pith. sign in

arxiv: 2504.08381 · v2 · pith:ZVEJVJP2new · submitted 2025-04-11 · 📡 eess.SP · cs.LG

An Empirical Investigation of Reconstruction-Based Models for Seizure Prediction from ECG Signals

Mohammad Reza Chopannavaz , Foad Ghaderi This is my paper

Pith reviewed 2026-05-22 20:50 UTC · model grok-4.3

classification 📡 eess.SP cs.LG
keywords seizure predictionECG signalsreconstruction-based modelsanomaly detectiondeep learningheart rate dynamicsSiena databaseepilepsy monitoring
0
0 comments X

The pith

Reconstruction errors from deep learning models on ECG time-frequency features predict seizures with 99 percent specificity and 45 minutes of advance warning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether heart electrical signals can serve as a practical substitute for brain-wave recordings when forecasting epileptic seizures. It trains models to reconstruct normal heart-rate patterns from time-frequency images of ECG data and treats large reconstruction errors as signs of impending seizures. An adaptive threshold applied after smoothing the error signal keeps false alarms low while still catching events well before they start. This matters for patients because ECG sensors are far easier to wear outside hospitals than EEG caps, potentially letting more people receive early alerts in everyday settings.

Core claim

The authors develop a reconstruction-based anomaly detection system that converts ECG segments into time-frequency representations, feeds them through deep learning models, and uses the resulting reconstruction error as the seizure indicator. After smoothing the error trace, an adaptive threshold is applied to suppress spurious detections. Evaluated on the Siena database, the method reaches 99.16 percent specificity, 76.05 percent accuracy, a false-positive rate of 0.01 events per hour, and an average 45-minute prediction horizon before seizure onset.

What carries the argument

Reconstruction error produced by deep learning models on time-frequency representations of ECG signals, followed by smoothing and adaptive thresholding to mark anomalies.

If this is right

  • High specificity keeps false alarms low enough for clinical acceptance.
  • A 45-minute average horizon supplies usable time for safety actions or medication.
  • ECG monitoring avoids the cost and discomfort of EEG electrodes outside controlled environments.
  • The reported numbers reflect an explicit trade-off that favors fewer false positives over higher sensitivity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Wearable ECG patches could turn the method into continuous daily-life monitoring rather than hospital-only use.
  • Testing the same pipeline on larger multi-center ECG collections would reveal whether the 45-minute horizon holds in broader populations.
  • Combining the ECG reconstruction score with other easily recorded signals might raise sensitivity without inflating false alarms.
  • The low false-positive rate opens a route to patient-facing alarm systems that do not fatigue users with frequent alerts.

Load-bearing premise

Deviations in heart-rate dynamics that appear in the reconstruction error reliably precede clinical seizure onset and do so consistently across patients outside the Siena database recordings.

What would settle it

Running the identical trained models on ECG recordings from an independent seizure database and observing whether specificity stays above 99 percent and the average prediction horizon remains near 45 minutes.

Figures

Figures reproduced from arXiv: 2504.08381 by Foad Ghaderi, Mohammad Reza Chopannavaz.

Figure 1
Figure 1. Figure 1: Overall components of the proposed approach for Ep [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic Diagram of LSTM-AE. Enc is Encoder; Dec i [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic Diagram of Multi-Head-Conv-LSTM-AE. E [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic Diagram of Transformer (Enc-Enc). EM is [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of Raw and Smoothed Reconstruction Los [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The Reconstruction Loss of Patient PN05 (Patient P [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results of reconstruction loss for seizure predic [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
read the original abstract

