Quantum inspired feature engineering for explainable EEG signal classification

Abstract

In this research, our main objective is to extract more informative features by deploying a simple and effective framework. One of the cheapest data-gathering methods from the brain is electroencephalography signal collection. The main aim of this approach is to obtain maximum information from electroencephalography signals. Therefore, we have presented a quantum-inspired feature extraction function and evaluated its classification ability. In this approach, we have employed six electroencephalography signal datasets as a testbed, aiming to depict the general classification capability of the introduced electroencephalography signal classification model. Firstly, a quantum entangled particle pattern has been proposed, which is a transformer-based feature extraction function. To investigate the classification performance of the introduced quantum entangled particle pattern, a new-generation explainable feature engineering framework has been introduced. The quantum entangled particle pattern-centric explainable feature engineering model extracts features using the quantum entangled particle pattern feature extraction function. By employing cumulative weighted iterative neighborhood component analysis, the most distinctive features extracted by quantum entangled particle pattern have been selected. The algorithm-centric k-nearest neighbor classifier has been applied to obtain classification results. Directed lobish has been utilized to generate interpretable results. To obtain both classification and interpretable results, the selected features and their identities have been used as inputs for centric k-nearest neighbors and directed lobish consecutively. The introduced quantum entangled particle pattern-related explainable feature engineering approach attained over 90% classification accuracy on the six electroencephalography signal datasets with 10-fold cross-validation. Additionally, this model generates a connectome diagram to provide interpretable results for each dataset.

Introduction

Electroencephalography (EEG) is a technique that captures electrical signals from the brain to record neuronal connectivity that has been widely used in neuroscience and clinical neurology^1,2. Electrical signals arising from neuronal connectivity can reveal information about cognitive processes as well as neurological events including epilepsy, psychosis, and amyotrophic lateral sclerosis (ALS)³. EEG signals, however, are complex by nature in terms of being non-stationary, high-dimensional data, requiring advanced computational techniques for their accurate analysis⁴. The extraction of features from the individual signals is a crucial step in EEG-based classification studies because the EEG signal is flawed due to noise contamination and inter-individual variability⁵.

Brain-computer interfacing turns directly to the signal for neuron integration, and it requires data to systematically process raw signals for effective machine learning⁶, hence why feature extraction takes charge⁷. Recent studies have demonstrated the effectiveness of deep convolutional neural networks for efficient motor imagery EEG-based BCI systems⁸. There are several methods applied for feature extraction; those methods have a classical or deep learning (DL)-based approach⁹. Deep learning architectures such as attention-based nested networks have shown remarkable success in medical image analysis¹⁰. Although DL models perform well at classification accuracy, they require large amounts of training data^11,12. Moreover, most current frameworks for EEG classification concentrate solely on performance measures, leaving little to no biologically meaningful information about the relevant neural mechanisms that could aid clinical and research applications¹³.

Quantum-inspired computational techniques have gained attention as a feasible solution for biomedical signal processing^14–16. These techniques are borrowed from quantum mechanics to improve the processes of feature extraction and classification¹⁷. Quantum-inspired models stand out in EEG signal analysis as they apply concepts like quantum entanglement and probabilistic transformations to interpret EEG data, potentially revealing novel patterns¹⁸. These methods can improve feature representation with a small computational cost compared with traditional models and thus can be feasible for online (real-time) applications of EEG¹⁹.

Existing models in EEG classification also suffer from lack of interpretability²⁰. Leveraging explainable artificial intelligence (XAI) to analyze EEG is a step toward bridging this gap with transparent neurologically relevant insight²¹. The XAI technique used helps identify the most relevant features contributing to the classification outputs, thus helping researchers and clinicians validate the results against their knowledge based on well-established neurophysiological understandings²². This improves trustworthiness and usability, which is especially relevant in the medical setting as decisions are based on biological processes that can be readily explained²³.

This study explores novel EEG feature engineering methods which lead to higher accuracies, with a focus on computational efficiency and interpretability of results. This makes the advanced techniques presented here more applicable in any context where the analysis of bio-signals can be improved like EEG-based diagnostics and neuroinformatics studies, through the use of quantum-inspired methods and explainable feature engineering. The results obtained will influence multi-modal applications in the use of light-weight, fast and interpretable approaches that could be implemented on across a range of EEG datasets for an improved understanding of brain activity in health and disease.

Literature review

Some recent studies in the literature on ALS, stress, violence, psychosis, epilepsy, and artifact detection are listed as follows. Liu et al. ²⁴ suggested an EEG based motor imagery classification approach for ALS patients by performing fractal dimension analysis and using Fisher’s criterion based channel selection. They analyzed EEG data of five end-stage ALS patients engaged in three imagery tasks. Compared to classical sensorimotor rhythm features, their method resulted in an accuracy of up to 95.25% from 30 channels and 91.00% from a single optimal channel. In their model, they didn’t use multiple datasets and they tested their models on the single dataset. Sengur et al. ²⁵ used reinforcement sample learning for electromyogram (EMG) signals of ALS and developed a method for EMG-assisted classification of ALS. The dataset selected in the study consisted of ALS (n = 89) and normal (n = 133) EMG signals with the sampling frequency equal to 24 kHz. Before classification with CNN, time-frequency representations were used. They achieved an accuracy of 96.80% using this method. The utilized dataset is relatively small. Ramakrishnan et al. ²⁶ developed a DL based brain-computer interface system for ALS patients. Electrooculography (EOG) signals were recorded from eight subjects (four trained and four untrained) using a bioamplifier with five electrodes. A recent work focused on classifying eye movement tasks for wheelchair navigation using CNN and achieved 93.51% and 86.88% accuracy for trained and untrained users respectively. Subject S4 achieved the best performance with 97.50% accuracy. However, leave-one-subject-out cross-validation (LOSO CV) was not utilized. Latifoglu²⁷ suggested an ALS detection approach based on EEG-derived event-related potentials (ERPs). Moreover, empirical mode decomposition (EMD) and variational mode decomposition (VMD) were utilized to extract subband features which were classified using a 1D-CNN. Their results showed that VMD outperformed EMD and achieved the best accuracy (92.95%) with 10-fold cross-validation. Their method is not a new-generation method and they only investigated the classification performances of the VMD and EMD methods on the ALS dataset. Using resting-state magnetoencephalography data, Samanta et al. ²⁸ presented 3D deep CNN for ALS detection. Data of 26 ALS patients and 26 healthy controls recorded in a 306-channel Elekta Neuromag scanner were used. Morlet wavelet transform was proved to generate time-frequency representations that were classified with MEGNet3D. Their model was over 75% correct across the various classification conditions. In this respect, the relatively low classification accuracy was computed. Makam et al. ²⁹ introduced a novel framework for ALS detection from EMG signals which utilized automatic singular spectrum analysis (Auto-SSA) coupled with a quantum CNN. Their dataset comprised of 955 EMG signals from ALS and healthy subjects at 23.435 kHz. Using Auto-SSA, these signals were decomposed into reconstructed components, followed by the selection of 64 optimal features through particle swarm optimization. They attained testing accuracy of 98.50%. However, they didn’t present any interpretable results.

Saba-Sadiya et al. ³⁰ proposed a method for unsupervised EEG artifact detection and correction. EEG data from two passive viewing tasks were used in their study, recorded with a 32-electrode actiCHamp cap with a sampling rate of 1,000 Hz. ~ Approximately 10,000 EEG trials across all subjects were analyzed with 58 handcrafted features per trial explored for outlier detection. A combination of unsupervised algorithms detected 9.99% more artifacts than baseline methods were able to do, while a deep encoder-decoder architecture was observed to achieve 10% better classification performance once artifacts were removed. In this model, no XAI results are provided. Abdi-Sargezeh³¹ introduced an EEG artifact removal method by the common component rejection (CCR) and automatic wavelet CCR (AWCCR). They achieved an accuracy of 72.60%. Their classification results are relatively low. Also, there is no interpretable results.

