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

cs.LGOct 8, 2026

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

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

Organizations: CergenX Ltd, Cork, Ireland · INFANT Research Centre, University College Cork, Ireland · Department of Paediatrics and Child Health, University College Cork, Ireland

Abstract

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 ( $Δκ= 0.025$ , 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 $ρ=0.801$ ), a moderate association with HIE grades ( $ρ=0.384$ ), a weaker association with seizure burden ( $ρ=0.244$ ), 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.

Figures & tables

Figure 1: Examples of low-voltage interval (LVI) annotations (A) and datasets used for detector development and clinical evaluation (B). Panels A(i–iii) show LVI annotations within different patterns of background activity in a single EEG channel (CZ-C3). Panel A(iv) shows an example of time-synchronised LVIs. All EEG segments are 60 s and use the same amplitude scale. Panels B(i) and B(iii) show flowcharts of the datasets used for detector development and clinical evaluation, respectively. Panel B(ii) shows examples of the probability and binary outputs from two machine-learning models: SVM, a support vector machine, and ConvNeXt, a convolutional deep-learning network. Two readers (R1 and R2) annotated the LVIs in this example.

	development set	validation set
	(n=51)	(n=72)
Gestational Age (weeks)	40+2 (39+2 – 40+5)	39+6 (39+0 – 40+5)
Birth Weight (kg)	3.5 (3.33 – 3.85) [n=26]	3.64 (3.37 – 3.83) [n=23]
Sex, Female	18 (35.3%)	28 (38.9%)
HIE	26 (51.0%)	55 (76.3%)
HIE: Mild / Moderate / Severe	9 / 13 / 4	26 / 18 / 11

Table 1: Demographic information on the EEG data sets used to develop and validate the low-voltage interval (LVI) detectors.

Figure 2: Performance of the low-voltage interval (LVI) detectors on the held-out validation set. A: Performance metrics calculated from annotations concatenated across all EEG records (top) and on a per-neonate basis (bottom), with medians and interquartile ranges shown for the latter. Consensus annotations from two human experts were used. B: Expert-level equivalence testing. The

\Delta\kappa

distribution is the difference in agreement (Cohen’s

\kappa

) between the detector and each human annotation compared with the agreement between the two human experts. Positive values of

\Delta\kappa

indicate an increase in agreement between the detection method (labelled as AI here) and experts compared with human-level agreement. Both ConvNeXt and SVM detectors reach expert-level agreement. Key: SVM, support vector machine; EDO, energy derivative operator.

Figure 3: Examples of clinical relevance of LVIs. Panels A and B: 6 LVI features generated from

1\text{\,}\mathrm{h}

of EEG with effect-size measure epsilon-squared (

\varepsilon^{2}

) and Spearman’s rank correlation coefficient (

\rho

). Panel A: LVI features are generated for 4 different EEG grades of background: normal/mild (0/1), moderate (2), and major (3) abnormalities, and inactive EEG (4). Panel B: HIE grades are the based on the modified Sarnat score, in addition to EEG from healthy neonates (normals) for comparison. Panel C: Two LVI features generated across time for the long-duration EEG recordings by generating features on a 1 h epoch shifted by

15\text{\,}\mathrm{min}

. Panel C(ii): the first 24 hours of the plot in panel C(i) in more detail.

Figure 4: Third example of clinical relevance of LVIs. A: LVI features generated from the first hour of EEG used to separate neonates with seizures from neonates without seizures. Area under the receiver operating characteristic curve (AUC) and Spearman’s rank correlation coefficient (

\rho

) used as discrimination measures. B: LVI features generated over the whole EEG recording for those neonates with seizures. Association between LVI features and seizure burden measured using

\rho

and discrimination of 4 categories of seizure burden measured with epsilon-squared (

\varepsilon^{2}

Figure 5: Effect of epoch duration on LVI features and biomarker performance. A: Difference in feature values for epochs

<

1 hour relative to values calculated on the full 1-hour epoch. Short-duration epochs are generated within the full-duration epoch at a 75% overlap rate. Absolute relative error is computed by comparing each short-duration to the full-duration epoch. The resulting error distribution is outlined with percentiles at the 50th (median), 75th, and 90th mark. This analysis uses the open-access EEG grades dataset. B: Performance metrics, either Spearman’s rank correlation coefficient or area under the receiver operating characteristic curve (AUC), for shorter-duration epochs compared to the full-duration epoch. This analysis is applied to three applications: separating EEG grades, separating HIE clinical grades, and predicting seizures. Performance metrics are calculated using one short-duration epoch, randomly sampled from each full-duration epoch and repeated

10,000

times. Median values are represented as markers and the 95% resampling interval as the shaded region.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

	application set
	( $n=391$ )
Gestational Age (weeks)	40+0 (39+1 – 41+0) [ $n=386$ ]
Birth Weight (kg)	3.5 (3.14 – 3.79) [ $n=151$ ]
Sex, Female	138 (43.4%) [ $n=318$ ]
HIE	222 (56.8%)
HIE: Mild / Moderate / Severe / Unknown	96 / 94 / 31 / 1

Appendix

Table A1 : Demographic information on the EEG data sets used for the application of the LVI detectors.

