A neuroanalytic multi-scale, multi-granularity decoupled distillation fuzzy classifier for interpretable epileptic EEG pattern detection

Introduction Electroencephalogram (EEG)-based recognition of epileptic seizures is a representative neuroanalytic pattern-detection problem at the intersection of artificial intelligence and computational neuroscience. Deep networks learn strong representations for this task but act as opaque function approximators, whereas Takagi-Sugeno-Kang (TSK) fuzzy classifiers express their reasoning as human-readable IF-THEN rules but have historically lagged behind deep models in accuracy. This study aimed to combine predictive strength with rule-level transparency. Methods A Multi-scale Multi-granularity Decoupled Distillation TSK Fuzzy Classifier (MMDD-TSK-FC) is proposed. One-dimensional convolutional neural networks (1D-CNNs) with different kernel widths are first trained as multi-scale teachers, so that both slow, trend-level and brief, transient-level EEG dynamics are captured. The temperature-softened outputs of each teacher are then transferred, through its own (decoupled) Kullback-Leibler distillation channel, to a TSK fuzzy student whose rule count is matched to the teacher's kernel width: wide kernels are paired with compact rule sets and narrow kernels with richer rule sets. A joint objective also penalises disagreement between each student and the ground-truth labels, and the predictions of the resulting multi-granularity students are fused by majority voting. The teachers are discarded after training, and the deployed students reason directly over the original input features. The method was evaluated on six binary tasks built from the Bonn and New Delhi Hauz Khas epileptic EEG corpora. Results Averaged over the six tasks, MMDD-TSK-FC improved accuracy by 5.18 percentage points over the single-granularity FCM-TSK-FC classifier (per-task range 3.05–8.09) and by 2.52 points over the distillation-based STSK-LLM-KD (range 0.66–4.51). Against the strongest baseline, ED-TSK-FC, the mean margin was 0.77 points, ranging from −0.31 on Task 2 to +2.49 on Task 6. Discussion Decoupled multi-scale distillation can bring a rule-based classifier to the accuracy level of deep models. Because each final decision can be traced to a small number of explicit rules defined on the input features, the decisions remain inspectable.

Authors

Institutions

Publication Details

Journal
Frontiers in Medicine
Published
2026-11-16
DOI
https://doi.org/10.3389/fmed.2026.1982807
Primary Topic
EEG and Brain-Computer Interfaces
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A neuroanalytic multi-scale, multi-granularity decoupled distillation fuzzy classifier for interpretable epileptic EEG pattern detection

Sami Ahmed Haider, Sanjay T. Singh, Fayez Alfayez, Haewon Byeon et al.
1 links
Frontiers in Medicine
EEG and Brain-Computer Interfaces
article

A neuroanalytic multi-scale, multi-granularity decoupled distillation fuzzy classifier for interpretable epileptic EEG pattern detection

Sami Ahmed Haider, Sanjay T. Singh, Fayez Alfayez, Haewon Byeon, Abdulrahman Mohammed Alamoudi, Masood Ur Rehman, Mukesh Soni
article en

Abstract

Introduction Electroencephalogram (EEG)-based recognition of epileptic seizures is a representative neuroanalytic pattern-detection problem at the intersection of artificial intelligence and computational neuroscience. Deep networks learn strong representations for this task but act as opaque function approximators, whereas Takagi-Sugeno-Kang (TSK) fuzzy classifiers express their reasoning as human-readable IF-THEN rules but have historically lagged behind deep models in accuracy. This study aimed to combine predictive strength with rule-level transparency. Methods A Multi-scale Multi-granularity Decoupled Distillation TSK Fuzzy Classifier (MMDD-TSK-FC) is proposed. One-dimensional convolutional neural networks (1D-CNNs) with different kernel widths are first trained as multi-scale teachers, so that both slow, trend-level and brief, transient-level EEG dynamics are captured. The temperature-softened outputs of each teacher are then transferred, through its own (decoupled) Kullback-Leibler distillation channel, to a TSK fuzzy student whose rule count is matched to the teacher's kernel width: wide kernels are paired with compact rule sets and narrow kernels with richer rule sets. A joint objective also penalises disagreement between each student and the ground-truth labels, and the predictions of the resulting multi-granularity students are fused by majority voting. The teachers are discarded after training, and the deployed students reason directly over the original input features. The method was evaluated on six binary tasks built from the Bonn and New Delhi Hauz Khas epileptic EEG corpora. Results Averaged over the six tasks, MMDD-TSK-FC improved accuracy by 5.18 percentage points over the single-granularity FCM-TSK-FC classifier (per-task range 3.05–8.09) and by 2.52 points over the distillation-based STSK-LLM-KD (range 0.66–4.51). Against the strongest baseline, ED-TSK-FC, the mean margin was 0.77 points, ranging from −0.31 on Task 2 to +2.49 on Task 6. Discussion Decoupled multi-scale distillation can bring a rule-based classifier to the accuracy level of deep models. Because each final decision can be traced to a small number of explicit rules defined on the input features, the decisions remain inspectable.

Frontiers in MedicineVol. 13
Lovely Professional University (IN), Galgotias University (IN), Majmaah University (SA), Heriot-Watt University (GB), Korea University of Technology and Education (KR), University of Jeddah (SA), University of Glasgow (GB)
EEG and Brain-Computer Interfaces
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon