Fusion Methods for Biosignal Analysis: Theory and Applications.

Abstract

Recent advances in data acquisition and biosignal processing are paving the way for the optimal integration or fusion of complementary data modalities in a wide variety of clinical settings. Data modalities include electrocardiography (ECG), electroencephalography (EEG), electrocorticography (ECoG), magnetic resonance imaging (MRI), functional MRI (fMRI), positron emission tomography (PET), and diffusion tensor imaging (DTI). Integration can be performed by exploiting the analyses sequentially or simultaneously, depending on issues related to synchronization, physical compatibilities, and standard clinical procedures. Fusion approaches aim at integrating analyses of data from different modalities, establishing synergic relationships for improved clinical hypothesis testing and medical diagnosis. The heterogeneous nature of data sources from different clinical analyses and acquisition modalities presents big challenges. The main objective of data fusion is to exploit complementary properties of several single-modality methods in order to improve each of them considered separately. In addition, fusion can enable or enhance the approximation to more complex structured results (e.g., hierarchical trees and topological networks). This broad field of research has been named in different ways, for instance, sensor data fusion; decision fusion; multimodal fusion; heterogeneous sensor fusion; mixture of experts; classifier combiners; and multiway signal processing. This special issue focuses on theoretical and application advances in fusion methods for biosignal analysis (FMBA). Fourteen submissions were received and each manuscript was reviewed by at least two external referees. Five papers were finally accepted for the special issue. The accepted papers cover important problems related to computational intelligence and neuroscience. In the paper opening the issue, H. Banville et al. propose a method for mental task evaluation based on the fusion of features extracted from EEG and near-infrared spectroscopy. The tasks evaluated are word generation, mental rotation, subtraction, singing and navigation, and motor and face imagery. The method is intended to improve classification performance for more efficient and flexible brain computer interfaces (BCIs). Y. Huang et al. develop a multimodal late fusion procedure that combines features from facial expression images and EEG data for emotion recognition. A two decision-level fusion is performed to classify four basic emotion states (happiness, neutral, sadness, and fear) and three emotion intensity levels (strong, ordinary, and weak). A general-purpose method for classification named dandelion algorithm (DA) is developed in the paper by X. Li et al. Different combinations of DA with other bioinspired methods (bat algorithm, particle swarm optimization, and enhanced fireworks algorithm) and a neural network-based method called extreme learning machine are implemented. Results show improvement in classification accuracy for different biomedical problems using data from EEG, ECG, and single photon emission computed tomography (SPECT).