Decoding Metacognitive Sensitivity from EEG using Deep Learning.
Metacognition, the ability to think about one's thinking processes, is vital for
professional performance, academic achievement, and mental health. However, its
ambiguous nature and subjective measurement techniques across various fields have
posed significant challenges to research. Cognitive neuroscience offers a unique
solution by providing objective measurements that link metacognition to brain activity,
thereby establishing a ground truth. Recently, the convergence of explainable artificial
intelligence (XAI) and perceptual decision-making, a subsection of metacognition
within cognitive neuroscience, has led to the development of the WaveFusion
framework. This innovative framework holds the potential to contribute to the
unification of the fragmented metacognition research fields.
The aim of this thesis was to enhance the WaveFusion framework, an explainable
deep learning model, to classify metacognitive sensitivity and confidence using EEG
data. The objectives were (1) to achieve a classification accuracy of 95% for
metacognitive sensitivity, (2) to improve the accuracy for metacognitive confidence to
97.5%, and (3) to identify key ambiguities and limitations in metacognition research.
This study utilized an EEG dataset with event-related potentials (ERP) responselocked for type 1 decisions. Data preprocessing addressed dataset imbalances
through augmentation and balanced batch sampling. EEG samples were transformed
into spectrograms and processed using the deep learning architecture comprising a
Lightweight Convolutional Neural Network (LWCNN), a Squeeze and Excitation
Network (SEN), and a classification network. The model was pre-trained using Subject
Aware Contrastive loss (SAC) and trained with binary cross-entropy loss. SEN
facilitated the models explainability by visualizing the created attention weights
through topoplots, providing insights into brain areas used for classification.
The WaveFusion model achieved high classification accuracy, reaching 99.7% for
metacognitive confidence and 99.1% for metacognitive sensitivity. These
improvements were due to a larger selection of electrodes, response-locked ERP
data, and increased dataset size. The WaveFusion model not only demonstrates high
classification accuracy but also offers enhanced explainability. This allows the
framework to contribute to three major ambiguities: (1) the relationship between
metacognition and executive functions, (2) its connection to consciousness, and (3)
the domain generality of metacognition. By leveraging the WaveFusion framework, we
can overcome limitations in cognitive neuroscience research through (1) utilizing
transfer learning to compare relationships, (2) employing automatic classification to
investigate ecological validity, and (3) expanding the framework for multimodality to
integrate insights across various fields.
Future research should focus on increasing data variability, addressing outlier
performances, and improving interpretability through advanced visualization
techniques to enhance the WaveFusion model’s robustness and applicability across
cognitive neuroscience domains.
Meer lezen