If you have cancer, diabetes, or even a bad cold, your condition can be monitored and quantified to a good extent. That data can inform a treatment choice, and the dosage can be adjusted based on the data quantifying the body’s response. Whether it’s a parent with a thermometer and children’s Tylenol or a tumor board determining an oncology regimen, such a data-driven approach is usually the best practice in 21st-century healthcare. Unfortunately, mental health care lags behind, without as much ability to diagnose and treat with informed precision.
My research attempts to address the need. In my lab, we develop AI approaches and brain-computer interfaces that could enable more personalized and accurate care.
Mental disorders such as major depression are a leading cause of disability worldwide, affecting tens of millions of people. Despite this, unfortunately, currently available treatments can fail a large proportion of patients—for example about 20-30% of people with major depression. How can treatments be improved for such patients?
Current scientific evidence suggests that mental disorders can reflect abnormal activity across networks of brain regions. If so, can we measure abnormal brain activity to track how symptom states evolve over time? Then, can we regulate the activity and these symptoms?
These are the questions that my lab is investigating as one possible way to improve treatment. But as you can imagine, it’s an extraordinary problem. While AI and machine learning (ML) are excellent for exploring large problems, the brain is one of the most complex systems there is. Also, obtaining data from the brain is challenging, leading to the need for AI methods that can address data scarcity.
There are three main challenges that we have to address toward achieving the above goals. How can we objectively measure symptoms, how do we quantify the effect of treatment on them, and how can we precisely tailor a treatment over time to regulate symptoms?
The first challenge: How can mood symptoms be measured from brain activity?
In our work, we first objectively track and decode a person’s mood symptoms from their brain signals. We want this decoding to provide objective, repeatable measures that can then be used in precisely tailoring a treatment to a patient’s own needs.
But how do we know how a mood symptom is represented in EEG? We develop ML models to interpret brain signals and relate them to symptoms. We have used these models to show successful tracking of mood symptom variation based on brain activity1. In the future, this approach may be augmented with other physiological measures, such as heart rate, or extended to other mental states, such as anxiety and chronic pain.
The second challenge: How do we know how a treatment dosage affects the brain signals and symptoms?
How does changing the dosage of a treatment change the response in a patient? This is the second challenge, and it is as important as the first. Even if we know a patient’s current symptom level, that is not enough. We also need to know what treatment dosage they need to make them feel better. Currently, we’re investigating the effect of deep brain stimulation (DBS) as one possible treatment modality, although our ML approaches are not limited to this.
We have been able to use a rigorous ML approach to predict how the brain will respond to different doses of DBS2. While the response to a given treatment may differ from person to person, our work can model the response in each individual. This approach could thus make it possible to personalize treatment over time for better results.
Due to the brain’s plasticity and ability to adapt, the effect of a treatment can also change over time. It is important to take such change into account when considering the ongoing dose of a treatment. Adaptive learning algorithms can make that possible, which is what we are also studying.
The third challenge: How do we adjust the treatment over time for better results?
The ultimate objective after solving the above two challenges is to adjust the dosage of treatment over time to precisely tailor it to a patient’s needs. We are creating model-based closed-loop systems to make this goal possible. Our decoders in the first challenge will measure the current symptom level. Then, the dose-response models in the second challenge will tell us how much dosage we need to regulate the symptom level toward a healthy state.
Said differently, in this third challenge, we put the knowledge and ML models from the first two challenges into action for better treatment outcomes. The closed-loop systems we develop can constitute a new generation of brain-computer interfaces for diverse mental disorders3.
It is very important to emphasize that this research is still in early days and so far has focused only on patients for whom no other available therapies work. We are demonstrating the potential for model-based approaches with AI to enable personalized mental health therapies.
This potential is not just limited to DBS but may one day extend to optimizing psychotherapy or pharmacotherapy interventions. There is much work ahead on ensuring safety and efficacy, and on development and validation of these approaches. Realizing such personalized therapies will take a truly interdisciplinary effort between psychiatrists, engineers, computer scientists, neuroscientists, neurosurgeons, behavioral scientists, and bioethicists.
My initial training as an engineer focused on information theory and AI. But I got fascinated by the brain during my Ph.D. studies and decided to use these computational tools to understand the brain and help develop treatments for brain disorders. The potential direct impact of such work on the betterment of people’s lives is what drives me each day.
Maryam M. Shanechi is a neuroengineer. As Dean’s Professor in the Viterbi School of Engineering at the University of Southern California (USC) and a member of the Neuroscience Graduate Program, she is the Founding Director of the USC Center for Neurotechnology. She received a B.A.Sc. degree in Engineering Science from the University of Toronto, S.M. and Ph.D. degrees in Electrical Engineering and Computer Science from MIT, and postdoctoral training in neural engineering and neuroscience at Harvard Medical School and UC Berkeley. Her research is at the intersection of AI, engineering, and neuroscience to develop closed-loop neurotechnology through decoding and control of neural dynamics. Her awards include the NIH Director’s New Innovator Award, NSF CAREER Award, ONR Young Investigator Award, ASEE’s Curtis W. McGraw Research Award, MIT Technology Review’s Top 35 Innovators Under 35, Popular Science Brilliant 10, Science News SN10, One Mind Rising Star Award, and a DoD Multidisciplinary University Research Initiative (MURI) Award. She was named a 2023 Blavatnik National Awards Finalist.