Quantum inspired wavelet and Fourier feature fusion for EEG based epilepsy and seizure detection

Abstract

The electroencephalography (EEG) signals are the cheapest approach to study the brain information, commonly used for epilepsy and seizure detection. This study presents a QFF-ML Net (Quantum Feature Fused Machine Learning Network). This novel approach integrates Quantum Wavelet Transform (QWT) and Quantum Fourier Transform (QFT) with a dimensionality reduction method. The raw EEG signals are not directly transmitted for remote monitoring; signal preprocessing and feature extraction are done using PCA applied dataset with QWT and QFT, then the quantum encrypted file is transferred via Ultra-reliable low-latency communication (uRLLC), which facilitates real-time remote diagnosis. This approach effectively reduces computational complexity while preserving critical seizure-related features, making it well-suited for bandwidth-constrained environments. Experimental findings show that QFF-ML Net achieved high performance across datasets, with accuracies of 92.3% on Bonn, 91.0% on CHB-MIT, and 88.3% on real-time 5G remote data, maintaining strong precision, recall, and F1-scores in all cases in EEG-based epilepsy and seizure detection on 5G-enabled remote healthcare networks.

Introduction

Epilepsy is a common chronic brain disorder affecting around 50 million people globally, with nearly 80% living in low and middle income countries. Despite effective treatments, up to 70% could live seizure-free if properly diagnosed. However, many remain untreated, especially in low-income regions, facing higher mortality risks and social stigma¹. Electroencephalography (EEG) signals are mostly used to capture brain activity and are widely applied in the detection of neurological disorders such as epilepsy, Alzheimer’s disease, and various sleep disorders. In that accurate detection of these conditions through EEG plays a crucial role in neurology^2,3. Epilepsy is classified into two main types partial seizures, which originate in a specific region of the brain, and generalized seizures, which affect the entire brain⁴ (Table 1).

Table 1.

Neurological and psychiatric disorders detectable using EEG.

Disorder category	Example disorders	EEG use
Seizure/Epilepsy	Generalized, Focal seizures	Diagnosis & monitoring
Sleep	Insomnia, Narcolepsy	Sleep stage analysis
Cognitive/Dementia	Alzheimer’s, Lewy Body, Vascular dementia	Cognitive decline marker
Brain Damage	Stroke, TBI	Localizing damage
Psychiatric	Depression, Schizophrenia, Bipolar	Abnormal rhythms/coherence
Developmental	ADHD, Autism	Neurofeedback / markers

This image compares a normal brain (left) with an epileptic brain (right), highlighting three abnormal regions. Type 1 represents Focal Cortical Dysplasia, a developmental brain abnormality. Type 2 shows Mesial Temporal Sclerosis, involving scarring in the temporal lobe. Type 3 depicts Tuberous Sclerosis Complex, a genetic condition causing brain lesions. These marked areas illustrate common structural causes of epilepsy (Fig. 1). Recently, the rapid development of Artificial Intelligence (AI) and 5G wireless networks enables the Internet of Medical Things (IoMT) for remote patient monitoring, remote consultation, remote diagnosis, and the new generation wireless communication enables low latency communication this enables remote surgery and so on in the medical field⁵.This research focuses on improving heart disease prediction using ECG sensor data by applying advanced machine learning techniques. After preprocessing the ECG signals with a Gaussian filter to remove noise and baseline drift, two neural network models—Multilayer Perceptron (MLP) and Gated Recurrent Units (GRU)—are used for analysis.

Fig. 1.

Comparison between a normal brain and epileptic brain with multiple lesion types.

The proposed GRU-MLP model achieves up to 99.5% accuracy, highlighting its effectiveness in early and reliable detection of heart disease⁶ ,The diagnosis and treatment of epilepsy continue to face numerous challenges. To address these, we developed a QFF-ML Net model that connects with remote telemedicine systems via 5G uRLLC.This framework enables real-time patient monitoring and data collection through 5G wireless networks. The collected EEG data is then processed by the QPCA-TL model, which is capable of detecting epilepsy even from low-sample inputs received from different user devices (dUEs).The following proposed methods are contributed in this article.

