U-NeXt: General-purpose event detection for neonatal EEG
Authors: Robert Hogan, Sean R. Mathieson, Aurel Luca, Soraia Ventura, Sean Griffin, Geraldine B. Boylan, John M. O'Toole
Organizations: CergenX Ltd, Cork, Ireland · INFANT Research Centre, University College Cork, Ireland · Department of Paediatrics and Child Health, University College Cork, Ireland
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.
Figures & tables
Figure 1: U-NeXt architecture. The 4-stage encoder (pale blue), adapted from ConvNeXt [ 8 ] , contains downsampling, normalisation, and convolution blocks. The 2-stage decoder (pale orange) contains upsampling and convolution blocks. The blocks use depthwise convolution ( d ), layer normalisation (LN), and Gaussian error linear unit (GELU) activation. Downsampling uses a 1×4 convolution; upsampling uses a 1×4 transposed convolution, both with stride 4. Skip connections concatenate encoder and decoder features and 2 decoder outputs provide deep supervision at different resolutions. W and D control width and depth, respectively.
ConvNeXt
U-NeXt
model
depth ( D )
width ( W )
# params
FLOP
# params
FLOP
Nano
1
1
38.7 k
1.9 m
50.1 k
638.78 k
Medium
3
4
1.7 m
84.3 m
1.9 m
23.8 m
XL
6
10
20.6 m
1 G
21.7 m
280.3 m
Table 1: Model size and compute for U-NeXt and equivalent ConvNeXt models [ 8 ] . Floating-point operation (FLOP) counts use one operation per multiply–add; U-NeXt uses 16 s inputs for direct comparison.
LVI ( n=51 )
Artefact ( n=98 )
Seizure ( n=202 )
Seizure+ ( n=76 )
Gestational age, weeks+days
40+2 (39+2 to 40+5)
39+6 (38+4 to 40+4) n=97
40+0 (39+2 to 40+5) n=197
40+0 (39+1 to 41+0)
Birth weight, kg
3.50 (3.33–3.85) n=26
3.31 (2.32–3.57) n=67
3.50 (3.27–3.76) n=77
–
Sex, female
18 (35.3%)
44 (45.4%) n=97
86 (42.8%) n=201
–
Clinical demographics
Primary diagnosis
Normal cohort
25 (49.0%)
0 (0%)
73 (41.0%)
–
Table 2: Cohort demographics: median (interquartile range) or number (percentage), unless stated otherwise; number of neonates ( n ) indicates incomplete data. LVI: low-voltage interval; HIE: hypoxic-ischaemic encephalopathy, with severity counts ordered mild/moderate/severe/unknown. The normal cohort comprises healthy postnatal-ward neonates [ 20 ] . EEG duration is per neonate; start is postnatal age at the first included EEG segment.
Figure 2: Comparison of U-NeXt and ConvNeXt on development-set validation data. Bars show pooled metrics; diamonds and whiskers show per-neonate medians and interquartile ranges. LVI used a Nano model with 10-fold cross-validation; artefact, Medium with 5-fold cross-validation; seizure, XL with a fixed 80:20 train/validation split. U-NeXt outperformed task-customised ConvNeXt baselines on every task despite using a common configuration.
Figure 3: (a) ConvNeXt and U-NeXt seizure probabilities with expert annotation over 3.5 hours from one recording. Shading marks positive annotations or predictions at a 0.5 threshold; scale bar indicates 10 minutes. U-NeXt produces a more coherent trace, accurate event boundaries, and better short-event sensitivity. (b) Sample sensitivity by annotated event duration across the three task-specific datasets. Dashed lines show relative sensitivity gains on task-specific right axes. Error bars: 95% percentile intervals from N=3,000 per-neonate bootstrap resamples.
Helsinki SVM
ConvNeXt
U-NeXt
U-NeXt+
Helsinki (n=79)
MCC
0.697
0.764
0.865
0.862
Pearson’s r
0.649
0.824
0.896
0.884
AUC
0.908
0.982
0.992
0.991
Sens./Spec. (%)
73.2/96.7
72.9/98.5
83.7/ 99.2
84.5 /99.0
PPV/NPV (%)
72.7/96.8
84.9/96.8
92.2 /98.1
90.7/ 98.2
Table 3: Seizure detection on held-out Helsinki and Cork datasets. U-NeXt models used whole Helsinki recordings and 10-minute Cork segments; Helsinki SVM used leave-one-neonate-out models. FP min/h: false-positive time per hour; FD/h: false detections per hour. Burden MAE: per-neonate mean absolute burden error (min/h). Δκ : change in inter-rater agreement when replacing one expert with the model; superscripts/subscripts give upper/lower 95% bounds from 10,000 neonate bootstrap resamples. Bold marks best value with U-NeXt models significantly improving on ConvNeXt.
Figure 4: (a) False-positive fragmentation in a seizure-free interval. ConvNeXt produces one detection; U-NeXt produces three but with 29.6% less false-positive time. The U-NeXt prediction is better but triples FD/h. (b) Impact of alarm configuration (minimum event duration and post-alarm silence period) on false alarms vs false detections in the prediction mask. Reasonable values can reduce false alarms by approximately 50% relative to FD/h. (c) Comparison of inter-rater FD/h and event sensitivity with model performance when using the same expert reference. The points and error bars show the mean and standard deviation across reference-rater choices; individual data points are shown for each reference choice.
TP
FP
FN
TN
Sens.
Spec.
PPV
NPV
MCC
Error MAE (min/h)
Cork ≥0.5 min/h
ConvNeXt
301
129
29
2103
91.2
94.2
70.0
98.6
0.766
6.53
U-NeXt
304
140
26
2092
92.1
93.7
68.5
98.8
0.760
4.84
U-NeXt+
309
96
21
2136
93.6
95.7
76.3
99.0
0.820
3.18
Table 4: Seizure-burden threshold classification against majority-consensus annotations. Segments are 1 h where possible (minimum 20 min; mean 59.4 min Cork, 53.9 min Helsinki). Error MAE: mean absolute predicted–reference burden difference (min/h), over FP/FN segments only.
Figure 5: (a) U-NeXt inference-time and compute gains over ConvNeXt for seizure detection on 1 hour of 18-channel EEG from Helsinki. (b) Effect of input segment size on U-NeXt seizure detection on Helsinki, relative to the 256 s training length. Burden metrics are calculated hourly. Shading shows interquartile ranges from N=5,000 neonate bootstrap resamples, each relative to its own 256 s baseline. On the logarithmic segment-size axis, performance improves rapidly from very short segments with limited EEG context, then stabilises with modest gains for longer segments.
Figure 6: Reduced-montage performance of U-NeXt+ on Cork (51 neonates). (a) Bipolar derivations and channel groups retained in each montage. (b) Changes in sample-based sensitivity, specificity, NPV, PPV, and event sensitivity, in percentage points. With two central channels, event sensitivity is 82.2% of its eight-channel value. A reference event is detected if it overlaps any predicted positive interval on the unanimous-consensus timeline. (c) Change in MCC. Changes are relative to the eight-channel reference montage. All configurations use maximum aggregation across retained channels. Shading shows pointwise interquartile range from N=5,000 neonate bootstrap resamples, each compared with its own eight-channel result.
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
CergenX Ltd, Cork, Ireland · INFANT Research Centre, University College Cork, Ireland · Department of Paediatrics and Child Health, University College Cork, Ireland