent-space machine-learning algorithm tailored for resting-state EEG robustly predicted patient response to sertraline. The findings were generalizable across different study sites and EEG equipment.
“We found that the use of the artificial intelligence algorithm can identify the EEG signature for patients who do well on sertraline,” study investigator Madhukar H. Trivedi, MD, professor of psychiatry at the University of Texas Southwestern Medical Center in Dallas, said in an interview.
“Interestingly, when we looked further, it became clear that patients with that same EEG signature do not do well on placebo,” he added.
The study was published online Feb. 10 in Nature Biotechnology (doi: 10.1038/s41587-019-0397-3).
Pivotal study
Currently, major depression is defined using a range of clinical criteria. As such, it encompasses a heterogeneous mix of neurobiological phenotypes. Such heterogeneity may account for the modest superiority of antidepressant medication relative to placebo.
While recent research suggests that resting-state EEG may help identify treatment-predictive heterogeneity in depression, these studies have also been hindered by a lack of cross-validation and small sample sizes.
What’s more, these studies have either identified nonspecific predictors or failed to yield generalizable neural signatures that are predictive at the individual patient level (Am J Psychiatry. 2019 Jan 1;176[1]:44-56).
For these reasons, there is currently no robust neurobiological signature for an antidepressant-responsive phenotype that may help identify which patients would benefit from antidepressant medication. Nevertheless, said Dr. Trivedi, detailing such a signature would promote a neurobiological understanding of treatment response, with the potential for notable clinical implications.
“The idea behind this [National Institutes of Health]–funded study was to develop biomarkers that can distinguish treatment outcomes between drug and placebo,” he said. “To do so, we needed a randomized, placebo-controlled trial that has significant breadth in terms of biomarker evaluation and validation, and this study was designed specifically with this end in mind.
“There has not been a drug-placebo study that has looked at this in patients with depression,” Dr. Trivedi said. “So in that sense, this was really a pivotal study.”
To help address these challenges, the investigators developed a machine-learning algorithm they called SELSER (Sparse EEG Latent Space Regression).
Using data from four separate studies, they first established the resting-state EEG predictive signature by training SELSER on data from 309 patients from the EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response in Clinic Care) study, a neuroimaging-coupled, placebo-controlled, randomized clinical study of antidepressant efficacy.
The generalizability of the antidepressant-predictive signature was then tested in a second independent sample of 72 depressed patients.
In a third independent sample of 24 depressed patients, the researchers assessed the convergent validity and neurobiological significance of the treatment-predictive, resting-state EEG signature.
Finally, a fourth sample of 152 depressed patients was used to test the generalizability of the results.
‘Fantastic’ result but validation needed
These combined efforts were aimed at revealing a treatment responsive phenotype in depression, dissociate between medication and placebo response, establish its mechanistic significance, and provide initial evidence regarding the potential for treatment selection on the basis of a resting-state EEG signature.
The study showed that improvement in patients’ symptoms was robustly predicted by the algorithm. These predictions were specific for sertraline relative to placebo.
When generalized to two depression samples, the researchers also found that the algorithm reflected general antidepressant medication responsivity and related differentially to a repetitive transcranial magnetic stimulation (TMS) treatment outcome.
“Although we only looked at sertraline,” Dr. Trivedi said, “we also applied the signature to a sample of patients who had been treated with transcranial magnetic stimulation. And we found that the signature for TMS [response] is different than the signature for sertraline.”
Interestingly, the antidepressant-predictive signature identified by SELSER was also superior to that of conventional machine-learning models or latent modeling methods, such as independent-component analysis or principal-component analysis. This SELSER signature was also superior to a model trained on clinical data alone and was able to predict outcome using resting-state EEG data acquired at a study site not included in the model training set.
The study also revealed evidence of multimodal convergent validity for the antidepressant-response signature by virtue of its correlation with expression of a task-based functional MRI signature in one of the four datasets.
The strength of the resting-state signature was also found to correlate with prefrontal neural responsivity, as indexed by direct stimulation with single-pulse TMS and EEG.
