PaperlayerView exact revision on arXiv

Source-linked AI summary

Expert-level detection of low-voltage intervals in EEG as a robust biomarker of neonatal encephalopathy

John M. O'Toole, Sean R. Mathieson, Aurel Luca, Soraia Ventura, Sean Griffin, Geraldine B. Boylan, Robert Hogan

arXiv:2610.12680v1cs.LGeess.SPphysics.med-ph

TL;DR

Neonatal EEG discontinuity is difficult to quantify automatically because IBI definitions depend on background context. This study introduces and detects amplitude-and-duration-defined LVIs, finding expert-level detection agreement and clinical associations with EEG grades, HIE severity, and seizure risk.

  • Problem

    IBI detection is difficult to generalize because voltage thresholds depend on EEG vigilance state or background pattern, limiting context-independent automated assessment.

    Source: §1 · p. 2

  • Method

    The study defines LVIs by amplitude and duration without requiring a burst pattern, develops machine-learning detectors, and evaluates LVI features across clinical EEG applications.

    Source: §2 · p. 2, §1 · p. 2, §2.2 · p. 4

  • Results

    LVI detection achieved expert-level equivalence, while LVI features stratified EEG and HIE grades and discriminated later seizures (AUC 0.88 for median duration and LVI percentage).

    Source: §4 · pp. 10, 12

  • Takeaways & Limitations

    Automated LVI quantification shows potential as a biomarker of neonatal encephalopathy and as an input to clinical decision support or application-specific algorithms.

    Source: §4 · p. 13

    Validation used EEG from the development centre and excluded major artefacts, so generalisability across centres and performance with significant artefacts remain uncertain.

    Source: §4 · p. 13

Abstract

from arXiv · show

Background: Discontinuity in neonatal EEG is a critical marker of brain dysfunction but can be difficult to quantify visually. Inter-burst intervals (IBIs) are a key characteristic of discontinuity, yet their definition does not easily translate to AI detection. Objectives: To introduce low-voltage intervals (LVIs), defined exclusively by amplitude and duration, as a context-independent alternative to IBIs, develop an AI detection method for LVIs, and assess their clinical utility as a marker of brain injury. Methods: We developed a ConvNeXt model to detect LVIs and compared its performance with existing methods. We then tested the clinical relevance of detected LVIs across four applications: 1) grading background EEG abnormalities on an open benchmark dataset, 2) classifying clinical hypoxic-ischaemic encephalopathy (HIE) grades and comparing with a normal EEG dataset, 3) seizure prediction, and 4) association with seizure burden. Results: For training and validation, 654 hours of per-channel EEG from 123 neonates were reviewed and 51,162 LVIs were annotated. On held-out validation, ConvNeXt had an AUC of 0.996 and a Matthews correlation coefficient of 0.926. Model-reader agreement exceeded inter-reader agreement ($\Delta \kappa = 0.025$, 95% CI: 0.007 to 0.045, $p<0.001$), attaining expert-level equivalence. On a combined dataset of 444 neonates, percentage LVI showed a strong association with EEG grades (Spearman's $\rho=0.801$), a moderate association with HIE grades ($\rho=0.384$), a weaker association with seizure burden ($\rho=0.244$), and a strong seizure prediction capacity (AUC $=0.872$). Conclusion: AI detection of LVIs achieves expert-level equivalence, whereas quantitative features do not. Automated LVI quantification provides a robust clinical biomarker of neonatal encephalopathy associated with background EEG activity, HIE severity, and seizure risk.

1 Introduction

Neonatal EEG can track brain dysfunction and evolving injury, but interpreting its complex patterns requires scarce specialist expertise. Because IBI thresholds vary with EEG context, the study introduces low-voltage intervals (LVIs) as an amplitude-and-duration-based alternative for automated detection and clinical evaluation.