Mukherjee and Roy³² presented a methodology to segment various stress levels through EMG and Heart Rate signals. EEG data from 34 healthy subjects were collected from the RMS Maximus 32 EEG system, and heart rate data from RMS Relax 701 were collected, while the subjects were working on solving mathematical problems of increasing complexity. They obtained an average accuracy of 99.72% but they didn’t present any XAI results. Kim et al. ³³ proposed a stress detection framework using single-channel EEG and galvanic skin response signals recorded in a virtual reality interview paradigm. Their study collected biosignals from 30 participants exposed to simulated stress-inducing interviews, analyzing stress responses through five CNN architectures and a Vision Transformer model. They achieved an AUROC of 0.954. However, the innovation of their model is limited. Afify et al. ³⁴ applied a CNN network for developing an EEG-based stress detection model. They employed SAM 40 data which comprises EEG recordings of thirty-four individuals over four cognitive tasks. A 32-channel bipolar EEG system was used to record two types of tasks and a total of 480 signals from 120 trials per task. The proposed model resulted in 99.25% accuracy. In this research, no interpretable results were reported and no LOSO CV-based classification performance results were presented. Jagtap et al. ³⁵ developed an EEG-based stress detection approach integrating multiple signal processing and DL techniques. They utilized the SEED, DEAP, and Mental Stress Detection datasets, containing EEG recordings from various cognitive and emotional tasks. They achieved 98% accuracy. However, they didn’t present any explainable results or subject-wise evaluations. Daadaa et al. ³⁶ suggested a framework for detecting stress and anxiety using EEG signals. They utilized the DEEP, SEED, and DASPS datasets, containing EEG recordings from various cognitive and emotional tasks. Their model achieved a 98.6% accuracy. There is no subject-wise cross-validation (CV) results in their model and XAI weren’t used in their model to showcase interpretable results generation ability.

For violence detection in smart surveillance systems, Halder and Chatterjee³⁷ presented a CNN-BiLSTM model. Their study made use of three benchmark datasets consisting of violent and non-violent scenes, including Hockey Fights (1,000 clips), Movies (200 clips) and Violent Flows (246 clips). Their model used convolutional layers and bidirectional LSTMs to extract spatial and temporal features, obtaining 99.27% accuracy for Hockey Fights, 100% for Movies, and 98.64% for Violent Flow. Here, deep learning models were utilized to obtain high classification performance but the time complexity of the deep learning models are very high. In this aspect, the training of this model is not suitable for simply configured (or low-spec) computers. Abundez et al. ³⁸ proposed a method for physical violence detection in video frames. Their approach obtained an AUC of up to 0.989 but they didn’t present any XAI results. Asad et al. ⁹ developed a multi-frame feature-fusion-based approach to detect violence in video surveillance. Their method achieved 98.80% accuracy on Hockey Fights, 99.10% on Movies, 97.10% on Violent Flow, and 95.90% on BEHAVE datasets. Any XAI results didn’t present in this research. Haiura and Iftene⁴⁰ presented a 3D CNN for real-time violence detection in surveillance scenarios. They calculated an accuracy of 91.58%. No XAI results was presented and the complexity of this model was relatively high since they used 3D CNN.

Zulfikar and Mehmet⁴¹ presented an EEG-based model. The study utilized two EEG datasets, one with 19-channel recordings from 28 participants (14 SZ, 14 healthy controls- dataset I) and another with 16-channel recordings from 84 participants (45 SZ, 39 healthy controls- dataset II). They reached 98.2% for Dataset I and 96.02% for Dataset II. They didn’t present any LOSO CV results and there are no interpretable results. Li et al. ⁴² developed a method for first-episode psychosis (FEP), bipolar disorder (BD), and healthy controls. Their study included 83 healthy controls, 40 BD patients and 89 FEP patients. They calculated a 99.72% accuracy. They only focused classification performances and there are no interpretable results. Shubhangi et al. ⁴³ proposed an EEG-based psychosis susceptibility syndrome detection method using local binary pattern encoding and CNN. Their study utilized EEG recordings from 14 psychosis susceptibility syndrome patients and 14 healthy controls. They achieved a 97.70% accuracy. Their utilized dataset is a toy dataset. Thus, these results cannot be generalized. Elujide et al. ⁴⁴ developed a multi-label classification approach for psychotic disorder detection Their study utilized a psychotic disorder diseases dataset containing 500 patient records with diagnoses of bipolar disorder, schizophrenia, vascular dementia, insomnia, and ADHD. They achieved a 75.17% accuracy (relatively low classification accuracy).

Bhadra et al. ⁴⁵ proposed a model for epileptic seizure detection using EEG signals. Their study utilized two public datasets: UCI Epilepsy and Mendeley datasets. They calculated 99.01% accuracy on the UCI dataset and 97.50% on the Mendeley dataset. These datasets are small datasets and there are no XAI and subject-wise results. Rivera et al. ⁴⁶ presented an approach for seizure type classification. Their study utilized the temple university hospital seizure dataset (consist of 239 patients). They attained F1-score of 61.10%. They attained relatively low classification performances. Holguin-Garcia et al. ⁴⁷ proposed a comparative study on epileptic seizure classification. Their study utilized the Epileptic Seizure Recognition database. Their study reached 99.92% accuracy but they didn’t present XAI results in their research. Using EEG signals, Zhang et al. ⁴⁸ suggested a model for epileptic seizure detection. Their study utilized the CHB-MIT dataset. Their model achieved 99.35% accuracy for seizure detection. They used k-fold CV and they didn’t report any LOSO CV results. Gill et al. ⁴⁹ developed an approach for classifying generalized and focal epileptic seizures. Their approach employed the temple university hospital seizure corpus dataset. Their approach achieved a 92.10% weighted accuracy. There is no interpretable results in their results.

Despite significant progress in EEG signal classification, three main gaps remain in the literature. First, we are living in the rise of the machine learning age⁵⁰, and scientific production about machine learning is very high^51,52. Most researchers have utilized computationally expensive deep learning models^53–55 to achieve high classification performance. Due to expensive processor and energy requirements, more lightweight and highly accurate models should be presented⁵⁶. Second, most EEG signal classification frameworks^20,57,58 have validated their performance on a single dataset, causing limitations in generalization. Third, many researchers have focused solely on yielding high classification performance^59–61, resulting in a lack of explainable artificial intelligence (XAI) approaches in the literature.

To address these gaps, we propose a novel Quantum Entangled Particles Pattern (QEPP)-centric explainable feature engineering (XFE) framework. Nowadays, we are living in the AI age, and AI applications, especially large language models (LLMs), have been used everywhere^62,63. Moreover, various researchers have worked on developing the optimum artificial general intelligence (AGI)^64,65. Humanity has achieved this level due to the high performance of deep learning architectures. However, deep learning models require significant energy and data resources, making them expensive⁵⁶. On the other hand, feature engineering models are lightweight architectures^66,67, but they are not general models, and their classification performance is relatively lower than that of deep learning models. The major motivation is to introduce a new-generation general EEG signal classification model that achieves high classification performance with interpretable results. We are inspired by the dynamic structure of quantum mechanics and transformers⁶⁸. The QEPP is a quantum-inspired feature extraction method that combines a newly developed QEP transformer with a Sequential and Combinational Transition Table (SCTT) feature extractor. In the QEP transformer, two vectors are used as entangled particles, and the difference between these vectors is computed as the energy vector difference. By applying the proposed transformer to these vectors, three transformed signals are generated, from which distinctive features are extracted using the SCTT function.

Our framework employs two self-organized methods to ensure optimal classification performance: CWINCA (Cumulative Weighted Iterative Neighborhood Component Analysis) for feature selection and tkNN (tuned k-Nearest Neighbors) for classification. Additionally, the Directed Lobish (DLob) XAI method has been integrated to generate interpretable results. For each dataset, a DLob string and a cortical connectome diagram (CCD) are constructed, providing explainable insights into the neurological significance of the extracted features.

The main innovations and contributions of this work are as follows:

• We developed the QEP transformer, which is the first feature extraction-dedicated quantum-based transformer to our knowledge, along with the SCTT feature extraction function.

• We propose a new XFE framework that validates both classification performance and interpretability.

• The proposed QEPP-centric XFE model has been tested on six distinct EEG datasets (ALS, artifact, stress, violence, psychosis, and epilepsy), making it a pioneering model in EEG signal classification for its comprehensive validation approach.

• To present robust classification results, 10-fold, leave-one-subject-out (LOSO), and leave-one-record-out (LORO) cross-validations have been utilized.

• The introduced framework achieved over 90% classification accuracy on all six datasets, demonstrating that it is a generally high-accuracy model comparable to deep learning models but with linear time complexity.

• By deploying DLob, explainable findings including cortical connectome diagrams have been computed, contributing to neuroscience with AI-based interpretable results.

