Dokare Indu, Gupta Sudha
Department of Electronics Engineering, K. J. Somaiya School of Engineering (formerly K. J. Somaiya College of Engineering), Somaiya Vidyavihar University, Vidyanagar, Vidyavihar East, Mumbai, Maharashtra 400077 India.
Department of Computer Engineering, Vivekanand Education Society's Institute of Technology, Chembur, Mumbai, Maharashtra 400074 India.
Cogn Neurodyn. 2025 Dec;19(1):85. doi: 10.1007/s11571-025-10269-3. Epub 2025 Jun 5.
The aim of this research is to combine Explainable AI (XAI) with advanced optimization techniques to provide a unique framework for seizure detection. This proposed work investigates how to enhance patient-specific and patient-non-specific seizure detection models by combining multiband feature extraction, SHAP-based feature selection, SMOTE, and a metaheuristic algorithm for hyperparameter tuning.The discrete wavelet transform (DWT) is used to decompose EEG signals to retrieve entropy-based and statistical information. Simulated Annealing (SA) is employed to optimize the Random Forest (RF) classifier's hyperparameters, and SHAP (SHapley Additive exPlanations) values are utilized for feature selection. Furthermore, a novel technique SHAP-RELFR has been demonstrated to select patient-non-specific features. Additionally, SMOTE is employed to handle imbalanced data. The proposed methodology is evaluated on the CHB-MIT and Siena datasets using both patient-specific and patient-non-specific feature selection approaches. Experimental findings demonstrate that the proposed methodology significantly improves the performance of seizure detection. The average accuracy, precision, sensitivity, specificity, F1-score, and AUC obtained for a patient-non-specific case are 96.58%, 95.19%, 94.52%, 98.02%, 94.72%, and 0.9452, respectively, using the CHB-MIT dataset. For the Seina dataset, the average accuracy, precision, sensitivity, specificity, F1-score, and AUC obtained for a patient-non-specific case are 94.81%, 94.51%, 94.04%, 96.87%, 94.28%, and 0.9400, respectively. Explainable AI combined with SMOTE and a metaheuristic optimization algorithm facilitates an enhanced seizure detection. The novel SHAP-RELFR method provides an effective patient-non-specific feature selection, enabling this approach to be applicable across diverse patients. This proposed framework offers a step toward enhancing clinical decision-making by providing interpretable and versatile seizure detection models.
本研究的目的是将可解释人工智能（XAI）与先进的优化技术相结合，以提供一个独特的癫痫发作检测框架。这项拟议的工作研究了如何通过结合多波段特征提取、基于SHAP的特征选择、SMOTE以及用于超参数调整的元启发式算法，来增强针对特定患者和非特定患者的癫痫发作检测模型。离散小波变换（DWT）用于分解脑电图（EEG）信号，以获取基于熵和统计的信息。采用模拟退火（SA）算法优化随机森林（RF）分类器的超参数，并利用SHAP（SHapley值加法解释）值进行特征选择。此外，一种新颖的技术SHAP-RELFR已被证明可用于选择非特定于患者的特征。此外，采用SMOTE来处理不平衡数据。所提出的方法在CHB-MIT和锡耶纳数据集上使用特定于患者和非特定于患者的特征选择方法进行评估。实验结果表明，所提出的方法显著提高了癫痫发作检测的性能。使用CHB-MIT数据集时，非特定于患者的情况下获得的平均准确率、精确率、灵敏度、特异性、F1分数和AUC分别为96.58%、95.19%、94.52%、98.02%、94.72%和0.9452。对于锡耶纳数据集，非特定于患者的情况下获得的平均准确率、精确率、灵敏度、特异性、F1分数和AUC分别为94.81%、94.51%、94.04%、96.87%、94.28%和0.9400。可解释人工智能与SMOTE和元启发式优化算法相结合，有助于增强癫痫发作检测。新颖的SHAP-RELFR方法提供了有效的非特定于患者的特征选择，使该方法能够适用于不同的患者。所提出的框架通过提供可解释且通用的癫痫发作检测模型，朝着增强临床决策迈出了一步。