1 Introduction

  • Neonatal encephalopathy can cause death and lifelong impairment, while EEG monitoring requires specialist interpretation that is rarely available around the clock.The paper identifies AI pattern matching as a potential aid for reviewing the large volumes of EEG generated during long-duration monitoring.Source: §1 · p. 2
  • IBI interpretation is context-dependent: voltage thresholds vary by background pattern, and existing detection methods largely trained and tested on a single state or pattern.For example, the cited thresholds are <25 µV for tracé discontinu and 25–50 µV for tracé alternant.Source: §1 · p. 2
  • The study defines LVIs by low amplitude and duration without requiring bursts, then evaluates automated detection across four applications of neonatal encephalopathy.These applications are EEG-background grading, clinical HIE grading, seizure prediction, and association with total seizure burden.Source: §1 · p. 2

2 Methods

The methods define LVIs by amplitude and duration independently of EEG state, develop detectors using annotated neonatal EEG, and evaluate their performance and clinical relevance across distinct datasets and tasks.

2 Methods

  • LVIs are individual-channel intervals below 25 µV for at least 2 seconds, without requiring preceding or subsequent bursts.They may occur across different background patterns; their channel extent is characterized separately from identification.Source: §2 · pp. 2–3

2.1 EEG data and LVI annotation

  • Development and validation used separate neonatal EEG datasets; two readers independently annotated validation recordings on two channels, while the development set covered all available bipolar channels.The development set used one-hour epochs from 51 neonates and validation used 20-minute epochs from 72 neonates, with minimal artefact in both.Source: §2.1 · p. 4

2.2 Detection Methods

  • The detector study compared quantitative-feature baselines and an existing IBI machine-learning method with a proposed convolutional neural network approach.The proposed approach adapted a convolutional neural network previously developed for neonatal seizure detection.Source: §2.2 · p. 4

2.2.1 Quantitative Features

  • Four baseline features represented amplitude or frequency-weighted energy, with thresholds tuned to balance sensitivity and specificity.A 2-second median filter smoothed all features except line-length.Source: §2.2.1 · p. 5

2.2.2 Machine Learning

  • The SVM combined eight EEG features, while the ConvNeXt Nano network learned from 2-second EEG segments; both methods were trained on the development set and tested on validation.SVM probabilities were smoothed and thresholded; ConvNeXt used data augmentation and limited post-processing.Source: §2.2.2 · p. 5

2.2.3 Detection performance

  • Detection was assessed with sample-based metrics across concatenated data and per neonate, plus an expert-level equivalence test comparing reader and detector agreement.The equivalence threshold was set to the margin of error in inter-reader agreement rather than an arbitrary non-inferiority margin.Source: §2.2.3 · p. 5

2.3 Application EEG datasets

  • Clinical evaluation used an open-access EEG-grading dataset and a separate clinical dataset of 18,824 hours from 391 neonates, with no overlap with detector-development EEG.The open-access dataset contained 169 one-hour epochs from 53 neonates.Source: §2.3 · p. 6

2.4 Application Experiments

  • Six temporal LVI features were tested against EEG and HIE grades, later seizures, and seizure burden, using rank-based group tests and Spearman correlation.Seizure prediction used the first EEG hour with a gap before seizure onset; seizure-burden features summarized overlapping one-hour epochs across recordings.Source: §2.4 · p. 6

3 Results

Automated LVI detectors outperformed quantitative features and achieved expert-level agreement on held-out validation. LVI features tracked EEG and HIE grades, seizure risk and burden, with robustness tests supporting their use across epoch durations and event lengths.

3.1 Part A: LVI detection performance

  • Across development and held-out validation sets, 654 h of per-channel EEG yielded 51,162 annotated LVIs; validation-set agreement between two human readers was Cohen’s κ = 0.82 (95% CI: 0.74 to 0.87).On validation, four of six features demonstrated non-inferiority or superiority to expert inter-rater variability (p < 0.05); maximum- and median-duration tests failed to reject the null, with errors of 0.100 s (95% CI: -0.686 to 0.856 s) and -0.172 s (95% CI: -0.188 to 0.594 s), respectively.Source: §3.1 · p. 7, §3.2 · p. 8
  • On held-out validation, AUC was 0.996 for both methods, while ConvNeXt’s MCC was 0.926 versus SVM’s 0.923; both ML detectors reached expert-level equivalence.ConvNeXt’s Δκ was 0.024 (95% CI: 0.011–0.040); quantitative features did not reach expert-level equivalence.Source: Figure 2 · p. 7, §3.1 · pp. 7–8