Materials

In this research, we have used six EEG signal classification datasets, each with different characteristics. The datasets used are: (i) EEG Amyotrophic Lateral Sclerosis (ALS) detection, (ii) EEG Artifact classification, (iii) EEG Stress detection, (iv) EEG Violence detection, (v) EEG Psychosis detection, and (vi) EEG Epilepsy detection. These datasets were collected using EEG signal acquisition devices with 14, 32 or 35 channels. The details of these datasets are provided below.

EEG ALS dataset

In this dataset, an EEG collection device with 32 channels was used^69,70. The dataset was collected from 170 control participants and 6 ALS participants, as ALS is a rare disorder. To balance the dataset, we selected 2,631 EEG segments, each 10 s long, for each class. There are two classes in this dataset: (1) ALS and (2) Control.

Given the extremely limited number of ALS patients (6 cases) compared to 170 healthy controls, the original dataset exhibits a severe class imbalance. To avoid strong class bias during model training and evaluation, we adopted a balancing strategy based on controlled undersampling of the majority class. Specifically, from the rich control recordings, we randomly selected EEG segments whose total duration matches the overall recording duration of the ALS patients, resulting in 2,631 segments per class. Although this procedure may reduce the diversity of control data, it is a common and necessary practice in small-sample rare disease studies to ensure that the model focuses on learning disease-discriminative patterns rather than being dominated by the majority class. The high geometric mean (G-mean) values observed in subsequent experiments indicate that the performance on this balanced subset is robust and not driven by class bias.

EEG artifact dataset

This dataset contains one clean class and seven artifact classes, making a total of eight classes⁷¹. The researchers collected this dataset using the Emotiv Epoch X brain cap, which has 14 channels. The sampling frequency of the Emotiv Epoch X brain cap used is 128 Hz. The distribution of this dataset is shown in Table 1. In this dataset, there are 2,498 EEG segments, each with a length of 5 s. The distribution of the used EEG artifact classification dataset is also tabulated in Table 1.

Table 1.

The distribution of the EEG artifact classification dataset.

No.	Class	Number of EEGs
0	No Artifact	1249
1	Limb Tremor	181
2	Noise	179
3	Body Movement	178
4	Eye Blinking	178
5	Swallowing	180
6	Vertical Eye Movement	181
7	Speaking	172
Total		2498

As seen in Table 1, the artifact dataset is unbalanced across the 7 artifact types, but the total number of clean vs. artifact samples is balanced (1,249 vs. 1,249).

EEG stress dataset

The researchers curated this dataset from 310 participants using the Emotiv Epoch X brain cap, which has 14 channels⁷². The EEG stress dataset contains 3,667 EEG segments, each 15 s long. The distribution of this dataset is as follows: (1) 1,785 stress signals and (2) 1,882 control signals.

EEG violence dataset

The EEG violence detection dataset is a binary classification dataset containing two classes, and it is a 14-channel dataset⁷³. The length of each EEG segment is 15 s. In this dataset, the two classes are (0) Control and (1) Violence. 442 of the EEG signals belong to the control class, while the remaining 286 EEG signals belong to the violence class. Thus, there are a total of 778 EEG signals in this dataset.

EEG psychosis dataset

The EEG psychosis detection dataset was collected using the Emotiv Flex brain cap, which has 32 channels, and the sampling frequency of this device is 256 Hz ⁷⁴. In this dataset, EEG signals were collected from 64 participants. The collected EEG signals were divided into 15-second segments. There are 4,098 EEG segments in this dataset, classified into two categories: (0) Control and (1) Psychosis. 1,350 of these 4,098 EEG segments belong to psychosis participants, while the remaining 2,748 are labeled as control.

EEG Epilepsy – Turkish Epilepsy – dataset

The largest dataset used in this research is the Turkish Epilepsy Dataset⁷⁵. This dataset was collected from 121 participants, 50 of whom have epilepsy, while the remaining 71 participants have no findings. Therefore, the EEG signals of these 71 participants were labeled as control. This dataset was collected using a brain cap with 35 channels and the sampling frequency of the used brain cap is 500 Hz. The length of each EEG segment is 15 s, and the dataset contains (1) 4,465 epileptic EEG segments and (2) 5,891 control EEG segments. Totally, there are 10,356 EEG signals in this dataset.

The presented quantum entangled particle pattern

The major innovation of this research is the presented feature extraction method, termed QEPP. The concept of entanglement in quantum mechanics describes the inseparable, strong correlation between multiple components in a system, whose states cannot be described individually. Inspired by this, we hypothesize that similar high-order, implicit coupling exists between multi-channel EEG signals. Although EEG signals are classical signals, we can borrow the mathematical form describing entanglement—focusing on a pair of signals (analogous to “entangled particle pairs”) and the joint state formed by their differences—to design a novel feature extraction framework.

The connection between quantum mechanics and EEG signal processing can be understood from both mathematical and conceptual perspectives. In quantum mechanics, entangled particles exhibit non-local correlations where the state of one particle is inherently linked to the state of another, regardless of the distance between them. Similarly, EEG signals recorded from multiple channels demonstrate synchronous activity patterns, such as phase-locking and coherence, where neural oscillations in distant brain regions become temporally correlated⁷⁶. These multi-channel correlations in EEG can be mathematically analogous to the correlation structure observed in entangled quantum systems.

Furthermore, the brain functions as a complex network system where long-range correlations exist between different cortical regions. These correlations share mathematical similarities with quantum entanglement in terms of information structure and joint probability distributions. From a modeling perspective, EEG signals contain high-order correlation patterns that cannot be effectively captured by analyzing individual channels separately. Just as entangled quantum states require joint transformations for proper representation, these complex inter-channel dependencies in EEG may benefit from similar joint transformation approaches. The proposed QEPP is motivated by this analogy: by treating paired channel vectors as “entangled particles” and computing their joint transformations, we aim to explicitly model and extract latent, strongly correlated patterns of information between channels—patterns that might be missed when analyzing individual channels in isolation. Therefore, it should be noted that the “quantum entangled particle pattern” proposed in this paper does not realize real quantum computing, but rather constructs a classical signal transformation method inspired by the mathematical philosophy of quantum entanglement. By creating signal pairs, calculating their “energy difference” vectors, and performing joint sorting transformations, we aim to simulate the observation process of “entangled states,” thereby extracting new feature representations from EEG signals that better reflect the collaborative work of brain networks.

To systematically elucidate the design principles and advantages of QEPP, we clarify its potential superiority from the following aspects:

Integrated association modeling approach: Compared to the two-stage paradigm of “segmentation followed by association” in traditional EEG feature engineering, QEPP is closer to a native, integrated association modeling approach. Traditional methods typically extract statistical features (such as power and entropy) independently for each channel and then construct inter-channel relationships based on these features. However, in this process, the instantaneous phase synchronization or nonlinear coupling information inherent in the original signal is often partially smoothed or lost during single-channel processing. In contrast, QEPP explicitly introduces the instantaneous relationships between channels into the modeling process by directly constructing channel differential signals () at the original signal layer. These instantaneous relationships participate in the subsequent joint sorting and encoding along with the original signal, thus embedding local attributes and relationship information from the very beginning of feature generation, reducing the attenuation of association information in the processing chain.

Nonlinear, order-preserving relational modeling: Traditional metrics used to characterize channel relationships (such as coherence or Pearson correlation) are essentially linear measures, with limited sensitivity to nonlinear synchronization mechanisms (such as phase locking and amplitude coupling) prevalent in EEG. The argsort-based joint encoding in QEPP can be viewed as a nonlinear, order-preserving relational modeling approach: it weakens absolute amplitude information, is robust to noise, and preserves and strengthens relative ordering relationships within and between channels. This representation, centered on ordinal relationships, is naturally suited to capturing nonlinear rank associations, helping to characterize competition, collaboration, and dynamic rearrangement processes in brain activity.

Inductive bias for holistic relationships: From the perspective of inductive bias, QEPP explicitly strengthens the modeling tendency of holistic relationships at the algorithmic structure level. By forcibly introducing a relation vector that is processed with equal weight to the original signal, and by jointly sorting the three vectors, the final state sequence no longer has a decomposable single-channel meaning; its semantics are jointly determined by the relative positions within the overall sequence. This “indivisible state” representation echoes the metaphorical idea in quantum entanglement that individual states depend on the overall definition. Subsequently, the statistical analysis of joint state transition patterns by SCTT further characterizes the dynamic reconstruction patterns of brain networks in the overall state space.

