Mental States in Healthy Cohorts and Epilepsy Patients
This research develops AI systems to monitor mental states (stress, workload, emotions) using wearable brain sensors, moving from lab settings to real-world use and clinical applications in epilepsy patients where stress triggers seizures. The work includes creating real-time emotion detection from EEG data, testing cognitive load monitoring in daily life, and applying these methods to epilepsy care. Additionally, the research develops mathematical models to analyze how anti-seizure medications affect seizure frequency over time, enabling more personalized treatment strategies.
Contact: Sidratul Moontaha
Human mental states such as stress, workload, and emotions strongly affect health, safety, and performance, with consequences ranging from burnout and cardiovascular risks to impaired driving and seizures in epilepsy patients. Reliable, real-time monitoring of these states with wearable sensors, particularly in daily-life settings, can provide broad societal benefits and critical clinical insights. Therefore, my research addresses these challenges in three steps: first, developing pipelines for real-time emotion recognition in laboratory conditions (EEGEMO); second, extending these systems to everyday stress and workload detection with healthy participants (UNIVERSE); and third, translating them to epilepsy patients, where stress is a major seizure trigger (EPIStress).
In parallel, I also investigate the need for robust seizure count models to better understand and predict long-term clinical outcomes. Here, I develop algorithms that quantify how different medication dosages influence seizure frequency, providing a foundation for more personalized treatment strategies
EEGEMO
Wearable EEG-based Emotion Classification in Real-time
This project presents research efforts to address the need for the transition from research-based, offline EEG analysis to real-world, real-time emotion classification systems. Since understanding human emotions is essential for affective computing applications, many works are available on emotion recognition based on various physiological measures. With the recent availability of wearable EEG devices, emotion recognition systems can be enriched by directly measuring brain activity with increased ease over earlier methods. However, data scarcity from wearable EEG devices is a meaningful obstacle, with the additional challenges of online classification in realistic settings where data is generated continuously. Live data streams are inherently non-stationary and pose several computational challenges. This is particularly true for EEG data from wearable devices, which contain numerous artifacts, high data velocity, and exhibit evolving data patterns.
Therefore, addressing these challenges, within this project, we collected physiological signals from two wearable EEG devices, Muse S and Neurosity Crown, and subjective assessments from 15 participants while undergoing emotion elicitation experiments. The EEG data is streamed to develop a real-time emotion classification pipeline. Our system can reliably detect valence and arousal levels from the EEG data from static data streams. The pipeline also shows the potential to implement it in live data streams by a small cohort experiment. The dataset description, results, and instructions for accessing the dataset are published here.
UNIVERSE
From Controlled to In-Situ Cognitive Load Detection
Along with emotional monitoring, understanding cognitive demands, such as mental workload or cognitive stress, is also a critical component of user-centered systems for developing adaptive technologies in healthcare (e.g., stress-sensitive seizure monitoring), education (e.g., intelligent tutoring systems, transportation (e.g., driver fatigue detection, air traffic control), workplace ergonomics (e.g., workload of performance-evaluated work), and mental health interventions (e.g., biofeedback). Therefore, this project extends our investigation into mental state detection by focusing on mental workload and stress, emphasizing generalizability beyond the laboratory setup.
The work begins with a feasibility study by curating data from wearable EEG from 11 participants in controlled conditions while inducing cognitive stress. The results show the ability to distinguish between high and low cognitive loads using data from low-cost wearable EEG devices and the participants' subjective feelings. The experimental paradigm, developed using PsychoPy, is robust enough to collect data from multiple wearable devices and various experimental paradigms. The publication and codes are available for other researchers to design similar experiments.
Building upon the feasibility study's findings, this chapter presents a more extensive experimental setup with multi-modal sensors involving 24 participants performing four distinct tasks, each with two difficulty levels in random order. Notably, the comprehensive study was conducted across two laboratory environments and was complemented by in Situ data collection. The results are published in a conference (in proceedings), and the dataset and descriptor are publicly available for other researchers.
EPIStress
Cognitive Stress Profiling in Epilepsy Patients
This project extends mental state detection to a clinical context by focusing on epilepsy patients. Epilepsy is a neurological condition where cognitive stress and emotional factors are frequently reported by patients as seizure triggers. Therefore, building on the mental workload detection framework proposed in the previous project (Universe) with the control group, this project extends the analysis of mental state detection to individuals with epilepsy. Physiological data in the lab and hospital settings, along with patient-reported seizure triggers, were collected during the study to detect cognitive stress in this population and examine patterns that might inform future personalized seizure risk forecasting systems. The dataset is publicly available for the researchers.
Seizure Count Time Series Modeling
This proposed method in this project leverages state space models combined with Kalman filtering tailored for discrete event data, modeling the daily number of seizures experienced by epilepsy patients. In practical use, the model incorporates time-dependent dosages of multiple anti-seizure medications (ASMs) as external control inputs, allowing the quantification of their effects—including potential superposition and delay effects—on seizure occurrence rates. This approach helps address the complexity arising from polytherapy, unexplained influences, and temporal correlations in seizure risk that are not explained solely by medication, offering insights into individual patient responses and underlying seizure dynamics. State space models also demonstrate robustness to typical observation errors in seizure diaries—missed seizures, misclassified events, and missing data—making them highly suitable for real-world clinical settings. The use of state space models in seizure count analysis provides a statistically powerful and flexible framework for examining medication effectiveness, tracking changes over time, and designing patient-specific interventions. Compared to traditional regression or non-dynamic models, the state space approach captures the stochastic and dynamic nature of seizure generation in the presence of therapeutic interventions and other unknown factors. The framework also allows objective comparison of ASM efficacy in individual patients, enabling data-driven clinical decisions. This modeling approach represents a significant advance for practical seizure count time series analysis, bridging the gap between sophisticated statistical modeling and the clinical realities of epilepsy management.