A dual-system approach for epilepsy diagnosis: integrating mamba-Bi-LSTM architecture with SHAP-based verification

Fuchang Luo; Jia Xie; Mufeng Chen; Quansheng Ren

arxiv: 2606.08240 · v1 · pith:3H7FBJS2new · submitted 2026-06-06 · 📡 eess.SP

A dual-system approach for epilepsy diagnosis: integrating mamba-Bi-LSTM architecture with SHAP-based verification

Mufeng Chen , Jia Xie , Fuchang Luo , Quansheng Ren This is my paper

Pith reviewed 2026-06-27 19:21 UTC · model grok-4.3

classification 📡 eess.SP

keywords epilepsyEEGdeep learningMambaBi-LSTMSHAPdiagnosis systemmedical AI

0 comments

The pith

A Mamba-Bi-LSTM model with SHAP verification achieves 98.7% accuracy in epilepsy diagnosis.

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

This paper develops a dual-system AI framework for epilepsy diagnosis that combines EEG signals with clinical text data. The main discrimination component integrates the Mamba architecture and Bi-LSTM to reach 98.7 percent accuracy and 0.992 F1 score. The verification component uses SHAP to provide explanations and feedback, improving overall accuracy by 5.1 percent with 220 millisecond processing time. A sympathetic reader cares because traditional EEG interpretation is slow and experience-dependent, and this system offers higher accuracy, speed, and interpretability for a condition with sudden seizures.

Core claim

The paper claims that the dual-system approach, with its Mamba-Bi-LSTM main model for heterogeneous data analysis and SHAP verification for explainability, surpasses existing methods by raising accuracy from 92.6% to 98.7% and F1 from 0.895 to 0.992, while the verification integration adds 5.1% to overall accuracy in an average of 220 ms.

What carries the argument

The dual-system intelligent diagnosis framework that fuses Mamba-Bi-LSTM for main discrimination of multi-source data and SHAP for verification and feedback.

If this is right

The system establishes a cross-modal database for EEG and clinical text fusion.
High-confidence predictions serve as automatic feedback to optimize the model.
The verification system enhances credibility through explainable diagnostic basis.
Processing time for feedback integration averages 220 ms.

Where Pith is reading between the lines

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

This could enable real-time clinical decision support if integrated into hospital systems.
Similar dual architectures might apply to diagnosis of other seizure-related or neurological conditions.
Continuous model improvement via feedback loops may reduce the need for large new labeled datasets.

Load-bearing premise

The cross-modal database and its test splits accurately represent real-world clinical populations and the accuracy gains hold without overfitting.

What would settle it

Evaluation on a new, independent dataset from multiple hospitals with varied patient demographics showing accuracy below 95% or no gain over existing methods would falsify the improvement claim.

read the original abstract

This study develops a medical AI-assisted diagnosis system based on deep learning, which provides intelligent diagnostic solutions for epilepsy, a disease that seriously threatens the life and health of patients. Epilepsy has sudden and unpredictable seizures. Traditional diagnostic methods mainly rely on doctors' manual interpretation of EEG, which is time-consuming and dependent by experience. In response to the above challenges, this study designed a dual-system intelligent diagnosis framework, which includes two core components: the main discrimination system and the verification system. The main discrimination system uses a deep learning model that combines the innovative Mamba architecture with the Bi-LSTM structure to integrate and analyze heterogeneous data to achieve extremely high diagnostic accuracy; the verification system provides an explainable diagnostic basis through the SHAP method to enhance the credibility of the results. This system establishes a cross-modal database to realize intelligent analysis of multi-source heterogeneous data-fusion EEG signals and clinical text data for epilepsy. The system outputs results based on diagnostic consistency and confidence levels, and high-confidence predictions can also be used as automatic feedback sources to optimize the model. The experimental results show that the accuracy of the main discriminant model of the intelligent diagnosis system for epilepsy has increased from 92.6% to 98.7% and the F1 score has increased from 0.895 to 0.992, all of which have exceeded the existing optimal methods; the average processing time for verification system feedback integration is only 220 ms, which increases the overall diagnostic accuracy by 5.1%.

