Deep learning for non-transvenous ICD screening: tailored neural network architectures, extended experiments, and explainability

Bartosz Morawski; Samuel J. Ward; Alex Touze; Lin-Thiri Toon; Paul R. Roberts; John Paisey; Alain B. Zemkoho

cs.LGOct 27, 2026

Deep learning for non-transvenous ICD screening: tailored neural network architectures, extended experiments, and explainability

Authors: Bartosz Morawski, Samuel J. Ward, Alex Touze, Lin-Thiri Toon, Paul R. Roberts, John Paisey, Alain B. Zemkoho

Organizations: School of Mathematical Sciences, University of Southampton, UK · School of Human Development and Health, Faculty of Medicine, University of Southampton, UK · Department of Cardiology, University Hospital Southampton NHS Foundation Trust, Southampton, UK

Abstract

This paper studies two non-transvenous implantable cardioverter-defibrillators (ICDs), namely the Subcutaneous ICD (S-ICD) and the Extravascular ICD (EV-ICD). Both avoid the complications of transvenous leads but are prone to oversensing, which can trigger inappropriate shocks associated with increased mortality. Oversensing risk is predicted by the ratio of T- or P-wave amplitude to R-wave amplitude (T:R, P:R), and these ratios vary over a 24-hour period. Yet S-ICD screening uses a single 10-second electrocardiogram (ECG) segment, and the EV-ICD has no formal screening protocol whatsoever. We present a 24-hour ambulatory screening framework for both devices which includes the first T:R analysis in an EV-ICD cohort and the first full evaluation of P:R prediction. We introduce AmpRatioNet, a family of 1D convolutional neural network architectures operating directly on the ECG signal, and benchmark it against the best existing approach - based on 2D phase space reconstruction images. Evaluation uses 4,605 clinician-annotated 10-second segments from 89 patients, with patient-level 10-fold cross-validation and bootstrap confidence intervals. AmpRatioNet-SE demonstrates superior performance on every metric in each setting with no device- or ratio-specific redesign. An explainability analysis against interpretable classical peak labelling algorithms shows that AmpRatioNet-SE's increase in error on oversensing-prone segments is up to 14 times smaller than theirs. Segment-level predictions are aggregated into a 24-hour timeline, extending S-ICD screening beyond 10 seconds and offering a plausible EV-ICD screening pathway pending device-level validation.

Figures & tables

Figure 1: Overview of the DeepICD modelling pipeline. (i) Shows the three sensing vector positions for S - ICD which are named primary 1–4, secondary 6–7 and alternate 2–3. Note that for EV - ICD , electrode 2 is lowered to be 8 cm superior to electrode 1. (ii) Illustrates the preprocessing steps. (iii) Shows a 1D and 2D CNN architecture that both output an amplitude ratio. (iv) Shows the full 24-hour timeline where the ratios for each 10-second sample are shown by the height of the bar. The ratios that exceed ${1}/{4}$ are coloured in yellow; those that exceed ${1}/{3}$ in orange; and those that exceed ${1}/{2}$ in red.

	AmpRatioNet-FCN		AmpRatioNet-RES		AmpRatioNet-SE
Stage	Operation	Output	Operation	Output	Operation	Output
Input	ECG signal	$1\times 5000$	ECG signal	$1\times 5000$	ECG signal	$1\times 5000$
Stem	Conv-BN-ReLU, $k=25$ , $s=2$	$32\times 2500$	Conv-BN-ReLU, $k=25$ , $s=2$	$32\times 2500$	Conv-BN-ReLU, $k=25$ , $s=2$	$32\times 2500$
Block 1	Conv-BN-ReLU, $k=15$ , $s=2$	$64\times 1250$	Residual block	$32\times 1250$	Residual block + SE	$32\times 1250$
Block 2	Conv-BN-ReLU, $k=7$ , $s=1$	$64\times 1250$	Residual block	$32\times 625$	Residual block + SE	$32\times 625$
Block 3	–	–	Residual block	$64\times 312$	Residual block + SE	$64\times 312$

Table 1: Comparison of the AmpRatioNet architectures used for raw ECG amplitude ratio regression. BN: Batch normalisation;

k

: kernel size;

s

: stride. All models receive a 10-second ECG segment sampled at 500 Hz. Residual blocks are defined in Table 2 , while the SE mechanism is defined in Table 3 .

Figure 2: Diagram of the AmpRatioNet-RES architecture, circles labelled with a plus and connected arrows denote skip connections. Convolutional layers are one dimensional as described in the dimension labels. The two-dimensional depictions in the diagram are for demonstrative purposes only. The AmpRatioNet-SE architecture is equivalent to this, with the addition of a squeeze-and-excitation block (see Table 3 ) within each ResBlock.