Given the ability of the algorithm to both predict outcome with sertraline and distinguish response between sertraline and placebo at the individual patient level, the investigators believe SELSER may one day support machine learning–driven personalized approaches to depression treatment.
“Our findings advance the neurobiological understanding of antidepressant treatment through an EEG-tailored computational model and provide a clinical avenue for personalized treatment of depression,” the authors wrote.
Yet, their work is far from over.
“Identifying this signature was fantastic, but you’ve got to be able to validate it as well,” Dr. Trivedi noted. “And luckily, we were able to validate it in the three additional studies.
“The next question is whether it can be broadened to other illnesses.”
Promising research
Commenting on the findings in an interview, Michele Ferrante, PhD, said he believes there may soon be a time during which algorithms such as this are used to personalize depression treatment.
“It’s well known that there are no good biological tests in psychiatry, but promising computational tools, biomarkers, and behavioral signatures for segregating patients according to treatment response are starting to emerge for depression,” said Dr. Ferrante, program chief of the Theoretical and Computational Neuroscience Program at the National Institute of Mental Health (NIMH).
“Precision in the ability to predict what patient will respond to each treatment will improve over time, I have no doubt,” added Dr. Ferrante, who was not involved with the current study.
However, he noted, such approaches are not without their potential drawbacks.
“The greatest challenge is to continuously validate these computational tools as they keep on learning from more heterogeneous groups. Another challenge will be to make sure that these computational tools become well-established, widely adopted, safe, and regulated by the [Food and Drug Administration] as Software as a Medical Device,” he said.
The current algorithm will also need to undergo further testing, said Dr. Ferrante.
“It has been validated on an external dataset,” he said, “but now we need to do rigorous prospective clinical trials where patients are selectively assigned by the AI to a treatment according to their biosignature, to see if these results hold true.
“Down the road, it would be important to implement computational models [that are] able to assign patients across the multiple treatments available for depression, including pharmaceuticals, psychosocial interventions, and neural devices.”
The study was funded directly and indirectly by the NIMH of the National Institutes of Health, the Stanford Neurosciences Institute, the Hersh Foundation, the National Key Research and Development Plan of China, and the National Natural Science Foundation of China.
Dr. Trivedi disclosed numerous financial relationships with pharmaceutical companies and device manufacturers. He has received grants/research support from the Agency for Healthcare Research and Quality, Cyberonic, the National Alliance for Research in Schizophrenia and Depression, the NIMH, and the National Institute on Drug Abuse.
A version of this article first appeared on Medscape.com.
Personalized treatment for depression may soon become a reality, thanks to an artificial intelligence (AI) algorithm that accurately predicts antidepressant efficacy in specific patients.
A landmark study of more than 300 patients with major depressive disorder (MDD) showed that a latent-space machine-learning algorithm tailored for resting-state EEG robustly predicted patient response to sertraline. The findings were generalizable across different study sites and EEG equipment.
“We found that the use of the artificial intelligence algorithm can identify the EEG signature for patients who do well on sertraline,” study investigator Madhukar H. Trivedi, MD, professor of psychiatry at the University of Texas Southwestern Medical Center in Dallas, said in an interview.
“Interestingly, when we looked further, it became clear that patients with that same EEG signature do not do well on placebo,” he added.
The study was published online Feb. 10 in Nature Biotechnology (doi: 10.1038/s41587-019-0397-3).
Pivotal study
Currently, major depression is defined using a range of clinical criteria. As such, it encompasses a heterogeneous mix of neurobiological phenotypes. Such heterogeneity may account for the modest superiority of antidepressant medication relative to placebo.
While recent research suggests that resting-state EEG may help identify treatment-predictive heterogeneity in depression, these studies have also been hindered by a lack of cross-validation and small sample sizes.
What’s more, these studies have either identified nonspecific predictors or failed to yield generalizable neural signatures that are predictive at the individual patient level (Am J Psychiatry. 2019 Jan 1;176[1]:44-56).