1. Introduced a QFF-ML Net model for efficient epilepsy and seizure detection.

2. Integrated with 5G-enabled uRLLC in ns-3 for real-time remote diagnosis.

3. Reduces human effort and cost in traditional epilepsy diagnosis.

4. Enables low-sample epilepsy detection in decentralized healthcare.

Literature survey

The new feature extraction method on EEG channels first converts EEG data into channel vectors, and calculate distance matrix representing pairwise channel distances. An average threshold is then applied to generate a counter or ”friendship” matrix, where values are marked as +1 or -1 based on their relation to the average. This matrix is transformed into a feature vector, capturing inter-channel relationships for classification. The FriendPat-centric XFE model is used on an EEG epilepsy detection dataset². Hybrid CNN-SVM and DNN-SVM models with Principal Component Analysis (PCA) are proposed to improve the epilepsy detection accuracy this approach achieved 99.42% accuracy when PCA was integrated. This was tested on the Epileptic Seizure Recognition dataset and the BONN dataset⁴. Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs) perform better accuracy in detecting epilepsy, but they lack data preprocessing and feature extraction steps and suffer from gradient vanishing in long sequences. To overcome this, Zhentao Huang used spatio-temporal feature fusion in EEG signal recognition with a dual attention mechanism⁷.Detecting epileptic seizures is time-consuming, and creating a dataset requires a large amount of human involvement. To overcome this, Xiaoshuang Wang et al. proposed a seizure detection approach using scalp EEG signals collected from individuals with frontal lobe epilepsy (FLE) and temporal lobe epilepsy (TLE).In this approach, data from 20 FLE patients and 20 TLE patients were utilized⁸. Discrete Wavelet Transform (DWT) is used for five-level decomposition of raw EEG signals to remove redundant features. Support Vector Machine–Recursive Feature Elimination is applied after the seizure states are classified using Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM), using the D–E group (Bonn dataset) and the Interictal–Ictal group (New Delhi dataset)⁹. In telemedicine applications, security is a major concern on patient records. For that, the EEG signal is encrypted using Chaotic Baker Map and Arnold Transform algorithms. The encrypted EEG data is classified and recognized by the IoMT system with CNN, and the proposed approach is compared with AlexNet, ResNet50, and SqueezeNet, but it outperforms well⁵. Childhood epilepsy syndrome classification and seizure detection: for this, LSTM is applied to the filtered EEGs of each channel to extract time characteristics. A 3D CNN is constructed to extract deep features of 21 channels. The extracted information is given as input into the fully connected layer of the dual-stream multitask network (DSMN) proposed module¹⁰. Kunxian Yan et al. proposed a Dynamic Temporal-Spatial Graph Attention Network for seizure detection.

This approach was validated on the Temple University Seizure Detection (TUSZ) dataset and achieved an accuracy ranging from 89% to 91%¹¹. Study¹² introduced a remote health monitoring system designed to support neurological care using a device that attaches to the outer ear. This small, non-invasive tool collects brain signals and sends them through secure cloud networks, making it easier for doctors to monitor patients from a distance. The signals are cleaned and analyzed with a smart processing method, helping identify early signs of neurological issues. The system shows strong potential for use in smart city healthcare, offering reliable support outside of hospitals. Focused on improving seizure detection through a global challenge using wearable EEG devices placed behind the ear. Participants worked with a large dataset from 42 epilepsy patients to create accurate detection models¹³. The winning solution used a decision tree approach with data enhancement techniques to boost accuracy. This effort highlighted the growing role of wearable devices and smarter data use in making epilepsy care more accessible and effective in everyday settings (Table 2).

Table 2.

Key findings and limitations from studies on EEG, URLLC, and IoT-based systems.

