A Reliable Deep Learning Framework for EEG Seizure Detection based on Uncertainty Quantification and Confidence Analysis

The use of deep learning for epileptic seizure detection from electroencephalographic (EEG) data has led to impressive advances in classification accuracy, being able to correctly distinguish between ictal and interictal phases. However, a crucial challenge remains in ensuring that these models are both transparent and reliably calibrated for safe clinical adoption. In fact, neural network models for seizure detection often operate as “black boxes,” yielding predictions without transparent reasoning regarding the level of certainty associated with each decision. In many of these systems, neural networks give out confidence levels that do not always show the real chance of making the right prediction. This reduces the trust in model decisions and hinders their adoption in real-world clinical settings. This limitation is particularly critical in healthcare, where false confidence can directly impact diagnosis and treatment as it can lead to high false positive rates or missed seizure events. Therefore, it is not sufficient to obtain high performance. It is also equally important to understand when and to what extent a model can be trusted, especially when used in medical decision-support systems. This thesis aims to introduce an innovative framework that explicitly integrates calibration and uncertainty estimation into the seizure detection pipeline to achieve a more realistic evaluation of model confidence. In the outlined context, this study proposes a framework that extends traditional deep learning pipelines with explicit uncertainty quantification and calibration analysis. The proposed system exploits Monte Carlo Dropout (MCD) to address the first issue, enabling the estimation of epistemic uncertainty by performing multiple stochastic forward passes through the network with dropout activated at inference time. Building upon this, the model calibration is achieved through a novel technique based on the estimation of the overlap area between the entropy distributions of the two classes. The resulting measure serves as the core calibration strategy of the framework. In addition, the Expected Calibration Error (ECE) is employed as a complementary evaluation metric, used to quantitatively assess the model’s confidence and the effectiveness of the calibration process. This combined approach offers a dual perspective, providing an extended characterization of the model’s behavior. It not only measures the overall predictive performance but also reveals how model confidence correlates with uncertainty across the decision space. Moreover, it enables a deeper analysis of how uncertainty quantification can contribute to improving the accuracy and reliability of the deep learning model. Extensive evaluations conducted on the public CHB-MIT dataset reveal that traditional numerical performance metrics may provide promising accuracy figures. At the same time, a combined analysis of confidence and uncertainty exposes regions of overconfidence and instability in the network’s prediction boundary. Overall, this work bridges the gap between model performance and reliability in EEG‐based seizure detection. It provides a solid methodological foundation for developing models that are not only accurate but also interpretable and trustworthy in clinical environments. In doing so, it opens the possibility of integrating the framework into a wider range of deep learning architectures.