The effect of brainwave-based feedback on anxiety and stress in college students

Approved 05/01/2023, Medical Research Ethics Committee of Hefei Fourth People's Hospital (316 Huangshan Road, Shushan District, Hefei, 230026, China; +86 (0)551 6361 6193; 2514331322@qq.com), ref: HFS-IRB-PJ-WL (2022002)

Study 1:
Prediction of cortisol and α-amylase levels using machine learning models based on electroencephalogram features.

Study 2:
1. Change in amplitude of anxiety-specific Event-Related Potential (ERP) components (e.g., N2, P3, or Late Positive Potential (LPP)) during the emotional regulation task.
Timepoint: From baseline (pre-intervention) to immediately after the completion of the closed-loop training sessions.
2. Change in high-frequency beta wave power during the emotional regulation task.
Timepoint: From baseline (pre-intervention) to immediately after the completion of the closed-loop training sessions.

Study 1:
1. Performance comparison of different machine learning models in predictive tasks: Specific metrics include accuracy, which evaluates the performance of multilayer perceptrons, support vector machines, and linear discriminant analysis in predicting elevated/low cortisol and α-amylase levels.
2. Importance contribution of EEG features to predictive models: Quantification of relative contributions from different frequency bands (e.g., Delta, Theta, Alpha, Beta, Gamma) to cortisol and α-amylase prediction using SHAP values. Quantification of relative contributions from different brain regions (e.g., right prefrontal cortex, occipital cortex, parietal cortex) to predicting these biomarker levels through SHAP values.
3. Regression prediction accuracy of machine learning models for cortisol and α-amylase levels: Evaluation metrics include R² (coefficient of determination) and mean absolute error to measure models' ability to predict continuous physiological parameter values.
4. Differences in EEG feature model prediction effectiveness across time points (first and second MIST tasks): Assess whether temporal factors during stress task interventions affect the classification performance of EEG feature-based models.

Study 2:
1. Change in self-reported anxiety scores as measured by standardized anxiety scales (e.g., GAD-7, SAS).
Timepoint: From baseline (pre-intervention) to immediately after the completion of the training sessions.
2. Change in behavioral performance during the emotional regulation task, including:
Reaction time to positive and negative stimuli.
Accuracy of emotion identification.
Timepoint: From baseline to post-intervention.
3.Change in scores from the closed-loop training task (where a higher score indicates better ability to suppress negative emotions and greater emotional stability).
Timepoint: Measured after each training block and compared across the training course.
4.Changes in other EEG/ERP components not listed as primary outcomes, such as P1, N1, and Error-Related Negativity (ERN/ΔERN), and their correlation with anxiety reduction.
Timepoint: From baseline to post-intervention.
5.Participant's subjective experience and strategy use, assessed through post-training questionnaires or interviews, regarding:
5.1. Attention concentration level.
5.2. Cognitive load.
5.3. Emotional state.
5.4. Fatigue level.
5.5. Self-regulation strategies developed.
Timepoint: After each training session or at the end of the entire intervention.
6.Performance and generalizability of the predictive EEG-based anxiety diagnostic model, evaluated by:
The model's accuracy, sensitivity, and specificity in classifying or predicting anxiety states.
Timepoint: Upon completion of data collection and analysis.