Layer	Type	Output	Kernel	Stride
1	Conv1D	$C_{\mathrm{out}}\times L$	15	1
2	BatchNorm	$C_{\mathrm{out}}\times L$	–	–
3	ReLU	$C_{\mathrm{out}}\times L$	–	–
4	Conv1D	$C_{\mathrm{out}}\times L$	15	1
5	BatchNorm	$C_{\mathrm{out}}\times L$	–	–
6	Shortcut	$C_{\mathrm{out}}\times L$	– / 1	– / 1

Table 2: Residual block used in the AmpRatioNet-RES and AmpRatioNet-SE architectures. The shortcut is an identity mapping when the input and output channel dimensions match; otherwise, a

1\times 1

convolution is used to match the channel dimension.

Layer	Type	Output size
1	Global Average Pooling	$C$
2	Dense $\left(\frac{C}{r}\right)$ + ReLU	$\frac{C}{r}$
3	Dense $(C)$ + Sigmoid	$C$
4	Reshape	$1\times C$
5	Channel-wise Scaling	$C\times L$

Table 3: Squeeze-and-excitation (SE) block used in AmpRatioNet-SE, with reduction ratio

r=8

. Adapted from [ 36 ] .

	ComplexCNN5		PSRNet32
Stage	Operation	Output	Operation	Output
Input	PSR image	$1\times 32\times 32$	PSR image	$1\times 32\times 32$
Block 1	Feature extraction block	$16\times 16\times 16$	Residual block	$16\times 32\times 32$
Downsample 1	–	–	Conv-BN-ReLU, $k=3\times 3$ , $s=2$	$32\times 16\times 16$
Block 2	Feature extraction block	$32\times 8\times 8$	Residual block	$32\times 16\times 16$
Downsample 2	–	–	Conv-BN-ReLU, $k=3\times 3$ , $s=2$	$64\times 8\times 8$

Table 4: Comparison of the two dimensional CNN architectures used for amplitude ratio regression from

32\times 32

Phase Space Reconstruction (PSR) images. BN: Batch normalisation;

k

: kernel size;

s

: stride. The ComplexCNN5 feature extraction block is defined in Table 6 , while the PSRNet32 residual block is defined in Table 5 .

Figure 3: Diagram of the PSRNet32 architecture, circles labelled with a plus and connected arrows denote skip connections.

Layer	Type	Output	Kernel	Stride
1	Conv2D	$C_{\mathrm{out}}\times H\times W$	$3\times 3$	1
2	BatchNorm	$C_{\mathrm{out}}\times H\times W$	–	–
3	ReLU	$C_{\mathrm{out}}\times H\times W$	–	–
4	Conv2D	$C_{\mathrm{out}}\times H\times W$	$3\times 3$	1
5	BatchNorm	$C_{\mathrm{out}}\times H\times W$	–	–
6	Shortcut	$C_{\mathrm{out}}\times H\times W$	– / $1\times 1$	– / 1

Table 5: Residual block used in PSRNet32. The shortcut is an identity mapping when the channel dimensions match; otherwise, a

1\times 1

convolution is used to match the output channel dimension.

Layer	Type	Output	Kernel	Stride
1	Conv2D	$2^{n+2}\times 2^{6-n}\times 2^{6-n}$	$(6-n)\times(6-n)$	1
2	BatchNorm	$2^{n+2}\times 2^{6-n}\times 2^{6-n}$	–	–
3	ReLU	$2^{n+2}\times 2^{6-n}\times 2^{6-n}$	–	–
4	Conv2D	$2^{n+2}\times 2^{6-n}\times 2^{6-n}$	$(6-n)\times(6-n)$	1
5	BatchNorm	$2^{n+2}\times 2^{6-n}\times 2^{6-n}$	–	–
6	ReLU	$2^{n+2}\times 2^{6-n}\times 2^{6-n}$	–	–

Table 6: ComplexCNN5 feature extraction block

n

, for

n\in\{1,\ldots,5\}

. The shortcut is an identity mapping except in the first block, where a

1\times 1

convolution matches the input channel dimension.

Setting	Architecture	RMSE [95% CI]	MAE [95% CI]	P95 AE [95% CI]	$R^{2}$ [95% CI]
EV-ICD P:R	ComplexCNN5	4.65 [3.52, 5.90]	2.89 [2.47, 3.41]	8.16 [6.55, 10.08]	0.602 [0.493, 0.665]
	PSRNet32	4.35 [3.01, 5.66]	2.52 [2.03, 3.12]	7.31 [5.39, 11.09]	0.651 [0.579, 0.716]
	AmpRatioNet-FCN	3.38 [2.75, 4.16]	2.29 [1.97, 2.71]	6.52 [5.57, 7.70]	0.790 [0.702, 0.844]
	AmpRatioNet-RES	2.15 [1.69, 2.71]	1.39 [1.14, 1.69]	4.53 [3.43, 5.78]	0.915 [0.867, 0.943]
	AmpRatioNet-SE	2.02 [1.57, 2.55]	1.24 [1.01, 1.53]	4.07 [3.12, 5.03]	0.925 [0.887, 0.950]
EV-ICD T:R	ComplexCNN5	4.58 [3.68, 5.49]	2.71 [2.31, 3.15]	8.25 [6.74, 9.77]	0.897 [0.855, 0.929]