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.

Desk Editor's Note private letter to a colleague

The 98.7% accuracy and 5.1% gain claims rest on an unreported dataset with no patient counts, splits, or validation details, so they cannot be assessed.

read the letter

The headline accuracy claims in this epilepsy diagnosis paper are difficult to evaluate because the abstract and methods provide no details on the dataset size, number of patients, recording lengths, or how the train/test splits were performed. Without that, the jump from 92.6% to 98.7% accuracy and the 5.1% overall gain could easily be the result of optimistic partitioning rather than a genuine advance.

What the paper does is apply the Mamba architecture combined with Bi-LSTM to fuse EEG signals with clinical text data in a dual-system setup, where the main model classifies and a verification module uses SHAP to provide explanations and potentially feedback for retraining. The dual-system idea with consistency checks is a reasonable engineering choice for building trust in a clinical tool.

It does a decent job of describing the model architecture and the role of SHAP for verification. The reported processing time of 220 ms for the verification step is a concrete number that could be useful if the rest of the setup were clear.

The soft spots are substantial though. There are no ablation results to isolate the contribution of Mamba versus Bi-LSTM, no error bars or statistical tests on the performance numbers, and no description of the cross-modal database at all. The baselines are mentioned as "existing optimal methods" but not specified or reimplemented. This makes the superiority claim hard to verify. Since everything is on data from the same collection, the circularity is a real issue for generalization claims.

This kind of work is aimed at researchers in biomedical signal processing and medical AI who might want to try similar fusion approaches. A reader looking for a new technique or first-principles insight won't find it here, but someone needing an example of multi-modal epilepsy classification could get some ideas if the paper had the missing details.

I would not recommend sending this to peer review until the authors add a full experimental section with dataset stats, validation procedures, and proper comparisons. It doesn't meet the bar for a serious referee in its current state.

Referee Report

3 major / 1 minor

Summary. The paper proposes a dual-system AI framework for epilepsy diagnosis: a main discrimination system combining Mamba architecture with Bi-LSTM for fusing EEG signals and clinical text data, plus a verification system using SHAP for explainable outputs. It claims this yields 98.7% accuracy (up from 92.6%), 0.992 F1 (up from 0.895), exceeds prior optimal methods, adds 5.1% overall accuracy via feedback integration, and runs verification in 220 ms on a cross-modal database.

Significance. If the performance claims were supported by adequate validation, the Mamba-Bi-LSTM fusion for heterogeneous epilepsy data and SHAP verification could offer a practical advance in automated diagnosis with improved explainability. No machine-checked proofs, reproducible code, or parameter-free derivations are present to strengthen the assessment.

major comments (3)

[Abstract] Abstract: The headline claims (accuracy rising from 92.6% to 98.7%, +5.1% overall lift, 220 ms latency) are asserted without any description of the cross-modal EEG+text database, including patient count, recording count, train/test split ratios, cross-validation procedure, or external cohorts; this directly undermines evaluation of the central superiority claim.
[Abstract] Abstract/Experiments (implied): No baseline implementations, ablation results, statistical significance tests, or error bars are referenced to isolate the contribution of the Mamba-Bi-LSTM architecture or the dual-system feedback loop from the reported 5.1% gain.
[Abstract] Abstract: Performance figures are obtained on data drawn from the same collection used to train and tune the model with no external benchmark referenced, so the quoted accuracy cannot be distinguished from optimistic partitioning or overfitting on the training distribution.

minor comments (1)

[Abstract] The abstract would benefit from a single sentence summarizing dataset scale and diversity to allow readers to gauge clinical relevance.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and commit to revisions that strengthen the description of our experimental protocol and validation strategy without altering the core claims.

read point-by-point responses

Referee: [Abstract] Abstract: The headline claims (accuracy rising from 92.6% to 98.7%, +5.1% overall lift, 220 ms latency) are asserted without any description of the cross-modal EEG+text database, including patient count, recording count, train/test split ratios, cross-validation procedure, or external cohorts; this directly undermines evaluation of the central superiority claim.

