Everyone experiences stress at some point, but it can often be invisible until it becomes overwhelming. For people recovering from substance use disorders or who struggle with mental illness, moments of overwhelming stress can have life-altering consequences.
Garvita Jain, a computer science master’s student at the University of Maryland, Baltimore County, is working to change that by combining brain-computer interface (BCI) technology, artificial intelligence, and cognitive behavioral therapy (CBT) into a wearable system designed to recognize stress as it happens and respond immediately.
It was during her coursework that Jain became inspired to create this system. In one course with her advisor, she learned about how BCI technology can be used to help people with mental health issues and how stress can be a contributing factor for people struggling with substance use disorder or mental health issues.
In another class about AI and machine learning where she learned about a headset that could detect physiological signals that indicate stress in a person’s body. She realized that using those signals along with AI, she could detect and monitor stress in real time, potentially allowing for interventions to happen sooner.
“Continuous monitoring could provide a solution early on when the stress is detected, rather than having to wait a week to see a professional,” Jain said. “That’s when I thought it would be good to utilize this technology to help people with mental health issues and substance use disorders.”
Traditional mental health assessments often rely on people recognizing and communicating their own stress levels. But during periods of intense stress, that isn’t always possible.
“Current assessment methods require the person to report how they’re feeling,” Jain said. “If a person is going through a relapse, they may not be able to report it at that very moment because they’re stressed and they don’t know what to ask for or how to ask for help.”
Jain’s solution replaces self-reporting with physiological data.
Garvita Jain
Using a commercially available Muse 2 headset, the system continuously monitors electroencephalogram (EEG) signals from the brain and photoplethysmography (PPG) signals related to heart activity. These biological signals are analyzed by AI that Jain trained to recognize different levels of stress.
“When we’re stressed, our heart rate goes up,” Jain said. “With EEG signals, alpha waves are suppressed when we’re under stress. These types of signals helped me determine whether a person was stressed.”
After the headset captures EEG and PPG signals, they are transmitted to a computer where the data is analyzed and AI then classifies the user’s stress level.
If stress is detected, the computer sends a signal to a mobile application called SeeBT, developed by another student in Jain’s research lab, that uses AI and cognitive behavioral therapy to help people when they are stressed. Its original design required users to initiate conversations manually. Jain modified the software so the intervention began automatically when stress was detected.
“The application triggers an AI conversation with the person so it can help mitigate stress,” Jain said. “I wanted the application to take action so that the person did not have to initiate the conversation.”
Building the system required more than simply collecting physiological data.
Jain first verified that the complete communication pipeline worked by using intentional eye blinks as surrogate triggers while wearing the headset. A single blink activated a mild stress notification, two blinks triggered moderate stress, and three blinks triggered severe stress.
Once the system reliably activated the mobile application, attention shifted to actual stress detection.
Jain collected her own EEG and PPG data across eight experimental sessions featuring alternating periods of rest and cognitively demanding tasks. Jain induced stress on herself by using progressively difficult arithmetic problems and the well-known Stroop word-color task, in which Jain had to identify the color of a word rather than reading the word itself.
Behavioral data, including response accuracy and completion time, were also incorporated into the model to improve classification.
“I didn’t have to tell the system how stressed I felt,” Jain said. “Self-reporting can sometimes be biased for many reasons.”
After preprocessing the data, including synchronizing EEG and PPG sampling rates, she then evaluated three machine learning algorithms, settling on the Support Vector Machine model because it provided the best combination of accuracy and real-time performance.
Although the project remains in its early stages, one moment convinced Jain that the concept could work.
While Jain was solving challenging math problems designed to induce stress, the system detected changes in physiological signals almost immediately.
“I would do some math questions, and suddenly the application would say, ‘Oh, you are stressed.’ Then I would relax for a bit, and the stress level would go down again,” Jain said. “Seeing it happen in real time was incredible.”
Throughout the project, Jain said she enjoyed learning about a range of fields, from neuroscience and psychology to artificial intelligence and computer science. She also learned a lot about stress.
“Everything I learned about stress, especially the physiological signals, was new to me,” Jain said. “It was really interesting to see how these subtle signals reflect how we’re feeling.”
She also learned from her own experience that this system could help a range of people who struggle to communicate what they’re feeling.
“As someone who is an introvert, it’s difficult for me to talk about these kinds of things,” Jain said. “If this technology can help people who also find it difficult to express how they’re feeling, it could make a real difference.”
While building a system that can identify stress is a significant milestone, the long-term goal is much broader.
Jain envisions a clinically validated system capable of delivering meaningful interventions at exactly the right moment, particularly for individuals recovering from substance use disorders.
“If an application could intervene in that moment, it could help stop the relapse,” she said.
Jain said that the next phase of the project will involve recruiting additional participants to improve the machine learning models and validate the system across diverse users. Because physiological signals vary from person to person, larger datasets will be essential before the technology can be deployed in real-world settings.
Jain also plans to collaborate with clinicians to strengthen the therapeutic component of the application.
“The long-term objective is not just to detect stress, but to help mitigate it as well,” Jain said.
Although Jain plans to pursue a career in industry after she graduates this year, she hopes to continue advancing technologies that bridge AI and healthcare.
“I want to work in a field where I can use the lessons learned and technology that I’ve built to create something that’s useful for people,” Jain said.
She believes the greatest promise of AI lies not simply in analyzing data, but in recognizing moments when someone needs help and responding before it’s too late.
“I hope to provide help when it is needed— not afterward, but at the exact moment it is needed,” Jain said. “That is one of the major impacts of this work and using this technology.”