Prediction of Cerebral Palsy in Newborns With Hypoxic-Ischemic Encephalopathy Using Multivariate EEG Analysis and Machine Learning

• Published in: IEEE access 2021, Vol.9, p.137833-137846
• Main Authors: Bakheet, Dalal, Alotaibi, Noura, Konn, Daniel, Vollmer, Brigitte, Maharatna, Koushik
• Format: Article
• Language: English
• Subjects: Biomarkers; Brain; Brain connectivity; Cerebral palsy; cerebral palsy (CP); Complexity; Complexity theory; Diameters; electroencephalogram (EEG); Electroencephalography; Empirical analysis; Empirical mode decomposition; empirical mode decomposition (EMD); Entropy; Feature extraction; Frequency ranges; Graph theory; Hypoxia; hypoxic-ischemic encephalopathy (HIE); Infants; Low-pass filters; Machine learning; Mathematical analysis; Medicin och hälsovetenskap; Multivariate analysis; Neurology; noise-assisted multivariate empirical mode decomposition (NA-MEMD); Parameters; Pediatrics; Permutations; Phase lag; Statistical analysis; weighted phase-lag index (WPLI)
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2021.3118076

This study was carried out to investigate whether the quantitative analysis of electroencephalogram (EEG) signals of infants with hypoxic-ischemic encephalopathy (HIE) can be used for early prediction of cerebral palsy (CP). We computed sample entropy (SampEn), permutation entropy (PEn), and spectral entropy (SpEn) measures to reflect the signals' complexity and the graph-theoretic parameters derived from weighted phase-lag index (WPLI) to measure functional brain connectivity. Both feature sets were calculated in the noise-assisted multivariate empirical mode decomposition (NA-MEMD) domain to characterize the tempo-spectral integration of information and thus provide novel insight into the brain dynamics. Statistical analysis results showed a general abnormality in the EEG of individuals with CP at the alpha-band component. Particularly, complexity measures were decreased, and graph-theoretic parameters specified by the diameter feature were increased in infants with CP compared to those with normal neurology. The proposed set of features have also been evaluated using the random under-sampling boosting (RUSBoost) classifier, which was trained and tested on the feature vectors of a cohort of 26 infants - 6 who developed CP by the age of 24 months and 20 with normal neuromotor outcome. A good performance of 84.6% classification accuracy (ACC), 83% sensitivity (SNS), 85% specificity (SPC) and 0.87 area under curve (AUC) was obtained using the entropy features extracted from the alpha-band component. A close result of 80.8% ACC, 67% SNS, 85% SPC and 0.79 AUC was also achieved using the diameter feature calculated from the same frequency range. Therefore, it was concluded that the obtained brain functions' characteristics successfully discriminate between the two groups of infants. These characteristics could be considered potential biomarkers of cerebral cellular damage and, therefore, could be employed in practical clinical applications for early CP prediction.