Table 7: Architecture comparison across all three task–cohort settings. Error metrics are

\times 10^{-2}

in dimensionless ratio units (lower is better);

R^{2}

is unitless (higher is better). Point estimates are pooled out-of-fold values over 10 patient-level folds; intervals are 95% patient-level cluster bootstrap intervals (

B=10{,}000

). Intervals are marginal: rows share the same patients and folds, so overlapping intervals do not imply equivalent performance.

Figure 4: Plot of Predicted vs True amplitude ratios for each dataset and ratio task. Predictions have been aggregated from the estimates made on validation folds in the controlled architecture experiments. 8 samples exceed the depicted amplitude ratio bound of 1, 4 for S-ICD T:R, 3 for EV-ICD T:R and 1 for EV-ICD P:R.

	Whole	Class $\mathcal{B}$	Class $\mathcal{A}$	Gap	Runtime
AmpRatioNet-SE	0.0362	0.0213	0.0422	0.0209	8.2 s
ComplexCNN5	0.0577	0.0382	0.0660	0.0278	6.7 s
NeuroKit2	0.1758	0.0365	0.2297	0.1932	110.4 s
Pan-Tompkins	0.1901	0.1266	0.2173	0.0907	122.8 s

Table 8: Validation RMSE of T:R ratio predictions on the S - ICD dataset.

	Whole	Class $\mathcal{B}$	Class $\mathcal{A}$	Gap	Runtime
AmpRatioNet-SE	0.0279	0.0167	0.0323	0.0155	7.7 s
ComplexCNN5	0.0458	0.0335	0.0511	0.0177	6.1 s
NeuroKit2	0.1879	0.0300	0.2408	0.2108	110.7 s
Pan-Tompkins	0.1996	0.1386	0.2252	0.0866	121.1 s

Table 9: Validation RMSE of T:R ratio predictions on the EV - ICD dataset.

Figure 5: The distribution of validation error for each model on the SICD and EVICD datasets partitioned into Class

\mathcal{A}

and Class

\mathcal{B}

demonstrating the deep learning model’s robustness to oversensing. The box plots represent the min, max, median and interquartile range of validation errors where in the case of NeuroKit2 and Pan-Tompkins on Class

\mathcal{A}

the top of the boxes are omitted because they are so large they skew the scale of the plot.

Figure 6: In the timelines (left), each bar height represents the average T:R ratio of a 10-second sample. These are sampled for an S - ICD patient across 24-hours and three recorded ECG vectors. The ratios that exceed

{1}/{4}

are coloured in yellow; those that exceed

{1}/{3}

in orange; and those that exceed

{1}/{2}

in red. The corresponding histograms (right) show the frequency of each ratio in buckets of width

{1}/{12}

. In the future, such an output may support the decision that Vector A represents the most appropriate electrode positions.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Model	EV-ICD	S-ICD
ComplexCNN5	4.55	5.95
PSRNet32	4.80	5.64
PSRNet64	5.44	5.91

Appendix

Table 10: Comparison of PSR image resolutions for T:R ratio prediction across datasets. Values are pooled out-of-fold RMSE (

\times 10^{-2}

) across all validation samples.

	EV-ICD		S-ICD
Variant	T:R	P:R	T:R
RES	3.409	2.082	4.162
Dilated	3.109	2.237	3.838
Inception	3.579	2.248	4.314
SE	2.796	2.068	3.696

Appendix

Table 11: Exploratory AmpRatioNet variant experiments across all prediction tasks. Values represent pooled out-of-fold RMSE (

\times 10^{-2}

) from 10-fold patient-level split cross-validation.

Figure 7: Training Distribution of T:R ratios for EV-ICD patients.

Figure 8: Training Distribution of P:R ratios for EV-ICD patients.

Figure 9: Training Distribution of T:R ratios for S-ICD patients.

Electrocardiography

Electrocardiography studies the heart’s electrical signals to measure cardiac function and support diagnosis, monitoring, and research.