Dataset	Reviewer	Number of	Prevalence	Duration (seconds)
		Events		median (IQR)
Development	R1	36,580	0.202	5.48 (3.55–9.31)
Validation	R2	5,845	0.306	4.59 (2.94–9.19)
Validation	R3	8,737	0.299	3.20 (2.36–6.06)

Appendix

Table A2 : Summary of LVI annotations for the development and validation datasets.

Figure A1 : Cross-validation testing results for the LVI detectors across the 10 cross-validation folds. Performance metrics are calculated from the predictions concatenated across all EEG records in the top row and on a per-neonate basis in the bottom row. For the per-neonate results, metrics are summarised by median with error bars representing the interquartile range.

	AUC	AP	$r$	accuracy	MCC	$k$	sens./spec.	PPV/NPV
				(%)			(%)	(%)
ConvNeXt	0.996	0.943	0.992	96.9	0.926	0.926	95.9 / 97.4	93.7 / 98.3
SVM	0.996	0.895	0.990	96.8	0.923	0.923	94.4 / 97.8	94.6 / 97.7
Peak-to-peak	0.983	0.673	0.943	93.0	0.845	0.838	97.5 / 91.1	81.8 / 98.9
EDO	0.988	0.014	0.971	91.1	0.813	0.800	97.9 / 88.4	77.5 / 99.0
Envelope	0.987	0.176	0.975	91.7	0.819	0.810	96.9 / 89.5	79.0 / 98.6

Appendix

Table A3 : Held-out validation performance of different LVI detectors. Data are concatenated across all EEG files. Metrics calculated using consensus annotations from two human experts.

	AUC	$r$	accuracy (%)	MCC
ConvNeXt	0.994	0.788	98.7	0.736
	(0.987 to 0.998)	(0.652 to 0.867)	(95.7 to 99.6)	(0.629 to 0.849)
SVM	0.993	0.559	98.6	0.703
	(0.985 to 0.996)	(0.335 to 0.750)	(95.2 to 99.6)	(0.556 to 0.812)
Peak-to-peak	0.955	0.319	93.8	0.470
	(0.905 to 0.975)	(0.154 to 0.503)	(90.1 to 97.9)	(0.359 to 0.676)

Appendix

Table A4 : Held-out validation performance of different LVI detectors. Data are summarised per neonate as median (interquartile range). Metrics calculated using consensus annotations from two human experts.

	$\Delta\kappa$	95% CI	$p$ -value
ConvNeXt	0.024	0.011 to 0.040	$<$ 0.001
SVM	0.020	0.008 to 0.034	$<$ 0.001
Peak-to-peak	-0.049	-0.073 to -0.025	0.774
EDO	-0.081	-0.104 to -0.058	0.999
Envelope	-0.074	-0.100 to -0.049	0.991
Line Length	-0.114	-0.140 to -0.087	$>$ 0.999

Appendix

Table A5 : Expert-level equivalence testing of low-voltage interval detection on the held-out validation set. Mean changes in Cohen’s kappa (

\Delta\kappa

) and 95% confidence intervals (CI) quantify the change in inter-rater agreement when one human annotation is replaced by detector predictions, relative to agreement between the two human experts. Positive values of

\Delta\kappa

indicate an increase in agreement between the detector and each human annotation. Only the ConvNeXt and the SVM detectors reach expert-level agreement. Distributions are generated using a bootstrapping process with 10k sampling iterations.

	Mean difference	95% CI	$p$ -value
Maximum duration (s)	0.100	-0.686 to 0.856	0.583
Median duration (s)	0.172	-0.188 to 0.594	0.777
Percentage (%)	-1.291	-2.818 to -0.245	0.023
Number of events	-24.489	-34.558 to -11.589	$<$ 0.001
Complexity	-0.036	-0.068 to -0.019	0.001
Periodicity	-0.001	-0.002 to -0.000	0.013

Appendix

Table A6 : Difference in LVI features calculated on the held-out validation set comparing the features generated from the ConvNeXt prediction mask to the average of the two human experts. Mean differences and 95% confidence intervals (CI) were generated from bootstrapped sampling across neonates. A one-sided non-inferiority test determines whether ML model error was greater than or equal to human inter-rater error; small

p

-values indicate that model error was consistently lower than, or comparable to, expert inter-rater variability.