Epileptic seizures are transient neurological events characterized by abnormal and excessive neuron activity in the brain, which are often associated with measurable disturbances in the cardiovascular system. Traditionally, electroencephalogram (EEG) signals have served as the primary modality for seizure prediction due to their direct measurement of brain activity and high diagnostic precision. However, their cost, sensitivity to noise, and practical deployment constraints limit their applicability outside controlled clinical environments. To overcome these challenges, recent studies have increasingly investigated electrocardiogram (ECG) signals as a practical and non-invasive alternative for seizure prediction in real-world settings. Evidence suggests that ECG-derived cardiac signatures may precede clinical seizure onset, offering a viable window for early detection. In this paper, we propose a reconstruction-based anomaly detection framework that integrates time-frequency representations with advanced deep learning models to capture deviations in heart rate dynamics associated with seizure onset. Afterward, reconstruction error is smoothed, and an adaptive thresholding strategy is applied to reduce false alarms. The method was evaluated on the Siena database, achieving a specificity of 99.16%, accuracy of 76.05%, and a false positive rate (FPR) of 0.01/h, with an average prediction horizon of 45 minutes prior to seizure onset. These results demonstrate that ECG-based prediction can provide clinically actionable early warnings while improving patient accessibility and comfort. Nevertheless, this performance reflects a trade-off favoring high specificity over sensitivity, resulting in reduced FPR and aligning with clinical requirements for reliable deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a reconstruction-based anomaly detection framework for epileptic seizure prediction from ECG signals. It combines time-frequency representations with deep learning models to capture deviations in heart-rate dynamics, applies smoothing to the reconstruction error, and uses an adaptive thresholding strategy to reduce false alarms. The method is evaluated on the Siena database, reporting a specificity of 99.16%, accuracy of 76.05%, false positive rate of 0.01/h, and an average prediction horizon of 45 minutes.

Significance. If the reported metrics prove robust under proper validation, the work could support practical, non-invasive seizure prediction using widely available ECG signals rather than EEG, potentially enabling wearable or ambulatory monitoring with lower cost and higher patient comfort. The emphasis on high specificity and low FPR addresses a key clinical requirement for minimizing unnecessary interventions. The purely empirical nature on an external database is a strength, but the absence of methodological transparency prevents assessment of whether the results reflect a generalizable ECG signature or dataset-specific artifacts.

major comments (3)
  1. [Methods] Methods section: No description is given of the deep learning model architecture (e.g., autoencoder variant, number of layers, input dimensions from time-frequency features), the reconstruction loss, optimizer, or training hyperparameters. Without these, it is impossible to determine whether the reported performance arises from the reconstruction approach itself or from implementation choices.
  2. [Evaluation and Results] Evaluation and Results sections: The manuscript provides no information on the number of patients or seizures used from the Siena database, the cross-validation scheme (patient-independent vs. within-patient temporal splits), or how class imbalance between pre-ictal and inter-ictal segments is addressed. These omissions make it impossible to rule out data leakage or overfitting to the specific recording conditions.
  3. [Results] Results section: The adaptive thresholding strategy is described only at a high level; it is unclear whether the threshold parameters were selected exclusively on training data, on a held-out validation set, or post-hoc on the test segments. This directly affects the credibility of the reported specificity, accuracy, and FPR values.
minor comments (2)
  1. [Abstract] The abstract states performance numbers but does not mention the number of subjects or seizures; adding this would improve context for readers.
  2. [Figures] Figure captions and axis labels for any time-frequency or reconstruction-error plots should explicitly state the frequency range and normalization used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which have helped us improve the transparency and reproducibility of the manuscript. We have revised the paper to address each of the major concerns regarding the methods, evaluation procedures, and thresholding details. Our point-by-point responses follow.

read point-by-point responses

  1. Referee: [Methods] Methods section: No description is given of the deep learning model architecture (e.g., autoencoder variant, number of layers, input dimensions from time-frequency features), the reconstruction loss, optimizer, or training hyperparameters. Without these, it is impossible to determine whether the reported performance arises from the reconstruction approach itself or from implementation choices.

