Researchers recently presented results from the first multicentre trial addressing the performance of a machine learning algorithm in detecting seizures in neonates.
Recognising seizures
Newborns often exhibit a range of unusual repetitive movements. Distinguishing whether these movements are due to seizures is vital. Seizures are often a sign of an underlying neurological condition and exposing newborns to unnecessary drugs would otherwise be harmful. However, diagnosis is often difficult for clinicians. Neonatal seizures are electrographic only and even if present, clinical signs are subtle and hard to distinguish from normal neonatal movements. While neonatologists often use amplitude-integrated electroencephalography (aEEG), scientists have reported its limitations. The gold standard for diagnosis of all seizures is continuous EEG (cEEG).
Regardless of the underlying cause, evidence has shown that seizures have a detrimental impact on neurodevelopment. Therefore, early recognition and treatment is necessary. Nevertheless, accurate diagnosis of seizures and prompt treatment remain challenging. Accurate interpretation of EEGs requires expert input, which may not be available in some neonatal intensive care units. EEGs are also very expensive and time consuming to use and interpret.
Automated detection
It has been proposed that automated seizure detection software should be incorporated at the bedside. As a result, experts have developed several algorithms for neonates. Some of these have been incorporated into EEG or aEEG systems for commercial use. In this study, published in The Lancet Child and Adolescent Health, researchers developed an EEG-based seizure detection software system called the Algorithm for Neonatal Seizure Recognition (ANSeR). The team wanted to evaluate the performance of this algorithm in real time. They achieved this by assessing the diagnostic accuracy for detecting seizures with and without the use of ANSeR as a support tool for clinicians.
Researchers conducted the trial in eight neonatal intensive care units across Ireland, the Netherlands, Sweden and the UK. Neonates with or at suspected risk of seizures requiring EEG were eligible for the study. Eligible neonates were randomly assigned to receive cEEG monitoring with ANSeR (algorithm group) or cEEG monitoring only (non-algorithm group). The primary outcome was the diagnostic accuracy (sensitivity, specificity and false detection rate) of healthcare professionals to identify neonates with electrographic seizures.
Algorithm vs Non-algorithm
The team found that out of the 128 neonates in the algorithm group, 25% presented with electrographic seizures compared to 29.2% (38 out of 130) neonates from the non-algorithm group. Sensitivity in the algorithm group was 81.3% and 89.5% in the non-algorithm group. Specificity was 84.4% in the algorithm group and 89.1% in the non-algorithm group. False detection rate was 36.6% in the algorithm group and 22.7% in the non-algorithm group. Although the diagnostic accuracy did not differ between the two groups, the team suggested that this is likely due to the high levels of experience in EEG at the recruiting hospitals. Nevertheless, the algorithm group was able to identify a higher percentage of hours in which seizures occurred (seizure hours). This suggests machine learning may be helpful in providing a real-time support tool for the recognition of seizures in neonates.
This study demonstrates that ANSeR is a safe machine learning algorithm that can accurately detect neonatal seizures. Whilst it did not enhance identification of individual neonates with seizures beyond EEG, it did improve the recognition of seizure hours. Healthcare professionals in this study used recognition of seizure hours as a support for seizure identification. This tool may provide a more beneficial opportunity for the use of machine learning in centres with less experience in interpreting neonatal EEG.
Image credit: By Image Team – canva.com