3.2 Part B: Clinical Relevance and Biomarker Potential

  • For background EEG grades, median-duration LVI had ρ = 0.821 and ε2 = 0.676; for HIE grades, LVI percentage had ρ = 0.384 and ε2 = 0.297.Both analyses found significant grade separation (p < 0.001); the HIE analysis included healthy-neonate EEG as a reference.Source: §3.2 · p. 8, Figure 3 · p. 9
  • Across the first 100 h after birth, maximum LVI duration decreased after about 10 h in severe HIE, while features for mild and moderate HIE appeared relatively stable.The analysis covered more than 13,000 h of multichannel EEG, with features computed on 1 h epochs shifted by 15 min.Source: §3.2 · p. 8, Figure 3 · p. 9

3.2.1 Robustness of LVI biomarkers

  • Feature values differed more from full-hour estimates as epochs shortened, but predictive performance stayed relatively consistent, especially for EEG and HIE grading.Seizure prediction varied more on shorter epochs, although median AUC estimates at 5 or 10 min were close to the 1 h estimate.Source: §3.2.1 · pp. 9–10

3.2 Part B: Clinical Relevance and Biomarker Potential

  • For first-hour seizure prediction, LVI percentage led the features with AUC = 0.872 and ρ = 0.475.The analysis excluded neonates whose seizure onset occurred less than 4 h after recording began.Source: §3.2 · p. 8, Figure 4 · p. 10
  • For seizure burden, five of six features had weak but significant associations; LVI periodicity had the highest correlation, ρ = 0.398, and largest stratified effect size, ε2 = 0.138.Burden was summed over each neonate’s seizures, with features computed on 1 h epochs shifted by 75%.Source: §3.2 · pp. 8–9, Figure 4 · p. 10

3.2.1 Robustness of LVI biomarkers

  • Restricting analysis to short-duration LVIs still produced non-zero performance at thresholds up to 4 s, and removing short events did not improve performance.LVI amplitude decreased as duration increased; the paper frames LVIs as spanning tracé discontinu, burst suppression, and isoelectric activity.Source: §3.2.1 · p. 10

4 Discussion

The discussion positions automated LVI detection as expert-equivalent and evaluates its biomarker value across EEG grades, HIE severity, and seizure outcomes, while noting important scope limits.

4 Discussion

  • Automated LVI detection achieved expert-level equivalence, and quantitative LVI features showed clinical potential as biomarkers across a 444-neonate dataset.The authors note that large-scale expert annotations were central to the detection result, while rigid quantitative features did not match human accuracy.Source: §4 · pp. 10, 13
  • ConvNeXt performed marginally better than the SVM on almost all metrics, although the difference was likely not clinically meaningful.The authors describe the ConvNeXt probability output as less noisy and its inference pipeline as simpler; both methods could run in real time on a standard computer.Source: §4 · p. 11
  • Six features characterize the LVI mask, combining intuitive event-count, duration, and percentage measures with complexity and periodicity for temporal dynamics.Complexity and periodicity were not top-ranked in clinical applications, but capture information the other features do not; the authors favor a small, interpretable set.Source: §4 · pp. 11–12
  • Shorter epochs showed larger feature-value differences from the full hour, but matched-duration discrimination remained on par with 1-hour epochs.A 10-minute epoch may suffice to summarize clinically important LVI activity; analysis duration should be consistent.Source: Figure 5 · p. 11, §4 · p. 12
  • Median LVI duration correlated with EEG grades at ρ = 0.82 (p < 0.001), with amplitude and continuity identified as common grading factors.The authors caution that end-to-end methods may learn broader EEG characteristics without accounting for all subtler waveform abnormalities.Source: §4 · p. 12
  • LVI features separated HIE grades and the normal cohort less strongly than EEG grades; LVI percentage correlated with HIE grade at ρ = 0.38 (p < 0.001).LVI percentage showed stable separation across the first 24 hours, while Sarnat scores may change as injury evolves.Source: §4 · p. 12
  • Median duration and LVI percentage predicted subsequent seizures with an AUC of 0.88, while the median time from the analysed epoch to first seizure was 8 hours.LVI activity across the full EEG recording had a significant but weak association with seizure burden; maximum duration correlated at ρ = 0.27.Source: §4 · pp. 12–13
  • The validation set came from the development centre and excluded major artefacts, and the study provides no clinical thresholds; generalization to traditional IBI events is uncertain.The authors call for wider clinical validation before setting cut-offs and note that LVI criteria differ from context-dependent IBI definitions.Source: §4 · p. 13

7 Competing Interests

All authors are affiliated with CergenX Ltd, a University College Cork spin-out developing newborn neuromonitoring technologies.