By clarifying the design motivation and information processing advantages at the above levels, we argue that the performance improvement brought about by QEPP is not coincidental, but rather that its mathematical structure and neural signal modeling assumptions are more closely aligned with the characteristics of the brain as a dynamic, integrative, and relationally driven system.

The graphical depiction of the recommended QEPP feature extraction model is shown in Fig. 1.

Fig. 1.

The graphical outline of the presented QEPP feature extractor and the schematizing of the quantum entanglement. Herein, TR: Transformed Signal.

Traditional EEG feature extraction methods typically process individual channels independently or rely on linear spatiotemporal assumptions. However, extensive neuroscience research indicates that brain function arises from the large-scale synchronization and collaboration of distributed neural networks, and this dynamic cross-brain region association exhibits nonlocal and nonlinear characteristics^77,78. Classical linear methods may have limitations in capturing such complex couplings.

The connection between quantum mechanics and EEG signal processing can be understood from both mathematical and conceptual perspectives to address these limitations. In quantum mechanics, entangled particles exhibit non-local correlations where the state of one particle is inherently linked to the state of another, regardless of the distance between them. Similarly, EEG signals recorded from multiple channels demonstrate synchronous activity patterns, such as phase-locking and coherence, where neural oscillations in distant brain regions become temporally correlated. These multi-channel correlations in EEG can be mathematically analogous to the correlation structure observed in entangled quantum systems.

Furthermore, the brain functions as a complex network system where long-range correlations exist between different cortical regions. These correlations share mathematical similarities with quantum entanglement in terms of information structure and joint probability distributions. From a modeling perspective, EEG signals contain high-order correlation patterns that cannot be effectively captured by analyzing individual channels separately. Just as entangled quantum states require joint transformations for proper representation, these complex inter-channel dependencies in EEG may benefit from similar joint transformation approaches. The proposed QEP transformer is motivated by this analogy: by treating paired channel vectors as “entangled particles” and computing their joint transformations, we aim to capture the intrinsic nonlocal and nonlinear correlations that exist across EEG channels, thereby extracting more informative features for classification.

Figure 1 demonstrates that the presented QEPP is a quantum-inspired feature extraction method. The steps of the introduced QEPP feature extraction method are:

S1: Divide the multichannel EEG signal into overlapping matrices.

1

Herein,: the overlapped matrix,: length of the EEG signal : number of channels, : the created overlapped matrix with a size of .

S2: Compute the difference vector using the generated matrix.

2

3

where : the computed vectors.

S3: Sort the generated vectors and obtain identities to create transformed signals.

4

Here, : transformed signals and we have generated three transformed signals. The given S1-S3 have been defined the presented QEP transformer.

S4: Extract feature deploying SCTT feature extractor and the mathematical definition of this feature extractor is given below.

5

Here,: transition table with a size of and six transition tables have been defined in this phase. Subsequently, we filled these transition tables deploying SCTT feature extraction function.

6

7

8

9

Herein, : the length of the transformed signals. Then, the feature vectors have been created employing matrix to vector transformations.

10

where : individual feature vector with a length of and six individual feature vectors have been created. These features have been merged to created final feature vector and the presented feature merging method has been explained below.

11

Here, : final feature vector with a length of .

The steps S1-S4 clearly define the presented QEPP feature extraction process. Steps S1-S3 define the QEP transformer, while S4 represents the SCTT feature extractor.

The proposed explainable feature engineering framework

In this research, we have presented an XFE model to investigate the classification ability of the QEPP feature extraction function. To demonstrate the maximum classification capability of the introduced QEPP feature extractor, we have used CWINCA and tkNN methods, as both are self-organized feature selection and classification methods. To obtain interpretable results, we have also used the DLob XAI method. In this regard, the introduced XFE framework consists of four essential phases: (i) QEPP feature extraction, (ii) CWINCA feature selection, (iii) tkNN-driven classification, and (iv) DLob-based XAI. The graphical outline of the presented QEPP-related XFE model is shown in Fig. 2.

Fig. 2.

The outline of the introduced QEPP-centric XFE framework. Herein, T: transformed signal and f: individual feature vector.

The phases of the presented QEPP-centric XFE framework are outlined below.

Phase 1: Feature extraction: In this phase, the QEPP feature extractor has been utilized, and the details of this feature extractor are explained in Sect.  3. The first phase of the recommended QEPP-driven XFE model is feature extraction, and the steps of this phase are demonstrated below.

Step 1: Extract features employing QEPP feature extractor.

12

where : the generated feature matrix, : the QEPP feature extractor and : the number of observations.

Phase 2: Feature selection: To obtain the most informative features, we have used the CWINCA feature selector. This feature selector uses distances to compute feature weights, and these weights are generated through cumulative weight computation. The start and stop indexes of the loop are then determined. In the iterative feature selection process, multiple feature vectors are selected, and the best feature vector is chosen using a greedy algorithm. The CWINCA method is both a self-organized and iterative feature selector. Therefore, it has been used to extract the most informative features. The feature selection phase is defined in Step 2.

Step 2: Choose the most informative features deploying CWINCA feature selector.

13

Herein,: the identities of the selected features, : the selected feature matrix, : the CWINCA feature selector, : the real output and 0.85 and 0.99 are the used threshold points to determine start and stop indexes of the loop. Also, : the used classification function.

To better outline and explain the CWINCA method used, the algorithm for CWINCA is provided in Algorithm 1.

Algorithm 1.

The procedure of the CWINCA feature selection function.

Algorithm 1 clearly demonstrates that the CWINCA feature selector is a self-organized and iterative feature selection method. Therefore, this feature selector has been employed to choose the most informative features. The outputs of this feature selection function (CWINCA) have been utilized in both classification and interpretable results generation.

Phase 3: Classification: In the feature selection phase, a distance-centric self-organized classifier has been utilized to ensure classification consistency. Therefore, the tkNN classifier has been used. The tkNN classifier generates both parameter-based and voted outcomes, as it employs iterative parameter change and iterative majority voting (IMV) together. In the final step, the best outcome is selected based on classification accuracy, similar to the CWINCA method.

The classification step of this research is presented in Step 3, and the pseudocode for the tkNN classifier is demonstrated in Algorithm 2.

Algorithm 2.

The procedure of the tkNN classifier.

Step 3: Classify the chosen features employing the tkNN classifier.

14

Herein, the classification outcome, : the employed tkNN classifier and : the utilized parameters. The pseudocode of the tkNN classifier is also depicted in Algorithm 2.

In this algorithm (see Algorithm 2), the number of parameters can be increased as needed.

Phase 4: XAI: The presented XFE framework utilizes the DLob symbolic language. The DLob symbolic language consists of 16 DLob symbols, and the explanations of these symbols are provided below.

Frontal Lobe (FL, FR, Fz).

FL (Frontal Left): Governs logical, planning, reasoning and problem-solving.

FR (Frontal Right): Facilitates creativity, emotional regulation, and intuition.

Fz (Frontal Midline): Plays a key role in attention, focus, and executive control.

Temporal Lobe (TL, TR).

TL (Temporal Left): Processes language, memory, and auditory input.

TR (Temporal Right): Responsible for emotional processing, memory, and non-verbal auditory signals.

Central Region (CL, CR, Cz).

CL (Central Left): Think of CL as the originator of deliberate motion. It fine-tunes and directs precise motor actions, ensuring that our left side movements are both intentional and coordinated.

CR (Central Right): This symbol harmonizes sensory feedback with motor commands, allowing the right side of our body to respond adaptively to changing environments and maintain balance.

Cz (Central Midline): Cz plans, coordinates, and synchronizes movements, serving as the central hub that unites the efforts of both CL and CR into smooth, purposeful actions.

Parietal Lobe (PL, PR, Pz).

PL (Parietal Left): Manages sensory integration and spatial awareness.

PR (Parietal Right): Handles sensory input and spatial orientation.

Pz (Parietal Midline): Crucial for integrating sensory data and spatial reasoning.

Occipital Lobe (OL, OR, Oz).

OL (Occipital Left): Processes visual details, including shapes and colors.

OR (Occipital Right): Specializes in visual-spatial recognition and perception.

Oz (Occipital Midline): Fundamental for overall visual processing.

Auditory Cortex (AL, AR).

AL (Auditory Left): Interprets verbal sounds and speech-related cues.

AR (Auditory Right): Focuses on non-verbal sounds, such as music and environmental noise.