For these reasons, there is currently no robust neurobiological signature for an antidepressant-responsive phenotype that may help identify which patients would benefit from antidepressant medication. Nevertheless, said Dr. Trivedi, detailing such a signature would promote a neurobiological understanding of treatment response, with the potential for notable clinical implications.
“The idea behind this [National Institutes of Health]–funded study was to develop biomarkers that can distinguish treatment outcomes between drug and placebo,” he said. “To do so, we needed a randomized, placebo-controlled trial that has significant breadth in terms of biomarker evaluation and validation, and this study was designed specifically with this end in mind.
“There has not been a drug-placebo study that has looked at this in patients with depression,” Dr. Trivedi said. “So in that sense, this was really a pivotal study.”
To help address these challenges, the investigators developed a machine-learning algorithm they called SELSER (Sparse EEG Latent Space Regression).
Using data from four separate studies, they first established the resting-state EEG predictive signature by training SELSER on data from 309 patients from the EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response in Clinic Care) study, a neuroimaging-coupled, placebo-controlled, randomized clinical study of antidepressant efficacy.
The generalizability of the antidepressant-predictive signature was then tested in a second independent sample of 72 depressed patients.
In a third independent sample of 24 depressed patients, the researchers assessed the convergent validity and neurobiological significance of the treatment-predictive, resting-state EEG signature.
Finally, a fourth sample of 152 depressed patients was used to test the generalizability of the results.
‘Fantastic’ result but validation needed
These combined efforts were aimed at revealing a treatment responsive phenotype in depression, dissociate between medication and placebo response, establish its mechanistic significance, and provide initial evidence regarding the potential for treatment selection on the basis of a resting-state EEG signature.
The study showed that improvement in patients’ symptoms was robustly predicted by the algorithm. These predictions were specific for sertraline relative to placebo.
When generalized to two depression samples, the researchers also found that the algorithm reflected general antidepressant medication responsivity and related differentially to a repetitive transcranial magnetic stimulation (TMS) treatment outcome.
“Although we only looked at sertraline,” Dr. Trivedi said, “we also applied the signature to a sample of patients who had been treated with transcranial magnetic stimulation. And we found that the signature for TMS [response] is different than the signature for sertraline.”
Interestingly, the antidepressant-predictive signature identified by SELSER was also superior to that of conventional machine-learning models or latent modeling methods, such as independent-component analysis or principal-component analysis. This SELSER signature was also superior to a model trained on clinical data alone and was able to predict outcome using resting-state EEG data acquired at a study site not included in the model training set.
The study also revealed evidence of multimodal convergent validity for the antidepressant-response signature by virtue of its correlation with expression of a task-based functional MRI signature in one of the four datasets.
The strength of the resting-state signature was also found to correlate with prefrontal neural responsivity, as indexed by direct stimulation with single-pulse TMS and EEG.
Given the ability of the algorithm to both predict outcome with sertraline and distinguish response between sertraline and placebo at the individual patient level, the investigators believe SELSER may one day support machine learning–driven personalized approaches to depression treatment.
“Our findings advance the neurobiological understanding of antidepressant treatment through an EEG-tailored computational model and provide a clinical avenue for personalized treatment of depression,” the authors wrote.
Yet, their work is far from over.
“Identifying this signature was fantastic, but you’ve got to be able to validate it as well,” Dr. Trivedi noted. “And luckily, we were able to validate it in the three additional studies.
“The next question is whether it can be broadened to other illnesses.”
Promising research
Commenting on the findings in an interview, Michele Ferrante, PhD, said he believes there may soon be a time during which algorithms such as this are used to personalize depression treatment.
“It’s well known that there are no good biological tests in psychiatry, but promising computational tools, biomarkers, and behavioral signatures for segregating patients according to treatment response are starting to emerge for depression,” said Dr. Ferrante, program chief of the Theoretical and Computational Neuroscience Program at the National Institute of Mental Health (NIMH).
“Precision in the ability to predict what patient will respond to each treatment will improve over time, I have no doubt,” added Dr. Ferrante, who was not involved with the current study.
However, he noted, such approaches are not without their potential drawbacks.