7 Competing Interests

  • All authors are affiliated with CergenX Ltd, which develops neuromonitoring technologies for newborns.G.B.B. and S.G. are co-founders; R.H., A.L. and J.O.T. are employees, and S.M. and S.V. were independent consultants.Source: §7 · p. 13

A.1 Quantitative Features for IBI Detection

The appendix describes amplitude-based IBI detection features, including thresholded amplitude and frequency-weighted energy measures. It also notes limitations and implementation choices relevant to their use as detectors.

A.1 Quantitative Features for IBI Detection

  • Amplitude-based IBI measures quantify peak amplitude or average power, including the ratio of signal outside a 25 µV peak-to-peak range over a 1 s window.The thresholded measure requires a threshold to convert continuous values into a binary mask; the specified threshold is 12.5 µV.Source: §A.1 · p. 17
  • Frequency-weighted energy measures include the nonlinear energy operator, which can oscillate widely and produce negative values.The operator remains used for burst or IBI detection and is implemented in the NicoletOne ICU Monitor and a modified form in the ANT Neuro nëo monitor.Source: §A.1 · p. 17
  • Line length sums absolute signal differences across a 1 s window, but the authors omit its original normalisation factor because it degraded LVI-detection performance.The method is defined as a frequency-weighted energy measure for neonatal EEG.Source: §A.1 · p. 17
  • The envelope derivative operator is presented as a nonnegative instantaneous energy measure that avoids band-pass pre-processing and post-processing.The authors state that it addresses some issues of the nonlinear energy method while retaining non-negativity.Source: §A.1 · p. 17
  • The signal envelope is defined using the Hilbert transform, which the passage identifies as H[·].The envelope is introduced as another way to measure amplitude.Source: §A.1 · p. 17
  • Adaptive thresholding performed poorly in preliminary generalisation tests and requires assumptions about the EEG epoch being processed.The authors contrast it with static thresholding for continuous, unbounded quantitative measures.Source: §A.1 · p. 17

A.2 Results

The appendix reports validation and clinical-application analyses of LVI detection, including expert-equivalence testing and feature comparisons across EEG grades, HIE grades, seizure prediction, and seizure burden. It also describes amplitude-duration and minimum-duration analyses.

A.2 Results

  • Held-out validation detector performance is reported both across concatenated EEG files and per neonate, with per-neonate metrics summarized by median and interquartile range.The cross-validation figure likewise presents pooled results and per-neonate summaries across 10 folds.Source: Figure A1 · p. 18, Table A3 · p. 19, Table A4 · p. 19
  • Only ConvNeXt and SVM reached expert-level agreement in held-out validation testing.Table A5 reports changes in Cohen’s kappa and confidence intervals from bootstrapping; positive Δκ indicates increased agreement relative to human-expert agreement.Source: Table A5 · p. 19
  • Table A6 compares ConvNeXt-derived LVI features with the average of two human experts using bootstrapped differences and a one-sided non-inferiority test.The caption defines small p-values as indicating model error consistently lower than, or comparable to, expert inter-rater variability.Source: Table A6 · p. 20
  • Table A7 measures separation of six LVI features across four EEG background grades and three HIE grades plus a normal cohort.It uses ε2 for differences across EEG grades and Spearman’s ρ for monotonic associations across grades.Source: Table A7 · p. 20
  • Table A8 evaluates six LVI features for seizure prediction in 203 neonates and association with seizure burden in 121 neonates.It reports AUC, ε2, and Spearman’s ρ as measures suited to the respective comparisons.Source: Table A8 · p. 20
  • Across all three datasets, LVI amplitude trends downward as duration increases, while Figure A3 examines duration-category rates and performance after excluding shorter LVIs.The duration-threshold analyses cover EEG grades, HIE grades, and seizure prediction.Source: Figure A2 · p. 21, Figure A3 · p. 22
Loading 2610.12680v1…