An adaptive hybrid swarm intelligent algorithm based on reinforcement learning for imbalanced EEG data processing in epileptic seizure detection
Epilepsy is a chronic neurological disorder that affects approximately 70 million people worldwide.With varied severity and impacts,it is characterized by recurrent seizures that occur without apparent triggers.Electroencephalography(EEG)is essential for diagnosing and monitoring epilepsy as it captures the brain's electrical signals,which are a direct result of neuronal firing.It is particularly effective for identifying seizures and classifying their types.A significant hurdle in EEG-based epilepsy detection is the scarcity of seizure(ictal)events compared with non-seizure(interictal)periods.This imbalance poses a challenge for developing detection models that can accurately pinpoint seizures within the vast majority of non-seizure EEG data.In the majority of unbalanced classification problems,including epilepsy detection,learning algorithms favor the majority class(non-seizure data),which often results in poor performance in identifying the minority class(seizure data).Although the models can predict non-seizure states with high accuracy in most cases,they often perform poorly in recognizing epileptic seizures.To address the issue,various methods have been proposed,such as undersampling the majority class,oversampling the minority class,or generating synthetic data.While these methods somehow improve their performances,they often cause information loss and blurred class boundaries,undermining model generalization and classification performances.To address these challenges,Swarm Intelligence Algorithms(SIAs)have emerged as a promising solution.Inspired by the collective behaviors of birds and ants,etc.,these algorithms have garnered keen academic interest for their efficiency in tackling complex and high-dimensional optimization problems.SIAs,including Particle Swarm Optimization(PSO),Ant Colony Optimization(ACO),and Artificial Bee Colony(ABC)algorithms,have achieved phenomenal success in tackling various unbalanced classification tasks.Compared with traditional optimization methods,SIAs prevent the loss of critical information,reduce the risk of overfitting and mitigate the issue of blurred decision boundaries in classification tasks.Despite their excellent performances in certain tasks,they still have limitations when dealing with the specific challenges of EEG data in epilepsy detection,particularly when addressing unbalanced data,where the characteristics of seizure and non-seizure categories are often highly complex and variable.To overcome these limitations,we propose an adaptive fusion swarm intelligence algorithm based on reinforcement learning(RL).The core concept of our approach is to leverage RL to dynamically select and combine various SIAs in response to the evolving complexity of classification problems during the optimization process.The RL agent is responsible for determining which swarm intelligence algorithm to apply at each stage of the optimization process,thereby enhancing the model's adaptability and flexibility in handling different aspects of the problem.This self-adaptive fusion approach helps the model respond more efficiently to the characteristics of both seizure and non-seizure data,enhancing the classification performances across different stages of the data.Additionally,our method incorporates a dual-population coevolutionary strategy.This strategy enhances the search process efficiency by enabling the algorithm to thoroughly explore the search space and more effectively identify the global optimal solution.The dual-population approach divides the population into two groups,each focusing on different aspects of the solution space.This partitioning facilitates more diverse exploration,thereby improving the overall optimization process and enhancing performances in balancing the representation of both classes.To assess the effectiveness of our method,we conduct extensive experiments on two public EEG datasets.Our results demonstrate the proposed adaptive fusion swarm intelligence algorithm outperforms traditional single-swarm intelligence methods in classification,particularly in identifying epileptic seizure events(the minority class).In conclusion,the adaptive fusion swarm intelligence algorithm based on reinforcement learning offers a promising solution to the class imbalance problem in EEG-based epilepsy classification.By synergizing the strengths of various swarm intelligence algorithms and leveraging reinforcement learning to dynamically select the most appropriate strategy,our method provides a more robust and efficient strategy for seizure detection,thereby enhancing the overall efficacy of epilepsy diagnostic systems.
国家科技期刊平台
NSTL
万方数据
