Early automated classification of neonatal hypoxic-ischemic encephalopathy − An aid to the decision to use therapeutic hypothermia

• Published in: Clinical neurophysiology 2024-10, Vol.166, p.108-116
• Main Authors: Lacan, Laure, Betrouni, Nacim, Chaton, Laurence, Lamblin, Marie-Dominique, Flamein, Florence, Riadh Boukhris, Mohamed, Derambure, Philippe, Nguyen The Tich, Sylvie
• Format: Article
• Language: English
• Subjects: Automated classification model; Electroencephalography - methods; Female; Humans; Hypothermia, Induced - methods; Hypoxia-Ischemia, Brain - classification; Hypoxia-Ischemia, Brain - diagnosis; Hypoxia-Ischemia, Brain - physiopathology; Hypoxia-Ischemia, Brain - therapy; Infant, Newborn; Machine Learning; Male; Neonatal EEG; Neonatal HIE; Perinatal asphyxia; Therapeutic hypothermia
• ISSN: 1388-2457; 1872-8952; 1872-8952
• DOI: 10.1016/j.clinph.2024.07.015

•We use qEEG analysis to mimic a visual EEG analysis approach in neonatal hypoxic-ischemic encephalopathy.•Machine learning models were developed and validated on two independent EEG datasets.•The proposed model is effective in discriminating neonates requiring therapeutic hypothermia in the first 6 h of life. The study aimed to address the challenge of early assessment of neonatal hypoxic-ischemic encephalopathy (HIE) severity to identify candidates for therapeutic hypothermia (TH). The objective was to develop an automated classification model for neonatal EEGs, enabling accurate HIE severity assessment 24/7. EEGs recorded within 6 h of life after perinatal anoxia were visually graded into 3 severity groups (HIE French Classification) and quantified using 6 qEEG markers measuring amplitude, continuity and frequency content. Machine learning models were developed on a dataset of 90 EEGs and validated on an independent dataset of 60 EEGs. The selected model achieved an overall accuracy of 80.6% in the development phase and 80% in the validation phase. Notably, the model accurately identified 28 out of 30 children for whom TH was indicated after visual EEG analysis, with only 2 cases (moderate EEG abnormalities) not recommended for cooling. The combination of clinically relevant qEEG markers led to the development of an effective automated EEG classification model, particularly suited for the post-anoxic latency phase. This model successfully discriminated neonates requiring TH. The proposed model has potential as a bedside clinical decision support tool for TH.