Author (Year)	Methodology (Model)	Feature engineering	Dataset used	Key findings	Limitations
Vos Gideon et al.14 (2025)	Literature review	EEG features (time, frequency)	60 studies on low-cost EEG	ML shows high accuracy	Lack of standardization
M VR et al.15 (2024)	Review of neuro devices	ML on neuroreceptor signals	Multi-sensor data	Enables cognitive health monitoring	Conceptual only
Hewen Bai et al.16 (2024)	Wireless EEG via Bluetooth	Signal filtering, noise reduction	Lab-collected 8-channel EEG	Reliable for home use	No clinical/seizure validation
Alkhaldi et al.17 (2024)	AI-based seizure classification	CNN, ML models	CHB-MIT, TUH datasets	96.8% sensitivity, 95.5% specificity	No real data testing
Razeena B. Shaik et al.18 (2024)	AI survey on URLLC	Beamforming, MIMO, allocation	Literature-based	Enhances URLLC in 5G/6G	No implementation detail
Moreno-Cardenas et al.19 (2024)	Economic modeling of networks	Game theory, slicing	Simulated interactions	DN model favors URLLC	Simplified/cheap simulation
Harun-Ar-Rashid et al.20 (2023)	IoMT via HL7, Raspberry Pi	HL7, encryption, FTP	Ultrasound images from hospital	Enables secure telemedicine transfer	Limited to HL7 systems
Milne-Ives et al.21 (2023)	Home EEG review for PwID	23-study analysis	Remote EEG use	Feasibility confirmed	Sparse sensitivity data
Sriraam et al.22 (2023)	Android seizure detection (SVM)	Matrix decomposition	5 EEG databases	97.5% specificity, low false alarms	Android-only, no deployment
Guiomar Niso et al.23 (2023)	Systematic survey	Channel, power, mobility	110 studies, 48 devices	EEG categorized by application	No experimental results
Gangadharan & Vinod24 (2022)	Drowsiness via SVM	Temporal EEG features	Portable EEG	78.3% accuracy	Signal quality impacts results
Madison M. Ives et al.25 (2022)	Scoping review (PRISMA)	PICOS framework	Remote epilepsy EEG	Gaps in adult epilepsy EEG use	Theory only
Biondi et al.26 (2021)	Wearable EEG analysis	HR, sleep, mood signals	Patient EEG + biometric data	Home seizure tracking improved	No AI modeling used
Campanile et al.27 (2020)	ns-3 simulation review	ns-3 extensions evaluated	128 papers (2009–19)	ns-3 widely used for IoT/Wireless	No real-world testing
Kim28 (2020)	URLLC design for 5G	Latency/reliability modeling	3GPP + theoretical data	Framework for low latency	No IoT validation
Siddiqi et al.29 (2019)	URLLC-IoT review	3GPP, latency models	Literature + specs	URLLC-IoT integration issues mapped	Theory only

Methods

Figure 2,illustrates 5G-enabled remote telemedicine system for epileptic and seizure recognition using EEG signals from the remote patient side. EEG data is collected from different users and transmitted via a 5G base station (gNB) using uRLLC to a central server for preprocessing and analysis.The QFF ML-Net model is deployed on the doctor’s side for remote diagnosis.It efficiently classifies patients EEG signals using quantum feature–fused remote EEG data.

Fig. 2.

5G-enabled remote EEG epilepsy detection via QFF-MLNet.

On the patient side, EEG signals are first acquired and preprocessed through denoising, filtering, and normalization after Principal Component Analysis (PCA) is applied for dimensionality reduction, followed by Quantum Wavelet Transform (QWT) and Quantum Fourier Transform (QFT) to extract complementary time–frequency and frequency-domain features. These features are concatenated into a single file and transmitted via the MQTT protocol over a 5G gNB (uRLLC) link, ensuring reliable and ultra-low-latency delivery.On the remote doctor’s side, the received feature file is processed using the proposed QFF-ML Net machine learning model,which performs classification and generates diagnostic results distinguishing between normal and abnormal EEG signals.The QWT provides localized time–frequency analysis, effectively capturing transient, non-stationary EEG patterns such as seizures, while the QFT efficiently extracts global frequency characteristics, making it well-suited for periodic or quasi-periodic EEG components. Combining features from both QWT and QFT yields complementary information—local detail from QWT and global spectral structure from QFT—thereby enhancing the discriminative power for EEG analysis (Table 3).

