An Integrated Calibration–Uncertainty Framework for Improving the Reliability of Deep Learning Models for Seizure Detection in EEG Signals

Background and objective: Deep learning models have demonstrated strong performance in automated seizure detection from EEG signals. However, these models may produce confident predictions even when incorrect, limiting their reliability barrier for clinical adoption. This study proposes an integrated calibration–uncertainty framework to enhance model reliability in EEG-based seizure classification. Methods: A CNN–BiLSTM model was trained to classify EEG epochs containing epileptic seizure activity. The framework leverages Expected Calibration Error (ECE) to assess global confidence reliability and Monte Carlo Dropout (MCD)-based uncertainty quantification to identify unreliable predictions. For each dropout rate, we evaluated both model calibration and the entropy-based separability between correctly (CC) and misclassified (MC) samples, computed as the Overlap Area between their uncertainty distributions. A multi-objective selection strategy was then used to automatically identify the configuration that best balances these complementary as- pects. Finally, a selective classification approach was implemented, using an uncertainty threshold to identify unreliable predictions and defer them for further clinical evaluation. Results: Varying the dropout rate significantly affected both calibration and uncertainty behaviour. The optimal balance was achieved at p = 0.1, yielding the lowest combined ECE and Overlap Area. The selective classification improved accuracy from 91.7% (baseline) to 99.6% while retaining ~74% of samples, outperforming models optimized for either calibration or uncertainty alone. Conclusions: The proposed dual perspective framework improves model robustness by integrating global confi- dence calibration with local uncertainty estimation, representing a practical step toward reliable AI deployment in clinical neurophysiology.