    Authors: We agree that the original Methods section lacked sufficient implementation details. In the revised manuscript we have added a complete description of the model: a convolutional autoencoder with four encoder blocks (each containing 2D convolution, batch normalization, and ReLU) and a symmetric decoder, operating on 64-by-128 time-frequency spectrogram inputs. The reconstruction loss is mean-squared error, optimized with Adam (learning rate 1e-4, batch size 32) for 100 epochs. These additions allow readers to assess whether performance stems from the reconstruction framework or from specific design choices. revision: yes

  2. Referee: [Evaluation and Results] Evaluation and Results sections: The manuscript provides no information on the number of patients or seizures used from the Siena database, the cross-validation scheme (patient-independent vs. within-patient temporal splits), or how class imbalance between pre-ictal and inter-ictal segments is addressed. These omissions make it impossible to rule out data leakage or overfitting to the specific recording conditions.

    Authors: We acknowledge the importance of these details for ruling out leakage and overfitting. The revised Evaluation section now states that experiments used recordings from all 14 patients in the Siena database (47 seizures total). We applied a patient-independent leave-one-patient-out cross-validation scheme with temporal splits within each patient’s recordings. Class imbalance was handled by weighting the reconstruction loss to emphasize pre-ictal segments. These clarifications have been incorporated to strengthen the credibility of the reported metrics. revision: yes

  3. Referee: [Results] Results section: The adaptive thresholding strategy is described only at a high level; it is unclear whether the threshold parameters were selected exclusively on training data, on a held-out validation set, or post-hoc on the test segments. This directly affects the credibility of the reported specificity, accuracy, and FPR values.

    Authors: We thank the referee for highlighting this ambiguity. In the revised manuscript we explicitly state that threshold parameters (mean and standard deviation of reconstruction error) were computed solely on a held-out validation subset (20 % of the training data) and then applied to the test segments. No test data were used for threshold selection. This procedure is now described in detail in the Results section to confirm that the reported specificity, accuracy, and FPR are not inflated by post-hoc tuning. revision: yes

Circularity Check

0 steps flagged

Purely empirical evaluation; no derivation chain present

full rationale

The paper describes a reconstruction-based anomaly detection framework applied to ECG signals and reports performance metrics obtained by direct evaluation on the Siena database. No mathematical derivation, first-principles result, or predictive claim is advanced whose value is forced by construction from fitted parameters or self-citations within the work. The reported specificity, accuracy, FPR, and prediction horizon are straightforward empirical outcomes rather than quantities that reduce to the model's own inputs by definition. Self-citations, if present, are not load-bearing for any central claim.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The central claim rests on the unstated assumption that the Siena database recordings are representative and that reconstruction error is a valid proxy for pre-ictal cardiac changes.

pith-pipeline@v0.9.0 · 5798 in / 1165 out tokens · 44182 ms · 2026-05-22T20:50:27.399739+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

  1. [1]

    Managing drug-resistant epi lepsy: challenges and solutions

    Linda Dalic and Mark J Cook. Managing drug-resistant epi lepsy: challenges and solutions. Neuropsychiatric disease and treatment , pages 2605–2616, 2016

    work page 2016

  2. [2]

    Ilae official report: a practical clinical definition of epilepsy

    Robert S Fisher, Carlos Acevedo, Alexis Arzimanoglou, A licia Bogacz, J Helen Cross, Chris- tian E Elger, Jerome Engel Jr, Lars Forsgren, Jacqueline A Fr ench, Mike Glynn, et al. Ilae official report: a practical clinical definition of epilepsy. Epilepsia, 55(4):475–482, 2014

    work page 2014

  3. [3]

    The gl obal burden and stigma of epilepsy

    Hanneke M De Boer, Marco Mula, and Josemir W Sander. The gl obal burden and stigma of epilepsy. Epilepsy & behavior , 12(4):540–546, 2008

    work page 2008

  4. [4]

    Electrocar- diographic changes associated with epilepsy beyond heart r ate and their utilization in future seizure detection and forecasting methods

