EEG asymmetry detection in patients with severe acquired brain injuries via machine learning methods. (January 2023)

• Record Type: Journal Article
• Authors: Corsi, Leonardo; Liuzzi, Piergiuseppe; Ballanti, Sara; Scarpino, Maenia; Maiorelli, Antonio; Sterpu, Raisa; Macchi, Claudio; Cecchi, Francesca; Hakiki, Bahia; Grippo, Antonello; Lanatà, Antonio; Carrozza, Maria Chiara; Bocchi, Leonardo; Mannini, Andrea
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: Disorders of consciousness -- Electroencephalography -- Diagnostic models -- Machine learning -- qEEG -- Brain symmetry; 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.2022.104260
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: EEG-derived symmetry indexes are markers of recovery for patients with acquired brain injury. Machine learning algorithms can be crossvalidated to automatically extract symmetry index. The solution detected asymmetry with a test accuracy of 85% (sensitivity 92%, specificity 80%) It can speed up analysis and improve quality of care in settings lacking skilled staff. Abstract: Lateral brain symmetry indexes, detected by electroencephalography (EEG), are markers of rehabilitative recovery widely used in patients with severe acquired brain injury (sABI). In this study, Machine Learning algorithms were cross-validated to detect consistent asymmetries, starting from a completely automated features extraction pipeline in the EEG recordings of 54 patients with sABI, classified by two expert neurophysiologists. Raw data were filtered and segmented in two-seconds non-overlapping epochs. Low data quality in frontal electrodes caused up to 40% of epochs rejection, whilst central and posterior electrodes contributed with the greatest number of artefacts-free epochs. Out of more than 3000 extracted features, ∼300 significantly differentiated symmetric and asymmetric EEG recordings, most of them extracted from pairs and lines of electrodes. The best performing solution (nested-cross-validated and optimized Support Vector Machine classifier) detected asymmetry with a test accuracy of 85% (sensitivity 92%, specificity 80%). The application of the proposed approach to our sample size supports the generalizability of our model and its translation to clinical practice. The algorithm, heading to automatic EEG analysis, has the potential to speed up analysis of long recordings and, to improve quality of care in settings lacking skilled staff.