Authors: We agree that the abstract and main text require expanded dataset documentation to support the performance claims. The revised manuscript will include a dedicated subsection detailing the cross-modal database construction, patient and recording counts, train/test split ratios, cross-validation procedure, and any external cohorts employed. revision: yes
Referee: [Abstract] Abstract/Experiments (implied): No baseline implementations, ablation results, statistical significance tests, or error bars are referenced to isolate the contribution of the Mamba-Bi-LSTM architecture or the dual-system feedback loop from the reported 5.1% gain.

Authors: We acknowledge the absence of explicit baseline comparisons, ablation studies, statistical tests, and error bars in the current version. The revision will incorporate these elements, including implementation details for baselines, ablation configurations, p-values or confidence intervals, and error bars to quantify the isolated contributions of the Mamba-Bi-LSTM fusion and feedback loop. revision: yes
Referee: [Abstract] Abstract: Performance figures are obtained on data drawn from the same collection used to train and tune the model with no external benchmark referenced, so the quoted accuracy cannot be distinguished from optimistic partitioning or overfitting on the training distribution.

Authors: The reported metrics derive from internal cross-validation on the collected multi-modal dataset. In revision we will explicitly describe the partitioning and cross-validation strategy and add a limitations paragraph discussing risks of distribution shift and overfitting, while explaining how the dual-system verification is intended to provide robustness against such issues. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical results lack derivation chain

full rationale

The paper presents an applied ML system (Mamba-Bi-LSTM + SHAP) whose headline claims are experimental accuracy numbers on an internal cross-modal EEG+text database. No first-principles derivation, uniqueness theorem, ansatz, or self-citation load-bearing step is described in the supplied text. The accuracy lift (92.6 % → 98.7 %) is reported as an empirical outcome rather than a quantity derived by construction from fitted parameters or prior self-citations. Because no load-bearing mathematical step reduces to its own inputs, the circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central performance claim rests on a large number of neural-network parameters fitted to the specific dataset, plus standard machine-learning assumptions about data distribution and the fidelity of SHAP attributions. No new physical or mathematical entities are postulated.

free parameters (2)

neural network weights and hyperparameters
All weights in the Mamba-Bi-LSTM model plus any fusion or regularization parameters are fitted to the epilepsy dataset to reach the stated accuracy.
data-fusion and confidence thresholds
Parameters controlling how EEG and text modalities are combined and when high-confidence outputs are fed back are chosen or tuned on the same data.

axioms (2)

domain assumption Training and test samples are drawn i.i.d. from the same distribution
Implicit in all supervised learning claims and required for the reported accuracy to generalize.
domain assumption SHAP attributions correctly identify the features the model actually uses
Taken as given for the verification system without additional validation.

pith-pipeline@v0.9.1-grok · 5814 in / 1556 out tokens · 24578 ms · 2026-06-27T19:21:59.107337+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 6 canonical work pages

[1]

Lipton, et al., Learning to diagnose with LSTM recurrent neural networks, arXiv (2015)

Zachary C. Lipton, et al., Learning to diagnose with LSTM recurrent neural networks, arXiv (2015)

2015
[2]

Lawhern, et al., EEGNet: a compact convolutional neural network for EEG-based brain – computer interfaces, J

Vernon J. Lawhern, et al., EEGNet: a compact convolutional neural network for EEG-based brain – computer interfaces, J. Neural Eng. 15 (5) (2018) 056013

2018
[3]

Pierre Vanabelle, et al., Epileptic seizure detection using EEG signals and extreme gradient boosting, J. Biomed. Res. 34 (3) (2020) 228 – 239

2020
[4]

Signal Process

Sheng Wong, et al., Channel-annotated deep learning for enhanced interpretability in EEG-based seizure detection, Biomed. Signal Process. Control 103 (2025) 107484

2025
[5]

Albert Gu, Tri Dao, Mamba: linear-time sequence modeling with selective State spaces, arXiv (2023)

2023
[6]

Rui Xu, et al., Visual Mamba: a survey and new outlooks, arXiv (2024)