“The greatest challenge is to continuously validate these computational tools as they keep on learning from more heterogeneous groups. Another challenge will be to make sure that these computational tools become well-established, widely adopted, safe, and regulated by the [Food and Drug Administration] as Software as a Medical Device,” he said.
The current algorithm will also need to undergo further testing, said Dr. Ferrante.
“It has been validated on an external dataset,” he said, “but now we need to do rigorous prospective clinical trials where patients are selectively assigned by the AI to a treatment according to their biosignature, to see if these results hold true.
“Down the road, it would be important to implement computational models [that are] able to assign patients across the multiple treatments available for depression, including pharmaceuticals, psychosocial interventions, and neural devices.”
The study was funded directly and indirectly by the NIMH of the National Institutes of Health, the Stanford Neurosciences Institute, the Hersh Foundation, the National Key Research and Development Plan of China, and the National Natural Science Foundation of China.
Dr. Trivedi disclosed numerous financial relationships with pharmaceutical companies and device manufacturers. He has received grants/research support from the Agency for Healthcare Research and Quality, Cyberonic, the National Alliance for Research in Schizophrenia and Depression, the NIMH, and the National Institute on Drug Abuse.
A version of this article first appeared on Medscape.com.
Personalized treatment for depression may soon become a reality, thanks to an artificial intelligence (AI) algorithm that accurately predicts antidepressant efficacy in specific patients.
A landmark study of more than 300 patients with major depressive disorder (MDD) showed that a latent-space machine-learning algorithm tailored for resting-state EEG robustly predicted patient response to sertraline. The findings were generalizable across different study sites and EEG equipment.
“We found that the use of the artificial intelligence algorithm can identify the EEG signature for patients who do well on sertraline,” study investigator Madhukar H. Trivedi, MD, professor of psychiatry at the University of Texas Southwestern Medical Center in Dallas, said in an interview.
“Interestingly, when we looked further, it became clear that patients with that same EEG signature do not do well on placebo,” he added.
The study was published online Feb. 10 in Nature Biotechnology (doi: 10.1038/s41587-019-0397-3).
Pivotal study
Currently, major depression is defined using a range of clinical criteria. As such, it encompasses a heterogeneous mix of neurobiological phenotypes. Such heterogeneity may account for the modest superiority of antidepressant medication relative to placebo.
While recent research suggests that resting-state EEG may help identify treatment-predictive heterogeneity in depression, these studies have also been hindered by a lack of cross-validation and small sample sizes.
What’s more, these studies have either identified nonspecific predictors or failed to yield generalizable neural signatures that are predictive at the individual patient level (Am J Psychiatry. 2019 Jan 1;176[1]:44-56).
For these reasons, there is currently no robust neurobiological signature for an antidepressant-responsive phenotype that may help identify which patients would benefit from antidepressant medication. Nevertheless, said Dr. Trivedi, detailing such a signature would promote a neurobiological understanding of treatment response, with the potential for notable clinical implications.
“The idea behind this [National Institutes of Health]–funded study was to develop biomarkers that can distinguish treatment outcomes between drug and placebo,” he said. “To do so, we needed a randomized, placebo-controlled trial that has significant breadth in terms of biomarker evaluation and validation, and this study was designed specifically with this end in mind.
“There has not been a drug-placebo study that has looked at this in patients with depression,” Dr. Trivedi said. “So in that sense, this was really a pivotal study.”
To help address these challenges, the investigators developed a machine-learning algorithm they called SELSER (Sparse EEG Latent Space Regression).
Using data from four separate studies, they first established the resting-state EEG predictive signature by training SELSER on data from 309 patients from the EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response in Clinic Care) study, a neuroimaging-coupled, placebo-controlled, randomized clinical study of antidepressant efficacy.
The generalizability of the antidepressant-predictive signature was then tested in a second independent sample of 72 depressed patients.
In a third independent sample of 24 depressed patients, the researchers assessed the convergent validity and neurobiological significance of the treatment-predictive, resting-state EEG signature.
