The Center for Data Innovation spoke with Archana Venkataraman, assistant professor of electrical and computer engineering at Johns Hopkins University. Venkataraman discussed how machine learning can help increase our understanding of complex neurological disorders.
This interview has been lightly edited.
Joshua New: You run the Neural Systems Analysis (NSA) Lab, which aims to use machine learning to gain a greater understanding of neurological disorders like epilepsy. Just how little do we understand these disorders?
Archana Venkataraman: We know almost nothing about these disorders beyond what we can observe behaviorally in patients. We do not know what causes autism, schizophrenia, major depression, Alzheimer’s disease, and the like. We do not know whether the factors are mostly genetic or mostly environmental. We do not have a good picture of what processes are being disrupted in the brain. And with very few exceptions, we do not have good treatment options. In fact, neurology and psychiatry are two massive fields of heuristics and trial-and-error when it comes to treating patients.
But this is not to say that we aren’t trying. There is plenty of amazing research being done in this space, from basic biology to genetics to imaging. However, the problem is just so complex that none of these individual efforts seem to be making an impact. This is where my lab come in. We are tackling these problems from an engineering standpoint by designing new algorithmic tools to parse and combine different types of data to get a more complete picture of these disorders.
New: Why is machine learning, more so than other data-driven approaches, useful for this kind of research?
Venkataraman: There are two benefits of using machine learning techniques. First, they can extract complex patterns from data that humans cannot see with their naked eye. The brain is incredibly complex, and so are many neurological disorders. Machine learning can help us overcome this complexity by going beyond isolated phenomena. Second, these methods are usually optimized for out-of-sample prediction. Ultimately, this is what we want, to be able to tailor and predict outcomes for each patient.
Now, these days, machine learning or artificial intelligence has become synonymous with deep learning. While my lab does use some deep learning methods, we must be careful about it. Deep learning is great when you have lots of training data and don’t mind a black-box solution. Our application is the complete opposite. We have limited training data, and we need interpretable algorithms to design treatments.
New: You developed a seizure-detection algorithm to track where and when seizures occur in patients’ brain. Why is this useful to understand?
Venkataraman: Epilepsy is one of the most prevalent neurological disorders worldwide; however, 20 to 40 percent of patients do not respond to anti-seizure medications. The only hope for these patients is if we can pinpoint what part of the brain is triggering the seizures. If so, doctors can surgically remove this area or implant a device to eliminate the seizures. Right now, seizure detection and localization are manual processes that require a trained clinician to scan through hours of electroencephalography (EEG) recordings or 3D magnetic resonance imaging (MRI) volumes. This process is time consuming and prone to human error. In fact, only 50 percent of patients who go for surgery remain seizure free after two years. My lab is designing new machine learning algorithms to automatically detect and localize seizures. We hope that this information will allow clinicians to make more informed decisions to improve outcomes in epilepsy.
New: Another one of your projects involves linking the manifestation of neurological disorders to altered neural communication patterns. What is the goal of this research?
Venkataraman: This project is more aligned with basic research than clinical translation. Neurological disorders are complex, as I mentioned earlier, and we now believe that they have system-level influences that bridge multiple areas of the brain. The goal of our work is to automatically identify which neural pathways in the brain are affected by a given neurological disorder. We are also using this approach to predict clinical severity and treatment outcomes on a patient-specific basis.
New: The NSA Lab also began work recently on an initiative to create an autism therapy that involves manipulating emotional cues in human speech. Can you describe this research?
Venkataraman: This is my pet project. It has nothing to do with my primary field of medical imaging. The idea is simple: people with autism have difficulty grasping social and emotional cues, particularly in speech. So, what if we could use machine learning to automatically amplify emotional cues in speech, to the point that someone with autism could perceive them? If so, this technology could form the basis for an assistive device or therapy. Now, it turns out no one knows how to manipulate emotions in speech. So perhaps misguidedly, I decided to tackle this problem. My lab has collected one of the largest annotated emotional speech databases and we have developed two emotion morphing strategies that are inspired by medical imaging. Our preliminary work outperforms state-of-the-art methods from the speech community.
I don’t know where this work will lead or how long it will continue, but it’s been a fun ride so far.