Table 3.

Experimental software and library setup for QFF-ML Net implementation.

Component	Library / Framework	Version / Notes
Programming Language	Python	3.10.8
Classical ML Ensemble	scikit-learn	1.7.2
Preprocessing (Scaling, SMOTE, Pipelines)	imbalanced-learn	0.14.0
Randomized Search CV	scikit-learn	1.7.2
Quantum Transformations (QFT, QWT)	PennyLane	0.42.3
Quantum Simulator Backend	PennyLane-Lightning	0.42.0
Data Handling	NumPy, Pandas	2.2.6, 2.3.3
Visualization	Matplotlib, Seaborn	3.10.6, 0.13.2

Patient-side QFF feature extraction and transmission

Acquisition of raw EEG signals occur at patient-side which are preprocessed through noise removal, normalization, and dimensionality reduction using Principal Component Analysis (PCA). This ensures that the input data is clean, compact, and suitable for advanced data analysis. Next, quantum feature extraction is performed, where the Quantum Wavelet Transform (QWT) captures localized time–frequency features, and the Quantum Fourier Transform (QFT) extracts global spectral features. These complementary feature sets are concatenated into a unified feature vector, , representing both temporal and frequency-domain characteristics of the EEG signals. Finally, the processed feature set is transmitted from the patient side to the doctor side using a lightweight MQTT protocol over a 5G gNB operating in uRLLC mode. The MQTT publisher on the patient device sends to the MQTT broker, which forwards it securely and with minimal latency to the doctor-side MQTT subscriber for further diagnostic processing.

The MQTT enable lightweight communication layer with efficient, low-latency, and reliable transmission of processed EEG feature vectors () between the patient-side device and the remote diagnostic server. Operating over a 5G gNB in uRLLC mode, MQTT’s publish–broker–subscribe mechanism minimizes network overhead and ensures near real-time data delivery. This design supports continuous monitoring and rapid diagnosis within the proposed QFF-ML Net framework.

Algorithm 1.

QFF feature extraction and transmission.

The EEG feature matrix, preprocessed from the raw data, was too large to be directly mapped onto the quantum circuit, as the number of qubits imposes a dimensional constraint. To address this, PCA was used to reduce the dimensionality corresponding to qubits, ensuring compatibility with the quantum feature space. PCA was prioritized over other dimensionality reduction techniques because it efficiently preserves the maximum variance of the data while providing orthogonal and linearly independent components. This property is advantageous for quantum transformations such as the Quantum Fourier Transform (QFT) and Quantum Wavelet Transform (QWT), as it avoids feature redundancy and stabilizes quantum encoding performance. Additionally, PCA’s deterministic and computationally efficient nature makes it suitable for preprocessing large EEG datasets.

Dataset description

The proposed QFF-MLNet remote telemedicine model tested with publicly available Bonn EEG dataset³⁰, and CHB-MIT Scalp EEG Dataset³¹,commonly used for epileptic and seizure detection. Bonn EEG³⁰ data consists of five classes (A, B, C, D, and E), each containing 100 EEG recordings. Each recording is a single-channel EEG signal with 4096 samples, stored in a text file and sampled at 173.61 Hz (Fig. 3).

Fig. 3.

EEG signal comparison on BONN Data normal vs abnormal brain states.

Sets A and B eye open and close healthy individuals. Sets C and D include interictal recordings—Set C from brain regions outside the epileptogenic zone and Set D from within the epileptogenic zone. Set E contains ictal (seizure) recordings. This dataset is widely used for benchmarking seizure detection algorithms, time-series classification methods, and analyzing neurophysiological patterns (Fig. 4).

Fig. 4.