In this work, channel-to-DLob transformation has been utilized as a specific algorithm, as all six datasets used in this study were collected with three different brain caps containing 14, 32, and 35 channels. Therefore, three different look-up tables (LUTs) have been used for the channel-to-DLob transformation.

By utilizing the generated DLob symbols, a DLob string has been created for each dataset. Using the extracted DLob string, Shannon Entropy and the CCD have been computed.

The final step of the presented model is provided below.

Step 4: Extract explainable outcomes employing the DLob symbolic language.

15

where : explainable outcome, : the DLob-based XAI results generation function, : the identities of the selected features and : the utilized LUT.

The utilized DLob-centric XAI method’s pseudocode has been given in Algorithm 3.

Algorithm 3.

The DLob-centric XAI method’s pseudocode.

By utilizing Algorithm 3, the explainable/interpretable results have been generated.

Experimental results

The major goal of this research is to present a general, highly accurate, and explainable feature engineering model. To achieve this objective, we have introduced the QEPP-centric XFE model. This model has a simple structure and attains high classification performance. Furthermore, the recommended QEPP-centric XFE framework generates interpretable results.

In this section, classification and interpretable results are evaluated. To demonstrate the general classification performance of the presented model, we have used six datasets, and their characteristics are summarized in Table 2.

Table 2.

The characteristics of the used datasets for tests.

No	Dataset	Number of classes	Number of observations	Number of channels
1	ALS	2	5268	32
2	Artifact	8	2498	14
3	Stress	2	3667	14
4	Violence	2	778	14
5	Psychosis	2	4,098	32
6	Epilepsy	2	10,356	35

Initially, we downloaded these datasets to our personal computer (PC). The PC used is a simple-configured system with 32 GB of main memory, a 3.2 GHz processor, and the Windows 11 operating system.

To program the introduced QEPP-centric XFE framework, MATLAB 2024a was utilized. This framework was implemented using M-files, which include: (i) Main functions, (ii) QEP transformers, (iii) SCTT feature extractor, (iv) CWINCA feature selector, (v) tkNN classifier, and (vi) DLob-based XAI.

The defined six functions were called from the main function. The proposed QEPP-related XFE framework is a parametric model, and the utilized parameters in this XFE model are listed in Table 3, along with the time burden of the methods used. The computational complexity of the QEPP-centric XFE framework is analyzed using Big O notation, considering the time burden of each phase. The parameters and methods used in the framework are also detailed. Table 3 tabulates the time complexity associated with each phase of the model.

Table 3.

Time burden (Big O Notation) of the presented model and the utilized parameters for the QEPP-centric XFE framework.

Phase	Method	Parameters	Time burden
Feature extraction	QEP	The size of the utilized input: , The generation method of third output: Subtraction, Identity generation method: Sorting in descending, The number of the transformed signal: 3.
	SCTT	The size of the transition tables: , The number of the transition tables: 6, The length of the features: .
Feature selection	CWINCA	Threshold values: 0.85, 0.99, The classification accuracy computation function: kNN with 10-fold CV, The best outcome selection method: Greedy algorithm,
Classification	tkNN	Number of the used distance: 3, Number of the used weights: 2, Number of the used k values: 10, Number of the parameter-based outcomes: 60, Number of the voted outcomes: 58, The number of total outcomes: 118, The best outcome selection method: Greedy algorithm,
XAI	DLob-based XAI generation	Number of DLob symbols used, For 14 channels: 8, For 32 channels: 13, For 35 channels: 12. The used statistical analysis methods: Shannon Entropy, transition table

**The explanations of the symbols are given as follows. Here, : length of the used EEG signals, : the number of the channels, : NCA’s time complexity coefficient, : range of the feature selection iteration, : kNN’s time complexity coefficient, : Greedy algorithm’s time complexity coefficient, : the number of iterations in the tkNN or number of the parameters, : the time complexity coefficient of the IMV and : the length of the selected features,.

Table 3 demonstrates both the utilized parameters and the time burden of the introduced QEPP-related XFE model. The total time complexity of this model is computed as: . This result openly indicates that the introduced QEPP-related XFE model has a linear time burden.

Classification results

To generate the classification results of the presented QEPP-centric framework, the tkNN classifier has been utilized, and 10-fold CV has been applied to obtain classification results. In this study, six datasets have been used.

To evaluate the classification performance of the QEPP-centric XFE framework on the utilized datasets, confusion matrices have been generated. The computed confusion matrices are presented in Fig. 3.

Fig. 3.

The generated confusion matrices by the introduced QEPP-centric XFE framework. The meaning of these numbers is explained in Table A3.

As seen in Fig. 3 and 100% classification accuracy has been achieved for four datasets: stress, violence, psychosis, and epilepsy. To evaluate the classification performance of the presented XFE model, classification accuracy and geometric mean have been utilized. The datasets considered in this study exhibit varying degrees of class imbalance, from moderate to extreme. Therefore, we report both classification accuracy and the geometric mean (G-mean) when evaluating model performance. While accuracy reflects the overall correct classification rate, the G-mean, defined as the geometric mean of sensitivity and specificity, provides an equal measure of performance on both majority and minority classes and is more robust for assessing discriminative capability under class imbalance. The computed classification performances are summarized in Table 4.

Table 4.

The classification results (%) of the recommended QEPP-driven XFE framework for the used six datasets. These results have been computed deploying 10-fold CV.

No	Dataset	Accuracy	Geometric mean
1	ALS	98.25	98.25
2	Artifact	90.07	82.29
3	Stress	100	100
4	Violence	100	100
5	Psychosis	100	100
6	Epilepsy	100	100

Table 4 obviously demonstrates that the introduced XFE framework achieved over 90% classification accuracy and over 80% geometric mean across all utilized datasets. Furthermore, the recommended QEPP-related XFE framework attained 100% classification accuracy for four datasets. For the moderately imbalanced datasets, namely Violence and Psychosis, no resampling strategy was applied in order to evaluate the proposed method under a class distribution closer to the original data. To compensate for the potential bias caused by class imbalance, we report the geometric mean (G-mean) in addition to accuracy as a core evaluation metric (see Table 4). Since G-mean equally reflects the performance on both majority and minority classes, it provides a more reliable assessment of discriminative performance under imbalance. The near 100% accuracy together with the high G-mean values indicates that the superior performance is not due to a trivial fit to the majority class.

Explainable results

The second output of the presented QEPP-related XFE framework is interpretable results (XAI). In this section, a connectome diagram has been created for each dataset, and histograms of the utilized DLob symbols have been presented. The computed connectome diagrams for these datasets are shown in Fig. 4.

Fig. 4.

The explainable results of the used datasets.

The computed Shannon entropies and their corresponding complexity values are presented in Table 5.

Table 5.

The computed information entropies of the generated DLob strings deploying the introduced QEPP-driven XFE framework.

No	Dataset	Entropy	Number of the Dlob symbols	Maximum entropy	Complexity ratio (%)
1	ALS	3.3548	14	3.8074	88.11
2	Artifact	2.6222	8	3	87.41
3	Stress	2.6967	8	3	89.89
4	Violence	2.4578	8	3	81.93
5	Psychosis	3.4825	14	3.8074	91.47
6	Epilepsy	2.2801	13	3.7004	61.62

Table 5 clearly showcases that epilepsy detection is the most predictable process, while psychosis detection is the most complex, with a computed complexity ratio of 91.47%.

Neuroscientific interpretation and functional network mapping

To strengthen the neuroscientific interpretability of the proposed QEPP-XFE framework beyond outcome visualization, the DLob connectome diagrams and Shannon entropy analyses were interpreted in the context of large-scale functional brain networks and disease-related neural mechanisms.

Overall, the highlighted DLob patterns and CCD connections consistently involve frontal, parietal, central, temporal, and occipital regions, which correspond to well-established functional systems in cognitive neuroscience, such as executive control, sensorimotor integration, sensory–visual processing, and temporal–limbic networks. This network-level interpretation is in line with recent EEG-based graph and attention models that explicitly relate learned connectivity patterns to functional brain systems, such as the fronto-parietal regulatory and control networks in emotion and cognition⁷⁹. This indicates that the proposed framework does not merely emphasize isolated channels, but captures system-level functional organizations relevant to cognition and neurological disorders.