	EEG background grades		HIE grades & normal cohort
feature	effect size	correlation	effect size	correlation
	( $\varepsilon^{2}$ )	( $\rho$ )	( $\varepsilon^{2}$ )	( $\rho$ )
Maximum duration (s)	0.521	0.714	0.269	0.311
Median duration (s)	0.676	0.821	0.248	0.349
Percent (%)	0.643	0.801	0.297	0.384
Number	0.523	0.630	0.164	0.250

Appendix

Table A7 : Measures of separation for the six LVI features across (a) the four EEG background grades and (b) the three clinical HIE grades and the normal cohort. The effect-size measure

\varepsilon^{2}

quantifies the difference across the four grades and Spearman’s correlation coefficient (

\rho

) quantifies the monotonic relationship across grades.

	Seizure prediction		Association with seizure burden
feature	AUC	correlation	effect size	correlation
		( $\rho$ )	( $\varepsilon^{2}$ )	( $\rho$ )
Maximum duration (s)	$0.850^{***}$	$0.447^{***}$	$0.116^{**}$	$0.258^{**}$
Median duration (s)	$0.841^{***}$	$0.437^{***}$	$0.136^{**}$	$0.294^{**}$
Percent (%)	$0.872^{***}$	$0.475^{***}$	$0.136^{**}$	$0.255^{**}$
Number	$0.816^{***}$	$0.404^{***}$	$0.120^{**}$	$0.163$

Appendix

Table A8 : Measures of association for the six LVI features for a) the seizure prediction task (203 neonates) and b) association with seizure burden (121 neonates). The area under the receiver operating characteristic curve (AUC) measures detection performance, the effect-size measure (

\varepsilon^{2}

) quantifies the difference between groups, and Spearman’s correlation coefficient (

\rho

) quantifies the monotonic relationship between the seizure-neonate groups and seizure burden.

Figure A2 : Association between the duration of LVIs and EEG amplitude. Across all three datasets, the trend indicates decreasing amplitude with increasing duration. Amplitude for each LVI was measured as the mean value of the EEG envelope. LVI events were grouped into half-octave bins on the logarithmic duration scale. For each bin, amplitude–duration pairs were weighted inversely by the number of events contributed by each neonate, ensuring that every neonate contributed equal total weight within that bin. The neonate-weighted median amplitude and median duration for each bin were plotted with an amplitude interval bounded by the weighted 5th and 95th percentiles. Bins with fewer than five neonates were excluded from the summary. A line was fitted to the un-binned data pairs using a Huber loss function to limit the influence of outliers. The model fitted amplitude as a linear function of a

\log_{10}

transformed duration. Amplitude–duration pairs were weighted inversely by each neonate’s total number of LVIs, giving every neonate equal total influence on the fitted line.

Figure A3 : Analysis of LVI duration. A: Distribution of the different categories of LVI duration across three datasets. The number of LVIs within a duration category is calculated per neonate, normalised to the total available EEG channel-hours. The stacked bars display the mean per-neonate event rate for each duration category within each grade or class. B: Performance of LVI features when setting a minimum-duration threshold on included LVI events. LVIs shorter than this threshold were removed, therefore testing the importance of, or robustness to, including shorter-duration LVIs. Three applications are included, namely the EEG grades of background activity, clinical HIE grades, and a seizure prediction task. In addition to the ordinal analysis across all grades for EEG and HIE, a secondary binary comparison is included to distinguish specific severity sub-groups: grade 0/1 versus 2 for EEG grades and mild versus moderate HIE.

Electroencephalography

Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.

See all 178 papers

Oct 23, 2026

U-NeXt: General-purpose event detection for neonatal EEG

We developed U-NeXt, a general framework for neonatal electroencephalography (EEG) event detection based on semantic segmentation. Its convolutional encoder-decoder directly segments individual channels using long input segments to preserve temporal context. A shared architecture and training framework were developed using low-voltage interval, artefact, and seizure datasets containing 51, 98, and 278 neonates, respectively, comprising over 11,000 hours of EEG and 82,000 annotated events. Performance was compared with task-specific ConvNeXt classifiers. For seizure detection, we also evaluated U-NeXt+, trained on an expanded dataset. Both U-NeXt variants and the ConvNeXt baseline were evaluated on two independent held-out seizure datasets, Helsinki and Cork, comprising 130 neonates, with recordings annotated by three experts. U-NeXt improved detection across all three event types, with particular gains in sensitivity for short events. U-NeXt+ achieved Matthews correlation coefficients of 0.862 and 0.891 on Helsinki and Cork, respectively, and improved hourly seizure burden correlations by 30-40% relative to ConvNeXt. U-NeXt+ met the criterion for expert-level performance on both datasets (p < 0.001). In separate computational benchmarks, U-NeXt reduced batch inference time relative to ConvNeXt by 254-fold on a central processing unit and 200-fold on a graphics processing unit. A common segmentation framework improves detection across diverse neonatal EEG event types while substantially reducing inference time. Accurate seizure burden estimates and performance comparable to expert reviewers support its potential use in continuous monitoring and clinical decision support.