CHB-MIT Scalp EEG Database.

The another datasets is CHB-MIT Scalp EEG Database³¹ it contains multichannel EEG recordings from scalp electrodes. Each record includes voltage differences between electrode positions (e.g., FP1-F7, C3-P3, C4-P4, CZ-PZ, F3-C3, F4-C4, F7-T7, F8-T8, P3-O1, P4-O2, T7-P7, etc.), with values represented as small floating-point numbers in the ±0.0000X range, indicating normalized or preprocessed microvolt-level signals. The target variable, Outcome, denotes seizure presence (1.0) or absence (0.0). The dataset is perfectly balanced, comprising 1,048,575 samples for each class, totaling over two million rows. Each row represents a timestamped EEG segment across all channels, enabling analysis of spatial and temporal brain activity patterns during seizure and non-seizure periods.

Data preprocessing

The EEG data collected from the patients were preprocessed using denoising, filtering, and normalization techniques, followed by dimensionality reduction through Principal Component Analysis (PCA) to extract the most informative features on both the datasets,

1

Figure 5 shows normal EEG signals with moderate fluctuations (Z, O, N) and abnormal EEG signals where seizure (S) exhibits sharp high-amplitude spikes compared to stable focal interictal (F) (Figs. 6, 7).

Fig. 5.

Preprocessed EEG Samples.

Fig. 6.

Preprocessed EEG samples comparison of normal vs abnormal.

Fig. 7.

Cumulative explained variance comparison.

The EEG feature of length requires 12 qubits for direct quantum integration, which is computationally expensive on classical simulators. To overcome this, the data are first normalized and then reduced in dimension using PCA. The feature vectors are padded from (500, 4096) to (500, 128), while the class labels remain unchanged. Finally, the concatenated dataset is constructed and used for supervised classification.

Quantum fourier transform on EEG signal

The EEG data vectors have length , applying a Quantum Fourier Transform (QFT) directly would require 12 qubits, which is computationally expensive on classical simulators. To address this, the EEG signals are first normalized and then reduced in dimension using PCA. After PCA, the dimension is 100, it can be zero-padded to the next power of two (128) to match the QFT input size,PCA components can be chosen directly as 128. This enables efficient application of the QFT with (since ) (Fig. 9).

Fig. 9.

QFT on EEG data comparison of multi class vs binary class.

Refer Fig. 10, each input sample consists of 128 features, which corresponds to a -dimensional vector. Therefore, the Quantum Fourier Transform (QFT) can be applied using a 7-qubit quantum circuit. Each sample is treated as an amplitude-encoded quantum state and normalized accordingly. The normalized vector is loaded into the quantum circuit using the QubitStateVector operation. A standard QFT is then applied over the 7 qubits, and the output can be extracted either as the final quantum state or via expectation value measurements.A classical EEG signal vector can be encoded into a quantum state via amplitude encoding, and then transformed using the Quantum Wavelet Transform (QWT). The QWT redistributes the amplitudes into localized time–frequency coefficients, capturing both temporal and spectral characteristics of non-stationary EEG signals in a quantum superposition. This enables parallel extraction of multi-scale features for efficient classification.EEG Signal Representation and Amplitude Encoding Let the sampled EEG signal be

2

3

where is the sampled EEG value at time index , and is chosen as a power of two to match the size of the quantum register.

Fig. 10.

Quantum circuit of QFT with based on Q-bits.

The EEG signal is normalized and encoded into a quantum state via amplitude encoding:

4

Here, is the Euclidean norm ensuring the quantum state is properly normalized. Quantum Fourier Transform (QFT) The Quantum Fourier Transform is defined as

5

Applying the QFT to the amplitude-encoded EEG signal :

6

The amplitudes after the QFT correspond to the discrete Fourier transform coefficients , which capture the global frequency features of the EEG signal.