Finally, a fourth sample of 152 depressed patients was used to test the generalizability of the results.
‘Fantastic’ result but validation needed
These combined efforts were aimed at revealing a treatment responsive phenotype in depression, dissociate between medication and placebo response, establish its mechanistic significance, and provide initial evidence regarding the potential for treatment selection on the basis of a resting-state EEG signature.
The study showed that improvement in patients’ symptoms was robustly predicted by the algorithm. These predictions were specific for sertraline relative to placebo.
When generalized to two depression samples, the researchers also found that the algorithm reflected general antidepressant medication responsivity and related differentially to a repetitive transcranial magnetic stimulation (TMS) treatment outcome.
“Although we only looked at sertraline,” Dr. Trivedi said, “we also applied the signature to a sample of patients who had been treated with transcranial magnetic stimulation. And we found that the signature for TMS [response] is different than the signature for sertraline.”
Interestingly, the antidepressant-predictive signature identified by SELSER was also superior to that of conventional machine-learning models or latent modeling methods, such as independent-component analysis or principal-component analysis. This SELSER signature was also superior to a model trained on clinical data alone and was able to predict outcome using resting-state EEG data acquired at a study site not included in the model training set.
The study also revealed evidence of multimodal convergent validity for the antidepressant-response signature by virtue of its correlation with expression of a task-based functional MRI signature in one of the four datasets.
The strength of the resting-state signature was also found to correlate with prefrontal neural responsivity, as indexed by direct stimulation with single-pulse TMS and EEG.
Given the ability of the algorithm to both predict outcome with sertraline and distinguish response between sertraline and placebo at the individual patient level, the investigators believe SELSER may one day support machine learning–driven personalized approaches to depression treatment.
“Our findings advance the neurobiological understanding of antidepressant treatment through an EEG-tailored computational model and provide a clinical avenue for personalized treatment of depression,” the authors wrote.
Yet, their work is far from over.
“Identifying this signature was fantastic, but you’ve got to be able to validate it as well,” Dr. Trivedi noted. “And luckily, we were able to validate it in the three additional studies.
“The next question is whether it can be broadened to other illnesses.”
Promising research
Commenting on the findings in an interview, Michele Ferrante, PhD, said he believes there may soon be a time during which algorithms such as this are used to personalize depression treatment.
“It’s well known that there are no good biological tests in psychiatry, but promising computational tools, biomarkers, and behavioral signatures for segregating patients according to treatment response are starting to emerge for depression,” said Dr. Ferrante, program chief of the Theoretical and Computational Neuroscience Program at the National Institute of Mental Health (NIMH).
“Precision in the ability to predict what patient will respond to each treatment will improve over time, I have no doubt,” added Dr. Ferrante, who was not involved with the current study.
However, he noted, such approaches are not without their potential drawbacks.
“The greatest challenge is to continuously validate these computational tools as they keep on learning from more heterogeneous groups. Another challenge will be to make sure that these computational tools become well-established, widely adopted, safe, and regulated by the [Food and Drug Administration] as Software as a Medical Device,” he said.
The current algorithm will also need to undergo further testing, said Dr. Ferrante.
“It has been validated on an external dataset,” he said, “but now we need to do rigorous prospective clinical trials where patients are selectively assigned by the AI to a treatment according to their biosignature, to see if these results hold true.
“Down the road, it would be important to implement computational models [that are] able to assign patients across the multiple treatments available for depression, including pharmaceuticals, psychosocial interventions, and neural devices.”
The study was funded directly and indirectly by the NIMH of the National Institutes of Health, the Stanford Neurosciences Institute, the Hersh Foundation, the National Key Research and Development Plan of China, and the National Natural Science Foundation of China.
Dr. Trivedi disclosed numerous financial relationships with pharmaceutical companies and device manufacturers. He has received grants/research support from the Agency for Healthcare Research and Quality, Cyberonic, the National Alliance for Research in Schizophrenia and Depression, the NIMH, and the National Institute on Drug Abuse.
A version of this article first appeared on Medscape.com.
FROM NATURE BIOTECHNOLOGY