• In ALS detection, the frequent involvement of central regions (CL, CR, Cz) and their transitions reflects alterations in bilateral motor coordination and interhemispheric integration, which are core characteristics of motor neuron degeneration. The additional engagement of frontal and parietal regions suggests compensatory recruitment of executive and attentional networks, a phenomenon widely reported in neuroimaging studies of ALS, where patients rely on higher-order cognitive resources to maintain motor performance.

• In stress detection, the dominance of frontal regions, particularly right frontal activity, points to the central role of prefrontal systems in emotional regulation and executive control under acute stress. The relatively limited cross-regional transitions indicate a more localized prefrontal engagement, consistent with models of stress processing where regulatory control is prioritized over distributed network integration.

• For violence-related EEG patterns, the concurrent involvement of frontal, parietal, and occipital regions reflects the joint engagement of executive, sensory, and visual processing systems. This multi-network activation is consistent with the need for integrated cognitive control, perceptual processing, and visual attention during threat-related or high-arousal conditions.

• In psychosis detection, the high complexity of the DLob sequences together with the absence of specific long-range connections in the connectome diagrams suggests disrupted functional integration rather than simple regional hypo- or hyperactivation. This pattern is in line with the dysconnectivity hypothesis in psychotic disorders, which proposes that symptoms arise from impaired coordination between distributed brain networks despite preserved local activity.

• In epilepsy detection, the dominance of temporal lobe regions combined with lower complexity values reflects more stereotypical and predictable activation patterns, consistent with the hypersynchronous and recurrent network dynamics commonly observed in temporal lobe epilepsy.

Importantly, these observations also provide an interpretative grounding for the quantum-inspired notions of wholeness and indivisibility adopted in QEPP. The proposed framework does not rely solely on single-channel activations, but rather on joint state configurations and their transitions captured by the SCTT. In several cases, discriminative information emerges from the collective configuration of multiple regions rather than from any single channel alone, indicating that the diagnostic patterns are fundamentally system-level and cannot be reduced to independent component-wise interpretations. In this sense, QEPP captures holistic, network-level dynamics of brain activity, moving the interpretability analysis from simple outcome visualization toward a mechanism-oriented, neuroscience-informed explanation.

From a clinical perspective, the interpretability outputs of the QEPP-XFE framework can be further discussed in terms of their plausibility with respect to established diagnostic knowledge and disease-related neurophysiological findings. It is important to emphasize that the following discussion does not aim to draw direct clinical conclusions or to claim clinical validation, but rather to evaluate whether the highlighted regions and connectivity patterns are consistent with current clinical and neuroscience understanding.

• Previous clinical and neuroimaging studies have shown that ALS is not limited to motor neuron degeneration, but is frequently accompanied by impairments in prefrontal executive functions and parietal sensory integration, as well as altered frontoparietal and interhemispheric connectivity. In this context, the prominence of central regions and frontal–parietal patterns in the DLob and connectome results appears clinically plausible, as it is consistent with known motor coordination deficits and compensatory recruitment of higher-order cognitive networks reported in ALS. The emphasis on interhemispheric central transitions can therefore be interpreted as reflecting altered bilateral motor network coordination, which is a well-documented aspect of ALS pathophysiology.

• For stress and violence datasets, the involvement of frontal regions, particularly right prefrontal areas, is clinically meaningful given the central role of the prefrontal cortex in emotion regulation, stress response, and executive control. In affective and social cognitive neuroscience, stress and impulsive or aggressive behaviors are often associated with insufficient prefrontal regulation and abnormal integration of sensory information. Accordingly, the concentration of discriminative patterns in frontal regions for stress, and the broader engagement of frontal, parietal, and occipital regions for violence-related EEG patterns, appears plausible as a reflection of coordinated executive, sensory, and visual processing during high-arousal or threat-related conditions.

• Psychotic disorders, including schizophrenia spectrum conditions, are increasingly conceptualized as disorders of large-scale brain network integration rather than focal regional abnormalities. The high sequence complexity together with the structurally disorganized connectivity patterns observed in the connectome diagrams are compatible with the dysconnectivity framework, which emphasizes impaired coordination between distributed functional networks. In this sense, the QEPP-XFE interpretability outputs appear consistent with the notion that psychosis involves widespread but poorly coordinated network activity rather than isolated regional dysfunction.

• Temporal lobe epilepsy is known to exhibit relatively well-defined and reproducible electrophysiological patterns, often characterized by hypersynchronous and stereotypical network dynamics. The dominance of temporal lobe-related patterns together with lower complexity and higher predictability in the proposed framework aligns well with the classical clinical understanding of epileptiform activity, where pathological network dynamics tend to be repetitive and less variable across time.

• EEG artifacts originate from diverse non-neural sources such as eye movements, muscle activity, and electrode-related effects, and therefore typically exhibit heterogeneous spatial and temporal patterns across the scalp. The relatively higher complexity and multi-regional involvement observed in the artifact-related interpretability results are in line with this clinical knowledge, where artifacts are recognized as varied and multi-source phenomena rather than manifestations of a single, localized neural process.

By comparing the interpretable outputs of the QEPP-XFE framework with established clinical and neuroscience knowledge, the presented results appear clinically plausible at a conceptual level. While this does not constitute clinical validation, it supports the credibility of the proposed interpretability approach and suggests that the highlighted regions and connectivity patterns are meaningfully related to known disease mechanisms and neurophysiological processes. This perspective strengthens the potential clinical relevance of the QEPP-XFE framework for future applied and translational studies.

Ablation study and comparative analysis

To comprehensively evaluate the contribution of each component in the proposed QEPP-centric XFE framework, we conducted extensive ablation studies and comparative experiments. These experiments address three critical aspects: (1) the discriminative power of QEPP features compared to other feature extraction methods, (2) the individual contributions of the XFE framework components (CWINCA and tkNN), and (3) the computational efficiency comparison with deep learning models.

Comparison of feature extraction methods

To demonstrate that the performance improvement originates from the QEPP feature extractor rather than solely from the post-processing pipeline, we compared QEPP with representative classical feature extraction methods while keeping the downstream processing (CWINCA + tkNN) identical. We selected three well-established feature extraction approaches as baselines:

• Wavelet Features (WF): Discrete wavelet transform coefficients extracted using db4 wavelet with 4-level decomposition, followed by statistical features (mean, variance, energy) from each sub-band.

• Functional Connectivity Features (FC): Phase Locking Value (PLV) and coherence-based connectivity matrices computed between all channel pairs.

• Time-Frequency Features (TF): Short-time Fourier transform (STFT) based power spectral density features across delta, theta, alpha, beta, and gamma bands.

All feature extraction methods were evaluated using the same CWINCA feature selector and tkNN classifier with 10-fold cross-validation. The comparative results on two representative datasets (Artifact and Epilepsy) are presented in Table 6.

Table 6.

Comparison of different feature extraction methods using identical downstream processing (CWINCA + tkNN) with 10-fold CV.

Dataset	Feature extraction	CWINCA + tkNN accuracy (%)	Geometric mean (%)
Artifact	Wavelet Features (WF)	81.35	72.18
Artifact	Functional Connectivity (FC)	78.42	68.93
Artifact	Time-Frequency Features (TF)	79.86	70.54
Artifact	QEPP (Proposed)	90.07	82.29
Epilepsy	Wavelet Features (WF)	92.47	91.83
Epilepsy	Functional Connectivity (FC)	89.25	88.16
Epilepsy	Time-Frequency Features (TF)	91.08	90.22
Epilepsy	QEPP (Proposed)	100.00	100.00

As shown in Table 6, the QEPP feature extractor consistently outperforms traditional feature extraction methods across both datasets when using identical downstream processing. This result demonstrates that the discriminative power of QEPP features is inherently superior to classical approaches, validating the effectiveness of the quantum-inspired feature extraction paradigm.

Ablation study of XFE framework components

To investigate the individual contributions of each component in the XFE framework, we conducted systematic ablation experiments with the following configurations:

Case 1

QEPP → CWINCA → tkNN.

Case 2

QEPP → tkNN (using all QEPP features).

Case 3

QEPP → CWINCA → Standard kNN (k = 5, Euclidean distance).

Case 4

QEPP → Standard kNN (no feature selection, standard classifier).

The ablation results on the Artifact and Epilepsy datasets are presented in Table 7.

Table 7.

Ablation study results showing the contribution of each XFE framework component with 10-fold CV.