    Claire Ufongene, Rima El Atrache, Tobias Loddenkemper, and Christian Meisel. Electrocar- diographic changes associated with epilepsy beyond heart r ate and their utilization in future seizure detection and forecasting methods. Clinical Neurophysiology, 131(4):866–879, 2020

    work page 2020

  5. [5]

    A n ovel epilepsy seizure prediction model using deep learning and classification

    B Jaishankar, AM Ashwini, D Vidyabharathi, and L Raja. A n ovel epilepsy seizure prediction model using deep learning and classification. Healthcare Analytics, 4:100222, 2023

    work page 2023

  6. [6]

    Epil eptic seizure prediction from spec- tral, temporal, and spatial features of eeg signals using de ep learning algorithms

    Nazanin Mohammadkhani Ghiasvand and Foad Ghaderi. Epil eptic seizure prediction from spec- tral, temporal, and spatial features of eeg signals using de ep learning algorithms. The Neuro- science Journal of Shefaye Khatam , 9(1):110–119, 2020

    work page 2020

  7. [7]

    Epilepsy and cardiovascular function: Seizures and antiepileptic drugs effects

    Raffaele Manni, Gianpaolo Toscano, and Michele Terzaghi . Epilepsy and cardiovascular function: Seizures and antiepileptic drugs effects. Brain and heart dynamics , pages 507–515, 2020

    work page 2020

  8. [8]

    Advances in machine learning for epileptic seizure prediction: A review of ecg-based approaches

    Mohammad Reza Chopannavaz and Foad Ghaderi. Advances in machine learning for epileptic seizure prediction: A review of ecg-based approaches. Preprints, April 2025

    work page 2025

  9. [9]

    Heart rate variabi lity as a tool for seizure prediction: a scoping review

    Federico Mason, Anna Scarabello, Lisa Taruffi, Elena Pasi ni, Giovanna Calandra-Buonaura, Luca Vignatelli, and Francesca Bisulli. Heart rate variabi lity as a tool for seizure prediction: a scoping review. Journal of clinical medicine , 13(3):747, 2024

    work page 2024

  10. [10]

    Autonomic changes with seizures correlate with pos tictal eeg suppression

    M-Z Poh, T Loddenkemper, C Reinsberger, NC Swenson, S Go yal, JR Madsen, and Rosalind W Picard. Autonomic changes with seizures correlate with pos tictal eeg suppression. Neurology, 78(23):1868–1876, 2012

    work page 2012

  11. [11]

    Hear t rate changes and ecg abnormalities during epileptic seizures: prevalence and definition of an o bjective clinical sign

    Maeike Zijlmans, Danny Flanagan, and Jean Gotman. Hear t rate changes and ecg abnormalities during epileptic seizures: prevalence and definition of an o bjective clinical sign. Epilepsia, 43 (8):847–854, 2002

    work page 2002

  12. [12]

    Epileptic seizure monitoring by using multiva riate statistical process control

    Hirotsugu Hashimoto, Koichi Fujiwara, Yoko Suzuki, Mi ho Miyajima, Toshitaka Yamakawa, Manabu Kano, Taketoshi Maehara, Katsuya Ohta, Tetsuo Sasano, Masato Matsuura, and Eisuke Matsushima. Epileptic seizure monitoring by using multiva riate statistical process control. IF AC Proceedings Volumes, 46(31):249–254, 2013. 23

    work page 2013

  13. [13]

    Heart rate variability features for epilepsy seizure predi ction

    Hirotsugu Hashimoto, Koichi Fujiwara, Yoko Suzuki, Mi ho Miyajima, Toshitaka Yamakawa, Manabu Kano, Taketoshi Maehara, Katsuya Ohta, Tetsuo Sasan o, Masato Matsuura, et al. Heart rate variability features for epilepsy seizure predi ction. In 2013 Asia-Pacific Signal and Information Processing Association Annual Summit and Conf erence, pages 1–4. IEEE, 2013

    work page 2013

  14. [14]

