Deep learning-based classification of multichannel bio-signals using directedness transfer learning. (February 2022)

• Record Type: Journal Article
• Authors: Bahador, Nooshin; Kortelainen, Jukka
• Is Part Of: Biomedical signal processing and control. Volume 72(2022)Part A
• Publication Date: 2022-02
• Subjects: Spectral-phase-spatial fusion -- Multivariate data -- Transfer learning -- Deep learning -- Bi-dimensional mapping; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103300
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Fusing spectral, phase and spatial information. Extracting high-level features representing information flow over EEG channels. Transforming time-series data into images. Transfer knowledge from deep models trained on large-scaled image dataset for time-series data analysis. Abstract: The problem with processing of multivariate/multichannel signals lies in adapting of existing classifiers on data. Reformulating time-series data as visual clues and assigning visual patterns to different categories help the classification of time series in a wide range of applications. These series-to-image transformations have benefits including better noise robustness and more options regarding augmentation. They also provide the possibility of achieving discriminative features by employing transfer learning paradigm in cases dealing with highly small training datasets. In this respect, this work aimed to encode spectral-phase information into a bi-dimensional map. Transferring knowledge was done using bi-dimensional transformation capitalizing on the direction and propagation pattern of one channel influence on the others. EEG data from patients diagnosed with delirium (N = 15) recorded using a 10-channel BrainStatus device were used for this analysis. Considering leave-one-subject-out cross-validation, classification outcomes demonstrated that directedness transfer learning via Alexnet yields a promising performance showing 97.17% precision and outperforming other approaches. Comparison with nine different deep networks pretrained on ImageNet database was included. Directedness transfer learning resulted in precision of 95.29 ± 1.46 (µ ± σ)% among all networks. For further evaluation, directedness bi-dimensional transformation was also compared with six other 2D maps. Applying different networks resulted in average precision of (91.99 ± 2.23)% for polar-, (91.69 ± 1.57)% for correlation-, (90.46 ± 1.71)% for Spectrogram-, (87.82 ± 2.16)% for Wavelet-, (84.24 ± 1.72)% for Wigner-Ville- and (82.84 ± 2.46)% for Mel-frequency Cepstrum maps. To conclude, the proposed technique shows significant benefit in compressing spatio-spectral patterns of multichannel signals in just a unified visual representation.