2024
[7]

N. Zhou, et al., Concordance study between IBM Watson for oncology and clinical practice for patients with cancer in China, Oncologist 24 (6) (2019) 812 – 819, https://doi.org/10.1634/theoncologist.2018-0255

work page doi:10.1634/theoncologist.2018-0255 2019
[8]

Pinto e Vairo, et al., Implementation of genomic medicine for rare disease in a Tertiary Healthcare system: mayo Clinic Program for rare and undiagnosed diseases (PRaUD), J

F. Pinto e Vairo, et al., Implementation of genomic medicine for rare disease in a Tertiary Healthcare system: mayo Clinic Program for rare and undiagnosed diseases (PRaUD), J. Transl. Med. 21 (1) (2023) 410, https://doi.org/10.1186/ s12967-023-04183-7

2023
[9]

Picone, The Temple University Hospital EEG data corpus, Front

Iyad Obeid, Joseph W. Picone, The Temple University Hospital EEG data corpus, Front. Neurosci. 10 (2016) 196

2016
[10]

Gemein, et al., Machine-learning-based diagnostics of EEG pathology, NeuroImage 220 (2020) 117021

Lukas A.W. Gemein, et al., Machine-learning-based diagnostics of EEG pathology, NeuroImage 220 (2020) 117021

2020
[11]

Fusion 58 (2020) 82 – 115

Alejandro Barredo Arrieta, et al., Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI, Inf. Fusion 58 (2020) 82 – 115

2020
[12]

Rishi Bommasani, Percy Liang, et al., The Foundation Model Transparency Index, Stanford HAI, 2023

2023
[13]

Ahmed Abdelhameed, Magdy Bayoumi, A deep learning approach for automatic seizure detection in children with epilepsy, Front. Comput. Neurosci. 15 (2021) 650050

2021
[14]

Radu Agwad, et al., Detection of epileptic seizure in EEG signals using machine learning and Deep learning techniques, J. Eng. Appl. Sci. 70 (1) (2023) 131, https://doi.org/10.1186/s44147-023-00353-y

work page doi:10.1186/s44147-023-00353-y 2023
[15]

Ralph G. Andrzejak, et al., Indications of nonlinear deterministic and finite dimensional structures in time series of brain electrical activity: dependence on recording region and brain State, Phys. Rev. E 64 (6) (2001) 061907

2001
[16]

Dissertation

Ali Hossam Shoeb, Application of Machine Learning to Epileptic Seizure Onset Detection and Treatment, Massachusetts Institute of Technology, 2009. Dissertation

2009
[17]

Methods Programs Biomed

Matthias Ihle, et al., EPILEPSIAE — A European epilepsy database, Comput. Methods Programs Biomed. 106 (3) (2012) 127 – 138

2012
[18]

Selvaraj, et al., EEG database of seizure disorders for experts and application developers, Clin

Thomas G. Selvaraj, et al., EEG database of seizure disorders for experts and application developers, Clin. EEG Neurosci. 45 (4) (2014) 304 – 309

2014
[19]

13 (4-5) (2000) 411 – 430

Aapo Hyv ¨arinen, Erkki Oja, Independent component analysis: algorithms and applications, Neural Netw. 13 (4-5) (2000) 411 – 430

2000
[20]

Koles, The quantitative extraction and topographic mapping of the abnormal components in the clinical EEG, Electroencephalogr

Zoltan J. Koles, The quantitative extraction and topographic mapping of the abnormal components in the clinical EEG, Electroencephalogr. Clin. Neurophysiol. 79 (6) (1991) 440 – 447, https://doi.org/10.1016/0013-4694(91)90163-x

work page doi:10.1016/0013-4694(91)90163-x 1991
[21]

Y. Zhao, X. Wang, et al., Residual and bidirectional LSTM for epileptic seizure detection, Front. Comput. Neurosci. 18 (2024), https://doi.org/10.3389/ fncom.2024.1415967

arXiv 2024
[22]

Li, et al., Epileptic seizure prediction via multidimensional transformer and recurrent neural network fusion, J

