Brain age as an estimator of neurodevelopmental outcome: A deep learning approach for neonatal cot-side monitoring

Abstract

The preterm neonate can experience stressors that affect the rate of brain maturation and lead to long-term neurodevelopmental deficits. However, some neonates who are born early follow normal developmental trajectories. Extraction of data from electroencephalography (EEG) signals can be used to calculate the neonate's brain age which can be compared to their true age. Discrepancies between true age and brain age (the brain age delta) can then be used to quantify maturational deviation, which has been shown to correlate with long-term abnormal neurodevelopmental outcomes. Nevertheless, current brain age models that are based on traditional analytical techniques are less suited to clinical cot-side monitoring due to their dependency on long-duration EEG recordings, the need to record activity across multiple EEG channels, and the manual calculation of predefined EEG features which is time-consuming and may not fully capture the wealth of information in the EEG signal. In this study, we propose an alternative deep-learning approach to determine brain age, which operates directly on the EEG, using a Convolutional Neural Network (CNN) block based on the Inception architecture (called Sinc). Using this deep-learning approach on a dataset of preterm infants with normal neurodevelopmental outcomes (where we assume brain age = postmenstrual age), we can calculate infant brain age with a Mean Absolute Error (MAE) of 0.78 weeks (equivalent to a brain age estimation error for the infant within +/- 5.5 days of their true age). Importantly, this level of accuracy can be achieved by recording only 20 minutes of EEG activity from a single channel. This compares favourably to the degree of accuracy that can be achieved using traditional methods that require long duration recordings (typically >2 hours of EEG activity) recorded from a higher density 8-electrode montage (MAE = 0.73 weeks). Importantly, the deep learning model's brain age deltas also distinguish between neonates with normal and severely abnormal outcomes (Normal MAE = 0.71 weeks, severely abnormal MAE = 1.27 weeks, p=0.02, one-way ANOVA), making it highly suited for potential clinical applications. Lastly, in an independent dataset collected at an independent site, we demonstrate the model's generalisability in age prediction, as accurate age predictions were also observed (MAE of 0.97 weeks).

Sensory event-related potential morphology predicts age in premature infants

“…A wide range of brain age models has been developed to trace the brain development of premature infants, encompassing structural connectivity (Brown et al, 2017;Kawahara et al, 2017), morphological (Liu et al, 2021) and electrophysiological data. For the latter, brain age models have previously been constructed in preterm infants using resting state EEGrecorded brain activity (Ansari et al, 2023;Lavanga et al, 2018;O'Toole et al, 2016;Pillay et al, 2020;Stevenson et al, 2017). Although the MAE achieved by our model is not as accurate as some resting state models (e.g., MAEs of approximately 1 week were achieved by Ansari et al (2023) and Liu et al (2021)), compared to these existing brain age models, our model has the advantage that it was constructed using electrophysiological responses of approximately 10 visual and 10 tactile stimuli from every recording.…”

“…Machine learning approaches can be used to accurately predict the post-menstrual age (PMA) of preterm infants from EEG (Ansari et al, 2023;Lavanga et al, 2018;O'Toole et al, 2016;Pillay et al, 2020;Stevenson et al, 2017), diffusion magnetic resonance imaging (MRI) (Brown et al, 2017;Kawahara et al, 2017) and structural MRI (Liu et al, 2021). These models may facilitate the early identification of infants with abnormal neurodevelopment, reducing the need for visual inspection of the EEG/MRI, which is subjective, requires trained clinical staff, and is time-consuming.…”

“…This so-called brain age can be seen as a maturation index of the neural system which is unlikely to reflect chronological age, which can be viewed as a continuous "ticking clock" (Salih et al, 2023). Previous EEG brain age models have focused on continuous ongoing EEG activity (i.e., non-evoked brain activity) (Ansari et al, 2023;O'Toole et al, 2016;Pillay et al, 2020;Stevenson et al, 2017). An alternative may be to construct models capturing evoked responses, giving specific information about the maturity of sensory processing.…”

A multi-task and multi-channel convolutional neural network for semi-supervised neonatal artefact detection

“…Finally, the relationship between the amount of detected artefacts in an EEG segment and the error made by a FBA estimation model was investigated to show the relevance of the proposed automated artefact detection method. To this end, a deep shared multi-scale inception network model was applied to the EEG recordings to estimate the FBA [30,31]. This FBA model provides one estimate of FBA for every 30 s of multi-channel EEG data.…”

Amplitude and Frequency Modulation of EEG Predicts Intraventricular Haemorrhage in Preterm Infants

“…A bandpass filter of 0-30 Hz was applied. To apply artifact correction (similar to approach in [35]) on the data, each recording was divided into 5-minute epochs. Data points that deviated from the epoch mean by more than 250 μV were discarded and replaced by linear interpolation, unless more than 7000 of such samples (10% of epoch length) were rejected.…”