Quantum wavelet transform on EEG signal

The role of QWT in EEG data is to provide a quantum-enhanced, multi-scale, time–frequency feature extraction method that captures transient and oscillatory patterns efficiently, making it particularly valuable for seizure detection, mental state analysis, and telemedicine applications. Custom PCA components is chosen as 128 to match with Q-bit. This enables efficient application of the QWT with (since ), Refer ?? , circuit applies the Haar wavelet transform on a quantum state by recursively using Hadamard gates on even qubits and CNOT gates between even-odd qubit pairs.It decomposes the input state into multiscale wavelet coefficients, capturing both coarse and fine features.The QWT efficiently encodes hierarchical information in the quantum state, useful for quantum signal processing and feature extraction (Figs. 11, 12).

7

Fig. 11.

QWT on EEG data comparison of multi class vs binary class.

Fig. 12.

Quantum circuit of QWT with based on Q-bits.

The classical discrete wavelet transform projects a signal onto wavelet basis functions:

8

where is the scale index and is the translation index. In the quantum domain, the QWT maps the quantum EEG state into a superposition of wavelet coefficients:

9

These coefficients encode localized time–frequency information, allowing the capture of transient EEG features.

10

where is a unitary matrix representing the discrete wavelet transform.

The transformed signal is then amplitude encoded as

11

with normalization factor

12

Doctor-side model training and diagnosis

On the doctor side, the received feature set is split into training and testing subsets for model development. Preprocessing steps include feature scaling and class balancing using SMOTE, followed by the training of four base models: Random Forest (RF), Gradient Boosting (GB), Logistic Regression (LR), and Support Vector Classifier (SVC). Each model undergoes hyperparameter tuning through a pipeline with RandomizedSearchCV, optimized using the ROC–AUC metric. The tuned models are then combined into a soft-voting ensemble, which is trained on the preprocessed data.QFF-ML Net is evaluated on the test set using accuracy, confusion matrix, classification metrics, ROC–AUC, and ROC curve visualization, thereby enabling reliable remote diagnostic decision-making.

Algorithm 2.

Novel QFF-ML net for remote diagnosis.

The proposed QFF-ML Net is a quantum-inspired machine learning framework designed for remote medical diagnosis refer algorithm2, At the patient side, raw biosignals are first preprocessed through noise filtering and dimensionality reduction using PCA. Quantum feature extraction is then applied, where QFT captures global spectral patterns while the QWT extracts localized time frequency features.

These features are fused through an attention-based mechanism to form a compact representation , which is transmitted to the doctor side, reducing communication overhead compared to raw data transfer. On the doctor side, machine learning pipelines are constructed with base models such as RF, GB, LR, and SVC. custom Hyperparameter optimization is performed using randomized search and the best-performing models are integrated into a soft-voting ensemble . The ensemble leverages the fused quantum features to produce the final diagnostic decision , which is evaluated in doctor side. By combining QFT and QWT features, QFF-ML Net captures both global and local signal characteristics, ensuring robustness against non-stationary biosignals like EEG. And the model integrated with 5G wireless communication to make real-time efficient design make in telemedicine applications in healthcare environments.

The QFF-ML Net achieved 92.6% accuracy (Fig. 13) with balanced performance across classes. For Class 0, precision, recall, and F1-score were 0.95, 0.92, and 0.93, while Class 1 achieved 0.88, 0.94, and 0.91, respectively refer Fig. 13a. Both macro and weighted averages were 0.92, confirming robustness. The ROC curve (Fig. 8b) shows an AUC of 0.952.

Fig. 13.

Performance analysis of the QFF-ML Net model in 5G assisted telemedicine systems.

Fig. 8.

2D PCA Projection of Normal vs Abnormal EEG Signals.

Refer Table4, which shows the tuned hyperparameter ranges for the base learners in the QFF-ML Net. Each model—RF, GB, LR, and SVC—was optimized using randomized search with cross-validation to balance model complexity and generalization. The selected parameters enhance learning efficiency and ensure robust ensemble performance, making the model suitable for implementation on edge devices.

Table 4.

Hyperparameter settings of base learners and their effectiveness in QFF-ML Net.