X. Li, et al., Epileptic seizure prediction via multidimensional transformer and recurrent neural network fusion, J. Transl. Med. 22 (2024), https://doi.org/ 10.1186/s12967-024-05678-7

work page doi:10.1186/s12967-024-05678-7 2024
[23]

Sun, et al., Epileptic seizure detection from EEG signals based on 1D CNN-LSTM deep learning model using discrete wavelet transform, Sci

H. Sun, et al., Epileptic seizure detection from EEG signals based on 1D CNN-LSTM deep learning model using discrete wavelet transform, Sci. Rep. 15 (2025), https:// doi.org/10.1038/s41598-025-18479-9

work page doi:10.1038/s41598-025-18479-9 2025
[24]

W. Hu, J. Wang, F. Li, et al., A modified transformer network for seizure detection using EEG signals, Int. J. Neural Syst. (2025), https://doi.org/10.1142/ S0129065725500030

2025
[25]

Jeong, K

H. Jeong, K. Lee, S. Kim, H.-C. Kang, D. Yang, Deep learning-based real-time seizure detection and multi-seizure classification on pediatric EEG, Front. Neurol. 17 (2026) 1726258, https://doi.org/10.3389/fneur.2026.1726258 . M. Chen et al. Biomedical Engineering Advances 11 (2026) 100218 18

work page doi:10.3389/fneur.2026.1726258 2026

[1] [1]

Lipton, et al., Learning to diagnose with LSTM recurrent neural networks, arXiv (2015)

Zachary C. Lipton, et al., Learning to diagnose with LSTM recurrent neural networks, arXiv (2015)

2015

[2] [2]

Lawhern, et al., EEGNet: a compact convolutional neural network for EEG-based brain – computer interfaces, J

Vernon J. Lawhern, et al., EEGNet: a compact convolutional neural network for EEG-based brain – computer interfaces, J. Neural Eng. 15 (5) (2018) 056013

2018

[3] [3]

Pierre Vanabelle, et al., Epileptic seizure detection using EEG signals and extreme gradient boosting, J. Biomed. Res. 34 (3) (2020) 228 – 239

2020

[4] [4]

Signal Process

Sheng Wong, et al., Channel-annotated deep learning for enhanced interpretability in EEG-based seizure detection, Biomed. Signal Process. Control 103 (2025) 107484

2025

[5] [5]

Albert Gu, Tri Dao, Mamba: linear-time sequence modeling with selective State spaces, arXiv (2023)

2023

[6] [6]

Rui Xu, et al., Visual Mamba: a survey and new outlooks, arXiv (2024)

2024

[7] [7]

N. Zhou, et al., Concordance study between IBM Watson for oncology and clinical practice for patients with cancer in China, Oncologist 24 (6) (2019) 812 – 819, https://doi.org/10.1634/theoncologist.2018-0255

work page doi:10.1634/theoncologist.2018-0255 2019

[8] [8]

Pinto e Vairo, et al., Implementation of genomic medicine for rare disease in a Tertiary Healthcare system: mayo Clinic Program for rare and undiagnosed diseases (PRaUD), J

F. Pinto e Vairo, et al., Implementation of genomic medicine for rare disease in a Tertiary Healthcare system: mayo Clinic Program for rare and undiagnosed diseases (PRaUD), J. Transl. Med. 21 (1) (2023) 410, https://doi.org/10.1186/ s12967-023-04183-7

2023

[9] [9]

Picone, The Temple University Hospital EEG data corpus, Front

Iyad Obeid, Joseph W. Picone, The Temple University Hospital EEG data corpus, Front. Neurosci. 10 (2016) 196

2016

[10] [10]

Gemein, et al., Machine-learning-based diagnostics of EEG pathology, NeuroImage 220 (2020) 117021

Lukas A.W. Gemein, et al., Machine-learning-based diagnostics of EEG pathology, NeuroImage 220 (2020) 117021

2020

[11] [11]

Fusion 58 (2020) 82 – 115

Alejandro Barredo Arrieta, et al., Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI, Inf. Fusion 58 (2020) 82 – 115