Dataset	Case	Components	Accuracy (%)	Δ vs. full model
Artifact	1	QEPP + CWINCA + tkNN	90.07	0
Artifact	2	QEPP + tkNN	84.23	− 5.84
Artifact	3	QEPP + CWINCA + Std kNN	86.71	− 3.36
Artifact	4	QEPP + Std kNN	82.15	− 7.92
Epilepsy	1	QEPP + CWINCA + tkNN	100.00	0
Epilepsy	2	QEPP + tkNN	96.28	− 3.72
Epilepsy	3	QEPP + CWINCA + Std kNN	97.85	− 2.15
Epilepsy	4	QEPP + Std kNN	94.62	− 5.38

The ablation results reveal several important findings:

1. Value of CWINCA (Case 1vs. Case 2): The CWINCA feature selector provides approximately 5.84% improvement by selecting the most discriminative features and reducing noise in the feature space.

2. Value of tkNN (Case 1vs. Case 3): The tkNN classifier contributes approximately 3.36% improvement over standard kNN through its parameter optimization and iterative majority voting mechanism.

3. Combined Effect (Case 1vs. Case 4): The full XFE framework achieves 7.92% improvement over the baseline configuration, demonstrating the synergistic effect of CWINCA and tkNN.

4. QEPP Baseline Performance: Notably, even Case 4 (QEPP + Standard kNN) achieves competitive performance (82.15% for Artifact, 94.62% for Epilepsy), which is higher to other feature extraction methods combined with the full CWINCA + tkNN pipeline (see Table 6). This finding strongly validates the inherent discriminative power of QEPP features.

Ablation study of XFE framework components

To quantitatively validate the “lightweight” claim of our framework, we compared the QEPP-centric XFE model with mainstream deep learning architectures commonly used for EEG classification. We selected three representative models:

• EEGNet: A compact CNN architecture specifically designed for EEG classification.

• CNN-LSTM: A hybrid architecture combining convolutional and recurrent layers.

• Lightweight Transformer: A reduced-parameter transformer model adapted for EEG signals.

All models were evaluated on the same hardware environment (CPU: Intel Core i7 @ 3.2 GHz, RAM: 32 GB, GPU: NVIDIA RTX 3080 for DL models) using identical data splits. The comparison results are presented in Table 8.

Table 8.

Comprehensive comparison of QEPP-XFE with deep learning models on Artifact and Epilepsy datasets.

Model	Dataset	Accuracy (%)	Training Time (s)	Inference Time (ms/sample)	Time complexity	Memory (MB)
EEGNet	Artifact	85.72	245	2.1	Exponential	156
EEGNet	Epilepsy	93.48	412	2.1	Exponential	168
CNN-LSTM	Artifact	83.19	387	4.7	Exponential	284
CNN-LSTM	Epilepsy	91.25	623	4.7	Exponential	312
Lightweight Transformer	Artifact	82.45	524	3.4	Exponential	245
Lightweight Transformer	Epilepsy	89.67	856	3.4	Exponential	278
QEPP-XFE (Proposed)	Artifact	90.07	18	0.8	Linear	45
QEPP-XFE (Proposed)	Epilepsy	100.00	32	0.8	Linear	52

The efficiency comparison reveals several advantages of the proposed QEPP-XFE framework:

1. The QEPP-XFE framework requires significantly less training time compared to deep learning models, as it does not involve iterative gradient-based optimization.

2. The inference time of QEPP-XFE is competitive with or faster than deep learning models, making it suitable for real-time EEG applications.

3. Unlike deep learning models that require GPU acceleration for efficient training, QEPP-XFE can be executed entirely on CPU with minimal memory footprint.

4. While achieving comparable or superior classification accuracy, QEPP-XFE demonstrates a favorable balance in the performance-efficiency-resource trade-off, validating its characterization as a lightweight model.

The comprehensive experiments presented in this section provide strong evidence for the following conclusions:

1. The QEPP feature extractor generates inherently more discriminative features compared to traditional methods (wavelet, functional connectivity, time-frequency), as demonstrated by controlled experiments with identical downstream processing.

2. Both CWINCA feature selection and tkNN classification contribute meaningfully to the overall performance, with their combination providing synergistic benefits.

3. The QEPP-XFE framework achieves competitive or superior performance compared to deep learning models while requiring substantially less computational resources, validating its suitability for resource-constrained and real-time applications.

Discussions

Our model achieved 98.25%, 90.07%, 100%, 100%, 100%, and 100% accuracy for ALS detection, artifact classification, stress detection, violence detection, psychosis detection, and epilepsy detection, respectively. Moreover, the introduced QEPP-centric XFE framework generates XAI results for each dataset used.

To implement the introduced model, the QEPP-centric XFE framework has been designed using cognitive steps. First, we introduced a new-generation transformer for feature engineering, leveraging the strength of transformers to propose a competitive EEG classification framework against DL models. To extract features, a new feature extractor (SCTT) has been introduced. By utilizing this feature extractor, different relationships have been extracted as features.

In the feature selection and classification phases, two self-organized methods, CWINCA and tkNN, have been used to achieve optimal classification performance. The extracted features also include channel information to enable the generation of XAI results. By applying channel-to-DLob symbol transformation, a DLob sentence has been created for each dataset, and interpretable results have been extracted using the generated DLob sentences.

The results demonstrate that the presented QEPP-based XFE framework achieved 100% accuracy for four datasets: stress detection, violence detection, psychosis detection, and epilepsy detection. The ablation studies reveal that QEPP features possess inherent discriminative power. Even when paired with a standard kNN classifier without any optimization, QEPP achieves 82.15% accuracy on the Artifact dataset, which is comparable to other feature extraction methods combined with sophisticated processing pipelines. This validates that the core innovation—the quantum-inspired feature extraction—is the primary driver of the framework’s success. For these datasets, leave-one-record/subject-out (LORO/LOSO) CV has been applied, and the computed results are depicted in Fig. 5.

Fig. 5.

The confusion matrices of the presented QEPP-driven model deploying leave-one record/subject out CV.

The computed classification performances of these datasets using LOSO/LORO CVs have been tabulated in Table 9.

Table 9.

The computed classification results deploying the presented QEPP-centric XFE framework.

Dataset	Accuracy	Geometric mean
Stress	75.35	75.37
Violence	99.45	99.55
Psychosis	98.24	98.55
Epilepsy	85.42	84.12

Table 9 clearly demonstrates that all classification results exceed 75% when using LOSO/LORO CVs. These results further validate the high classification performance of the presented QEPP-based XFE framework.

To position the QEPP-based framework within the literature, a comparative results table is provided in Table 10.

Table 10.

The comparative results for same datasets.

Dataset	Study	Method	XAI	Split ratio	Accuracy
ALS	80	VGG19	No	5-fold CV	80.00
ALS	81	Transformer-based method	No	70:15:15	99.33
ALS	Our method	QEPP-centric XFE	Yes	10-fold CV	98.25
Artifact	71	Directed Lobish, transition table pattern CWINCA	Yes	10-fold CV	95.40
Artifact	Our method	QEPP-centric XFE	Yes	10-fold CV	90.07
Stress	72	Quadruple Transition Pattern	Yes	1. 10-fold CV 2. LOSO CV	1. 92.94 2. 73.63
Stress	82	Cubic pattern	Yes	1. 10-fold CV 2. LOSO CV	1. 96.29 2. 76.17
Stress	73	Channel-based minimum and maximum pattern	Yes	1. 10-fold CV 2. LOSO CV	1. 92.86 2. 73.30
Stress	Our method	QEPP-centric XFE	Yes	1. 10-fold CV 2. LOSO CV	1. 100.0 2. 75.35
Violence	73	Channel-based minimum and maximum pattern	Yes	1. 10-fold CV 2. LORO CV	1. 99.86 2. 99.31
Violence	Our method	QEPP-centric XFE	Yes	1. 10-fold CV 2. LORO	1. 100.0 2. 99.45
Psychosis	74	Zipper pattern, iterative NCA	Yes	1. 10-fold CV 2. LORO CV	1. 99.95 2. 96.12
Psychosis	Our method	QEPP-centric XFE	Yes	1. 10-fold CV 2. LORO CV	1. 100.0 2. 98.24
Epilepsy	75	Hypercube pattern	No	LOSO CV	87.78
Epilepsy	83	EpilepsyNet	No	10-fold CV	85.00
Epilepsy	84	Xception	No	10-fold CV	87.42
Epilepsy	85	VGG16	No	10-fold CV	91.13
Epilepsy	68	Transformer	No	10-fold CV	85.00
Epilepsy	86	MobileNet	No	10-fold CV	91.66
Epilepsy	87	CWT-based DCNN	No	10-fold CV	95.99
Epilepsy	88	Archimedean Spiral and Swin Transformer	No	10-fold CV	97.98
Epilepsy	Our method	QEPP-centric XFE	Yes	1. 10-fold CV 2. LOSO CV	1. 100.0 2. 85.42

Per Table 10, our method shows competitive performance and provides explainable outputs.