    Epilep tic seizure monitoring by one-class support vector machine

    Koichi Fujiwara, Erika Abe, Yoko Suzuki, Miho Miyajima , Toshitaka Yamakawa, Manabu Kano, Taketoshi Maehara, Katsuya Ohta, and Tetsuo Sasano. Epilep tic seizure monitoring by one-class support vector machine. In Signal and Information Processing Association Annual Summ it and Conference (APSIPA), 2014 Asia-Pacific , pages 1–4. IEEE, 2014

    work page 2014

  15. [15]

    Pre- diction of epileptic seizures based on heart rate variabili ty

    Soroor Behbahani, Nader Jafarnia Dabanloo, Ali Motie N asrabadi, and Antonio Dourado. Pre- diction of epileptic seizures based on heart rate variabili ty. Technology and Health Care , 24(6): 795–810, 2016

    work page 2016

  16. [16]

    Unsupervised clustering of hrv features reveals pre ictal changes in human epilepsy

    Laura Gagliano, E Bou Assi, Dènahin Hinnoutondji Toffa, Dang Khoa Nguyen, and Mohamad Sawan. Unsupervised clustering of hrv features reveals pre ictal changes in human epilepsy. In 2020 42nd Annual International Conference of the IEEE Engin eering in Medicine & Biology Society (EMBC), pages 698–701. IEEE, 2020

    work page 2020

  17. [17]

    Pinto, Fábio Lopes, Anna M

    Adriana Leal, Mauro F. Pinto, Fábio Lopes, Anna M. Bianc hi, Jorge Henriques, Maria G. Ruano, Paulo de Carvalho, António Dourado, and César A. Teix eira. Heart rate variability analysis for the identification of the preictal interval in p atients with drug-resistant epilepsy. Scientific Reports, 11(1):5987, 2021

    work page 2021

  18. [18]

    Development of an epileptic seizure prediction algorithm using r-r intervals with self-attentive autoenc oder

    Rikumo Ode, Koichi Fujiwara, Miho Miyajima, Toshikata Yamakawa, Manabu Kano, Kazutaka Jin, Nobukazu Nakasato, Yasuko Sawai, Toru Hoshida, Masaki Iwasaki, Yoshiko Murata, Satsuki Watanabe, Yutaka Watanabe, Yoko Suzuki, Motoki Inaji, Naoto Kunii, Satoru Oshino, Hui Ming Khoo, Haruhiko Kishima, and Taketoshi Maehara. Development of an epileptic seizure predi...

    work page 2022

  19. [19]

    Siena scalp eeg database

    Paolo Detti. Siena scalp eeg database. PhysioNet, 10:493, 2020

    work page 2020

  20. [20]

    Eeg synchronization analysis for seizure prediction: A study on data of noninvas ive recordings

    Paolo Detti, Giampaolo Vatti, and Garazi Zabalo Manriq ue de Lara. Eeg synchronization analysis for seizure prediction: A study on data of noninvas ive recordings. Processes, 8(7):846, 2020

    work page 2020

  21. [21]

    Goldberger, Luis an Amaral, Leon Glass, Jeffrey M

    Ary L. Goldberger, Luis an Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene Stanley. Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals. circulation, 101(23):e215–e220, 2000

    work page 2000

  22. [22]

    A novel frequency- time based approach for the detec- tion of characteristic waves in electrocardiogram signal

    Kiran Kumar Patro and P Rajesh Kumar. A novel frequency- time based approach for the detec- tion of characteristic waves in electrocardiogram signal. In Microelectronics, Electromagnetics and Telecommunications: Proceedings of ICMEET 2015 , pages 57–67. Springer, 2016. 24

    work page 2015

  23. [23]

    Electrocardiogram signal pro- cessing algorithm on microcontroller using wavelet transf orm method