Robert Hogan, Sean R. Mathieson, Aurel Luca +4

CergenX Ltd, Cork, Ireland · INFANT Research Centre, University College Cork, Ireland · Department of Paediatrics and Child Health, University College Cork, Ireland

Electroencephalography EEG Seizure Detection Semantic Segmentation

Ask PDF

Nov 18, 2025

Interpretable Multimodal EEG-ECG Framework for Neonatal Seizure Prediction and Patient-Adaptive Transfer Learning

Neonatal seizures carry substantial risks of acute brain injury and long-term neurodevelopmental impairment, motivating the need for timely identification of impending seizure activity. While continuous electroencephalogram (EEG) monitoring remains the clinical gold standard, reliable patient-independent seizure prediction in neonates remains largely under-explored. We propose an interpretable multimodal framework that combines EEG and electrocardiogram (ECG) signals for neonatal seizure prediction. Our method uses mel-frequency cepstral coefficient (MFCC) representations processed by a lightweight (approx. 70k-parameter) convolutional neural network with squeeze-and-excitation attention. The framework was evaluated on the Helsinki neonatal EEG-ECG dataset and the Siena scalp EEG-ECG dataset. Under strict leave-one-patient-out (LOPO) evaluation, patient-independent prediction proved challenging (mean accuracy 58.5%, F1-score 60.3%), highlighting substantial inter-subject variability in neonatal EEG. However, fine-tuning the pretrained model with only five minutes of patient-specific data per class improved performance to 95.87% accuracy and 95.85% F1-score, demonstrating the feasibility of rapid patient-adaptive transfer learning. SHapley Additive exPlanations (SHAP) identified informative EEG channels that showed partial spatial agreement with clinically reported seizure origins. These findings suggest that fully patient-independent neonatal seizure prediction remains challenging, but that patient-adaptive fine-tuning with a small amount of data may provide a practical pathway toward personalized neonatal seizure prediction systems.

Sithmini Ranasingha, Agasthi Haputhanthri, Hansa Marasinghe +5

Department of Electronic and Telecommunication Engineering, University of Moratuwa, Sri Lanka · Department of Electrical and Computer Engineering, University of British Columbia, Canada · Department of Biomedical Engineering, University of Melbourne, Australia +1

EEG Seizure Detection Cross-Subject EEG Decoding Transfer Learning

Ask PDF

Aug 25, 2026

Attention-Enhanced Dual-Branch ConvNeXt-BiLSTM Network for Subject-Independent EEG Seizure Detection

Automated seizure detection from scalp electroencephalography (EEG) is difficult because seizure morphology varies among patients and seizure samples are substantially outnumbered by non-seizure samples. This paper presents an attention-enhanced dual-branch network that jointly learns time--frequency and temporal representations from the same EEG segment. A continuous wavelet transform converts each segment into a scalogram processed by an ImageNet-pretrained ConvNeXt-Tiny backbone and squeeze-and-excitation attention. In parallel, a bidirectional long short-term memory network followed by multi-head self-attention models the raw signal. The two feature vectors are concatenated and classified by a weighted multilayer perceptron. Experiments use 14 subjects from the CHB-MIT scalp EEG database with subject-wise partitioning performed before overlapping segmentation. The model obtains

97.88\%

accuracy and

97.51\%

F1-score over ten across-subject splits, and

97.51\%

accuracy,

96.59\%

F1-score, and

98.02\%

area under the ROC curve under 14-fold leave-one-subject-out validation. Removing temporal attention causes the largest ablation loss. The model requires 28.26 million parameters and 4.56 GFLOPs, with a measured network-only inference latency of 4.64 ms per segment.

Maimuna Chowdhury, Sk. Imran Hossain

Dept. of Computer Science and Engineering Khulna University of Engineering & Technology Khulna, Bangladesh

EEG Seizure Detection Electroencephalography Convolutional Neural Networks

Ask PDF

Abstract

Electroencephalography

Explore similar work

U-NeXt: General-purpose event detection for neonatal EEG

Interpretable Multimodal EEG-ECG Framework for Neonatal Seizure Prediction and Patient-Adaptive Transfer Learning

Attention-Enhanced Dual-Branch ConvNeXt-BiLSTM Network for Subject-Independent EEG Seizure Detection