We have generated explainable results by deploying the QEPP-driven XFE model. Based on the interpretable results, the following findings have been obtained:

ALS Detection

The most activated lobes are the frontal and parietal lobes. This demonstrates that ALS can cause defects in these regions. Additionally, strong activations in CL, CR, and Cz indicate frequent hemispheric transitions in ALS detection. Moreover, CL, CR and Cz can represent motor coordination deficits.

Artifact Classification

The frontal lobe is the most activated. Activations are also observed in the temporal, parietal, and occipital lobes, highlighting the complex nature of artifact classification.

Stress Detection

FR is the most activated DLob symbol, indicating that stress is an internal cognitive process. Also, FR dominance represents executive and emotional overload.

Violence Detection

The most frequently activated DLob symbols are FR, FL, PR, and OR. This suggests that violence affects cognitive, sensory, and visual activations.

Psychosis Detection

The frontal, parietal, and occipital lobes are affected. The most complex DLob sentence belongs to this dataset, reflecting the cognitive and perceptual disruptions in psychosis. The connectome diagram of the psychosis detection shows that widespread disruptions affect multiple brain areas.

Epilepsy Detection

The temporal lobe is the most frequently activated, demonstrating that most epilepsies originate in the temporal lobe. Epilepsy detection is the most predictable classification task.

These findings validate the effectiveness of the QEPP-based XFE framework in both classification performance and interpretability.

The most important points of this research have discussed as below.

Findings:

• QEPP is a novel quantum-inspired feature extraction method that integrates a QEP transformer and a SCTT to increase EEG signal processing.

• Multichannel EEG transformation is achieved using quantum entanglement principles, where paired vectors and their differences generate three transformed signals. The presented QEP transformer improves feature representation.

• The introduced XFE framework is uses two self-organized methods and these methods are CWINCA and tkNN. Due to these self-organized methods, high classification performances have been yielded.

• Six diverse EEG datasets (ALS, artifact, stress, violence, psychosis, epilepsy) validate the framework, demonstrating over 90% classification accuracy in all cases. Also, four datasets attained 100% accuracy with 10-fold CV.

• The created DLob strings and cortical connectome diagrams highlight brain region activations.

• ALS Detection: Frontal and parietal lobes dominate. Heavy activation in CL, CR, and Cz and it indicates frequent hemispheric transitions. ALS disrupts motor and cognitive integration.

• Artifact Classification: Frontal lobe is dominant. Temporal, parietal, and occipital lobes are also active. Artifacts cause widespread neural noise. EEG complexity increases due to non-neural signals.

• Stress Detection: FR is the most activated symbol. Stress is an internal cognitive process. Dominant frontal activation suggests executive and emotional load.

• Violence Detection: FR, FL, PR, and OR dominate. Violence engages cognitive, sensory, and visual networks. Frontal involvement represents decision-making and impulse control.

• Psychosis Detection: Frontal, parietal, and occipital lobes are affected. The most complex DLob sequence. Disruptions in logical reasoning, sensory integration, and perception. However, some symbols (such as TR-TL, TL-Fz, OL-Pz) have no connection. This situation demonstrated that psychosis participants cannot use their brain effectively.

• Epilepsy Detection: Temporal lobe is the most active. Most epileptic seizures are temporal epilepsies. The most predictable EEG pattern is the epilepsy detection.

• All datasets have their own unique connectome diagrams.

Advantages:

• The presented QEPP-based XFE framework is a lightweight model but this model attained high classification performances.

• This model yielded over 90% classification accuracy on all six datasets used.

• Our model attained over 75% classification performances with LOSO/LORO CV.

• This model is an XFE model and the interpretable results were computed. Actually, the generated cortical connectome diagrams are unique.

Limitation:

• Performance dropped with LOSO CV (e.g., 75.35% for stress).

Future directions:

• Extension to other biomedical signals: We plan to extend the QEPP-driven XFE framework to other biomedical time-series signals such as ECG and EMG for cardiac and neuromuscular disorder detection, in order to evaluate the generalization capability of the proposed representation across different physiological signal modalities.

• QEPP-enhanced hybrid deep learning architectures: A concrete future direction is to integrate QEPP as a structured inductive bias module into lightweight deep learning models by making key operations (e.g., ordering and ranking) differentiable using approximate differentiable sorting operators. This would enable end-to-end training of a QEPP-enhanced hybrid architecture and allow systematic comparisons with standard lightweight CNN models (e.g., EEGNet) in terms of performance, efficiency, and interpretability.

• Language-model-based structured interpretation: To improve the readability and clinical usability of DLob representations, custom language models or structurally constrained large language models can be developed to translate DLob symbols and connectivity patterns into neuroanatomically consistent and clinically meaningful textual descriptions, thereby bridging the gap between visual explainability and narrative clinical reporting.

• Exploration of alternative quantum-inspired representations: Beyond the current QEPP formulation, other quantum-inspired models and diagrammatic representations can be investigated to design new-generation, interpretable, and computationally efficient signal classification frameworks, and their effectiveness can be systematically compared with the proposed approach.

Potential applications:

• Stress/anxiety detection in workplaces or clinics.

• Real-time EEG monitoring for epilepsy/ALS patients with explainable insights.

• This framework can help to develop new generation drugs for brain-related disorders.

• We can adapt this framework for cardiac (arrhythmia) or respiratory signal classification.

Conclusions

This study introduced a quantum-inspired feature extraction method (QEPP) and a new explainable feature engineering framework for EEG signal classification. The proposed model achieved 98.25% accuracy for ALS detection, 90.07% for artifact classification, and 100% accuracy for stress, violence, psychosis, and epilepsy detection using 10-fold cross-validation. The framework uses a transformer-based QEP and SCTT extractor together with CWINCA and tkNN. It is lightweight and runs in linear time.

The model also produced clear interpretable results. The DLob-based approach generated DLob strings and cortical connectome diagrams that reveal brain region activations. The computed Shannon entropies and complexity ratios (up to 91.47%) further support the interpretability. These results demonstrate that the proposed QEPP-based framework is an effective and efficient tool for EEG signal classification and analysis.

The introduced QEPP-centric XFE framework contributes to feature engineering by proposing a transformer-based feature extraction function and this work creates a competitive alternative to DL models. This XFE framework is validated on six EEG datasets, making it a general and high-performance EEG classification model. Moreover, the interpretable results highlight the neurological significance of the extracted features. The DLob-based findings indicate the dominant brain lobes involved in different EEG signal classifications, aligning with known neurological patterns.

The high classification accuracy and interpretable findings position this model as a valuable tool for neurological disorder detection, cognitive research, and clinical applications. The framework bridges the gap between performance and interpretability. This model is a practical choice for real-world EEG-based diagnostics and AI-driven neuroscience studies.

Author contributions

Conceptualization, FAA, MSNY, KAA, MJ, OFG, MB, SD, TT; methodology, FAA, MSNY, KAA, MJ, OFG, MB, SD, TT; software, SD, TT; validation, OFG, MB, SD, TT; formal analysis, FAA, MSNY, KAA, MJ, OFG; investigation, FAA, MSNY, KAA, MJ; resources, OFG, MB; data curation, FAA, MSNY, KAA, MJ; writing—original draft preparation, FAA, MSNY, KAA, MJ, OFG, MB, SD, TT; writing—review and editing, FAA, MSNY, KAA, MJ, OFG, MB, SD, TT; visualization, FAA, MSNY, KAA, MJ, OFG, MB, SD, TT; supervision, TT; project administration, TT.

Funding

This research study and the Article Processing Charge (APC) were supported by Ongoing Research Funding program (ORF-2026-1392), King Saud University, Riyadh, Saudi Arabia.

Data availability

The datasets used and analyzed during the current study are publicly available and were obtained from third-party sources. The original sources of the datasets are cited in the manuscript. No new data were generated in this study.

Declarations

Competing interests

The authors declare no competing interests.