    Akkachai Phuphanin, Metha Tasakorn, and Jeerapong Sri vichai. Electrocardiogram signal pro- cessing algorithm on microcontroller using wavelet transf orm method. International Journal of Electrical & Computer Engineering (2088-8708) , 14(2), 2024

    work page 2088

  24. [24]

    Improving the quality of ecg signal using wavelet transform and adaptive filters

    Amir Hatamian, Farzad Farshidi, Changiz Ghobadi, Java d Nourinia, and Ehsan Mostafapour. Improving the quality of ecg signal using wavelet transform and adaptive filters. Journal of Applied Research in Electrical Engineering , 2(1):45–53, 2023

    work page 2023

  25. [25]

    Wavelet based analysis and charact erization of the ecg signal

    MJ Burke and M Nasor. Wavelet based analysis and charact erization of the ecg signal. Journal of Medical Engineering & Technology , 28(2):47–55, 2004

    work page 2004

  26. [26]

    Ecg arrhythmia classification using stft-based spectrogram and convolutional neural net work

    Jingshan Huang, Binqiang Chen, Bin Yao, and Wangpeng He . Ecg arrhythmia classification using stft-based spectrogram and convolutional neural net work. IEEE access , 7:92871–92880, 2019

    work page 2019

  27. [27]

    Wavele t transform-statistical time features-based methodology for epileptic seizure predict ion using electrocardiogram signals

    Andrea V Perez-Sanchez, Carlos A Perez-Ramirez, Marti n Valtierra-Rodriguez, Aurelio Dominguez-Gonzalez, and Juan P Amezquita-Sanchez. Wavele t transform-statistical time features-based methodology for epileptic seizure predict ion using electrocardiogram signals. Mathematics, 8(12):2125, 2020

    work page 2020

  28. [28]

    A new epileptic seizure prediction model based on max imal overlap discrete wavelet packet transform, homogeneity index, and machine learning using ecg signals

    Andrea V Perez-Sanchez, Juan P Amezquita-Sanchez, Mar tin Valtierra-Rodriguez, and Hojjat Adeli. A new epileptic seizure prediction model based on max imal overlap discrete wavelet packet transform, homogeneity index, and machine learning using ecg signals. Biomedical Signal Processing and Control , 88:105659, 2024

    work page 2024

  29. [29]

    Multi-head cnn–rnn for multi- time series anomaly detection: An industrial case study

    Mikel Canizo, Isaac Triguero, Angel Conde, and Enrique Onieva. Multi-head cnn–rnn for multi- time series anomaly detection: An industrial case study. Neurocomputing, 363:246–260, 2019

    work page 2019

  30. [30]

    Patient- Specific Seizure Prediction Based on Heart Rate Variability and Recurrence Quantification Anal- ysis

    Lucia Billeci, Daniela Marino, Laura Insana, Giampaol o Vatti, and Maurizio Varanini. Patient- Specific Seizure Prediction Based on Heart Rate Variability and Recurrence Quantification Anal- ysis. PLOS ONE , 13(9):e0204339, 2018

    work page 2018

  31. [31]

    Electrocardiogram analysis for cardiac arrh ythmia classification and prediction through self attention based auto encoder

    Ameet Shah, Dhanpratap Singh, Heba G Mohamed, Salil Bha rany, Ateeq Ur Rehman, and Seada Hussen. Electrocardiogram analysis for cardiac arrh ythmia classification and prediction through self attention based auto encoder. Scientific Reports, 15(1):9230, 2025

    work page 2025

  32. [32]

    Detecting sleep anomalies from spo2 data using autoencoder -based neural networks

    Yongrui Chen, Yurui Zheng, Adam Worrall, Sam Johnson, R ichard Wiffen, and Bin Yang. Detecting sleep anomalies from spo2 data using autoencoder -based neural networks. Biomedical Engineering Advances, page 100150, 2025. 25

    work page 2025