See all 111 papers

Sep 8, 2026

Physics-Informed Deep Learning for False Ventricular Tachycardia Alarm Reduction in the ICU

False ventricular tachycardia (VT) alarms are a leading contributor to alarm fatigue in intensive care units. We propose a deep learning framework combining a 1D SE-ResNet with ICU-realistic data augmentations and a physics-informed auxiliary reconstruction task based on the three-element Windkessel hemodynamic model, implemented as a differentiable forward simulation. By requiring the network's latent representation to produce physiologically plausible arterial pressure waveforms, artifact-driven ECG patterns are penalized while true VT remains coherent across modalities. Evaluated on the VTaC benchmark under a strict real-time protocol (10-second pre-alarm window), our method achieves a 5-point Challenge Score improvement over prior state-of-the-art. Ablation studies confirm that the physics-informed objective is the primary performance driver, providing gains in accuracy, 2x label efficiency, and more localized and clinically meaningful ECG segments.

Athanasios Papastathopoulos-Katsaros, Alexandra Stavrianidi, Zhandong Liu

Department of Pediatrics, Baylor College of Medicine, Houston, TX, USA · Jan and Dan Duncan Neurological Research Institute, Texas Children’s Hospital, Houston, TX, USA · Institute for Analysis and Numerics, University of Münster, Germany +1

Electrocardiogram Classification Physics-Informed ML

Ask PDF

Jul 27, 2026

Automated ECG Interval Measurement and Wave Delineation Using Fast Fourier Convolution ResNet

Accurate measurement of ECG intervals, including PR, QRS duration, and QT/QTc, is central to cardiac diagnosis, yet the published ECG delineation literature evaluates performance almost exclusively as fiducial-point timing errors on small curated databases, rather than as clinical interval accuracy on large unselected cohorts. We bridge this gap by evaluating a complete end-to-end pipeline on 10,646 clinical 12-lead ECGs and reporting the first large-scale interval measurement accuracy study with full statistical characterisation, including bias, 95% limits of agreement (Bland-Altman), bootstrap confidence intervals, and rhythm-stratified error analysis. The underlying delineation is performed by a Fast Fourier Convolution ResNet (FFCResNet), adapting local temporal convolutions with global spectral processing via FFT and augmented with register tokens for contextual feature learning. Three per-wave models (P, QRS, and T) are trained on six public databases with ECG-specific augmentation. On 10,646 ECGs, the system achieves a QT MAE of 17.5 ms [95% CI: 16.9-18.2], with a Bland-Altman bias of +8.5 ms (LoA: -68.5 to +85.5 ms); a QRS duration MAE of 14.8 ms [95% CI: 14.6-15.0], with a bias of +12.6 ms (LoA: -12.3 to +37.6 ms); and a ventricular rate MAE of 0.8 beats/min. All biases are statistically significant by the Wilcoxon signed-rank test (p < 0.001) but remain within or near published inter-observer variability bounds for sinus rhythms. Rhythm-stratified analysis reveals substantially higher QT errors for supraventricular tachycardias (SVT MAE: 75.0 ms; AVRT MAE: 85.3 ms) than for sinus bradycardia (SB MAE: 9.3 ms) and sinus rhythm (SR MAE: 8.9 ms), providing an honest characterisation of the deployment scope. Wave segmentation achieves internal Dice scores of 95.5%, 98.2%, and 96.1% for P, QRS, and T waves, respectively, and cross-database Dice scores of 78.1%, 85.5%, and 74.2%.

Farhan Adam Mukadam, Harshit Mishra, Nachiket Makwana +3

Indian Institute of Science, Bengaluru, India · Gauze, India · Waikato District Health Board, Hamilton, New Zealand +1

Electrocardiography Convolutional Neural Networks Electrocardiogram Interpretation

Ask PDF

Aug 16, 2026

Characterising cardiac tissue properties with graph neural networks

Characterising electrophysiological properties of cardiac tissue efficiently and accurately from spatially sparse intracardiac measurements is clinically important for localising ablation targets and improving arrhythmia treatment. We developed a graph neural network-based framework trained on synthetic electrogram signals on 2D flat surfaces to identify areas of interest in the context of cardiac ablation for premature ventricular complexes (PVCs). Our method achieved an average precision of 0.96, 0.97, and 0.95 for the detection of single-patch fibrosis, rapid depolarisation and high excitability, respectively. The trained model can then be applied to 2D curved surfaces with few-shot fine-tuning, demonstrating its generalisation capability. Future work will develop this framework further for clinical use in PVC ablation.

Ching-En Chiu, Yoo Ri Kim, Magdi Saba +2

Department of Electrical and Electronic Engineering, Imperial College London, London, United Kingdom · Department of Internal Medicine, Chonnam National University College of Medicine, Gwangju, Republic of Korea · St. George’s University Hospital, London, United Kingdom +2

Graph Neural Networks

Ask PDF

Abstract

Electrocardiography

Explore similar work

Physics-Informed Deep Learning for False Ventricular Tachycardia Alarm Reduction in the ICU

Automated ECG Interval Measurement and Wave Delineation Using Fast Fourier Convolution ResNet

Characterising cardiac tissue properties with graph neural networks