Detection of Transient Bursts in the EEG of Preterm Infants using Time-Frequency Distributions and Machine Learning

Area of Science:

• Neuroscience
• Computational Biology
• Medical Technology

Background:

• Short-duration bursts in electroencephalogram (EEG) recordings are crucial indicators of neurological maturation in premature infants.
• Accurate detection of these bursts is vital for monitoring brain development in vulnerable preterm populations.

Purpose of the Study:

• To evaluate a feature-less machine learning approach for detecting spontaneous activity bursts in the EEG of preterm infants.
• To assess the efficacy of transforming time-series EEG data into time-frequency distributions (TFDs) for direct machine learning analysis.

Main Methods:

• EEG data were collected from infants with a gestational age below 30 weeks within the first 3 days of life.
• A gradient boosting machine was trained directly on time-frequency distributions (TFDs) of the EEG, bypassing traditional feature extraction.
• Parameter optimization for TFD kernel, Doppler, and lag windows was performed using nested cross-validation.

Main Results:

• The machine learning model achieved a median area under the receiver operator characteristic curve of 0.881 for burst detection.
• Detection performance was found to be sensitive to the Doppler window length but not the lag window length.
• Analysis revealed that a wideband region below 15 Hz in the TFD was critical for burst detection.

Conclusions:

• A feature-less machine learning approach using TFDs is effective for detecting EEG bursts in preterm infants.
• This method contributes to the development of automated brain-health indices for monitoring preterm infant neurodevelopment.
• Identifying critical frequency bands in TFDs enhances the understanding of EEG maturation markers.