2020

[12] [12]

Rishi Bommasani, Percy Liang, et al., The Foundation Model Transparency Index, Stanford HAI, 2023

2023

[13] [13]

Ahmed Abdelhameed, Magdy Bayoumi, A deep learning approach for automatic seizure detection in children with epilepsy, Front. Comput. Neurosci. 15 (2021) 650050

2021

[14] [14]

Radu Agwad, et al., Detection of epileptic seizure in EEG signals using machine learning and Deep learning techniques, J. Eng. Appl. Sci. 70 (1) (2023) 131, https://doi.org/10.1186/s44147-023-00353-y

work page doi:10.1186/s44147-023-00353-y 2023

[15] [15]

Ralph G. Andrzejak, et al., Indications of nonlinear deterministic and finite dimensional structures in time series of brain electrical activity: dependence on recording region and brain State, Phys. Rev. E 64 (6) (2001) 061907

2001

[16] [16]

Dissertation

Ali Hossam Shoeb, Application of Machine Learning to Epileptic Seizure Onset Detection and Treatment, Massachusetts Institute of Technology, 2009. Dissertation

2009

[17] [17]

Methods Programs Biomed

Matthias Ihle, et al., EPILEPSIAE — A European epilepsy database, Comput. Methods Programs Biomed. 106 (3) (2012) 127 – 138

2012

[18] [18]

Selvaraj, et al., EEG database of seizure disorders for experts and application developers, Clin

Thomas G. Selvaraj, et al., EEG database of seizure disorders for experts and application developers, Clin. EEG Neurosci. 45 (4) (2014) 304 – 309

2014

[19] [19]

13 (4-5) (2000) 411 – 430

Aapo Hyv ¨arinen, Erkki Oja, Independent component analysis: algorithms and applications, Neural Netw. 13 (4-5) (2000) 411 – 430

2000

[20] [20]

Koles, The quantitative extraction and topographic mapping of the abnormal components in the clinical EEG, Electroencephalogr

Zoltan J. Koles, The quantitative extraction and topographic mapping of the abnormal components in the clinical EEG, Electroencephalogr. Clin. Neurophysiol. 79 (6) (1991) 440 – 447, https://doi.org/10.1016/0013-4694(91)90163-x

work page doi:10.1016/0013-4694(91)90163-x 1991

[21] [21]

Y. Zhao, X. Wang, et al., Residual and bidirectional LSTM for epileptic seizure detection, Front. Comput. Neurosci. 18 (2024), https://doi.org/10.3389/ fncom.2024.1415967

arXiv 2024

[22] [22]

Li, et al., Epileptic seizure prediction via multidimensional transformer and recurrent neural network fusion, J

X. Li, et al., Epileptic seizure prediction via multidimensional transformer and recurrent neural network fusion, J. Transl. Med. 22 (2024), https://doi.org/ 10.1186/s12967-024-05678-7

work page doi:10.1186/s12967-024-05678-7 2024

[23] [23]

Sun, et al., Epileptic seizure detection from EEG signals based on 1D CNN-LSTM deep learning model using discrete wavelet transform, Sci

H. Sun, et al., Epileptic seizure detection from EEG signals based on 1D CNN-LSTM deep learning model using discrete wavelet transform, Sci. Rep. 15 (2025), https:// doi.org/10.1038/s41598-025-18479-9

work page doi:10.1038/s41598-025-18479-9 2025

[24] [24]

W. Hu, J. Wang, F. Li, et al., A modified transformer network for seizure detection using EEG signals, Int. J. Neural Syst. (2025), https://doi.org/10.1142/ S0129065725500030

2025

[25] [25]

Jeong, K

H. Jeong, K. Lee, S. Kim, H.-C. Kang, D. Yang, Deep learning-based real-time seizure detection and multi-seizure classification on pediatric EEG, Front. Neurol. 17 (2026) 1726258, https://doi.org/10.3389/fneur.2026.1726258 . M. Chen et al. Biomedical Engineering Advances 11 (2026) 100218 18

work page doi:10.3389/fneur.2026.1726258 2026