Model	Hyperparameter	Effectiveness / Purpose
RF	Number of Estimators ()	100–500; reduces variance and stabilizes predictions
	Maximum Depth ()	10–20; prevents overfitting while capturing feature interactions
	Minimum Samples Split ()	2–10; ensures smooth tree growth and better generalization
GB	Number of Estimators ()	100–200; balances learning speed and model stability
	Learning Rate ()	0.05–0.2; small rates improve robustness to noise
	Maximum Depth ()	3–7; controls complexity and ensures efficient computation
LR	Regularization Strength (C)	0.01–10; controls overfitting by penalizing large coefficients
	Penalty	L2; stabilizes convergence
	Solver	lbfgs; ensures numerical stability and efficient optimization
SVC	Regularization Parameter (C)	0.1–10; balances margin softness for noisy data
	Kernel Type	rbf, linear; supports both linear and nonlinear decision boundaries
	Gamma	scale, auto; adjusts influence of individual training points

Remote telemedicine system

EEG signals are assumed to be collected using an EEG headset or biosensor and are either stored in the above formats or streamed in real-time via a buffer. The raw EEG data from UE (sensor node) cannot be directly transmitted over a 5G wireless network it stored in different formats such as .mat, .csv, or .txt, Before transmission, the data must pass through encoding, compression, and packetization.The receiver (Node 1) doctor successfully receives concatenated quantum features it transmitted via MQTT over 5G uRLLC link, validating reliable end-to-end data delivery for remote telemedicine and diagnosis refer Fig. 14. To reduce data redundancy and bandwidth usage, a quantum-inspired EEG processing pipeline is applied using QWT and QFT for compression and data transmission (Fig. 15).

Fig. 14.

Transmission and reception of concatenated QWT and QFT feature files in NS-3.44 simulation.

Fig. 15.

The network topology visualization in NetAnim.

In the NS-3.44 simulation using the CTTC LENA-NR module³², concatenated feature files are transmitted from the patient UE and successfully received at the doctor UE over uRLLC 5G link. 5G NR band n78 carrier frequency () used in the simulation setup the range of 3300–3800 MHz (FR1, TDD). The central frequency of 3.5 GHz 5G deployments,Ns3 simulation setup given in Table 5.

Table 5.

Key NS-3.44 simulation setup and URLLC performance metrics for 5G-MQTT based telemedicine.

Parameter	Configuration	Value / Result
Simulator	Tool	NS-3.44 with CTTC LENA-NR
Operating Band		n78 @ 3.5 GHz
System Bandwidth	FR1	50 MHz
Network Topology	UEs + gNB	2 Patients + 1 Doctor + 1 gNB
QoS Profile	Service Type	uRLLC (3GPP)
Packets Sent	Count	4000
Packets Received	Count	3999
Packet Loss Rate	Percentage	0.025%
Average Jitter	Delay	1 ms
Throughput	Data Rate	0.64 Mbps

The MQTT protocol enable efficient and reliable transmission of the extracted QFF feature set between the patient and the remote doctor in a 5G-enabled environment.

Its lightweight publish–subscribe mechanism significantly reduces communication overhead and latency, making it well-suited for real-time healthcare applications. In this framework, the patient device functions as an MQTT publisher that transmits the preprocessed and quantum-fused features () to an MQTT broker over the 5G network. The broker then relays this data to the remote doctor’s system acting as an MQTT subscriber. By leveraging 5G’s ultra-reliable low-latency communication (uRLLC) and MQTT’s asynchronous delivery, the framework ensures minimal delay, secure data flow, and consistent QoS during feature transmission.

The QWT extracts time-frequency components of the EEG signal, after enabling denoising ,filtering and quantuam compression by reducing signal redundancy and noise.The QFT captures frequency domain features, helping to compress periodic patterns in the signal. The outputs of both QWT and QFT are concatenated, and the resulting compressed data is then transmitted over the 5G wireless network using the MQTT protocol for efficient low-latency delivery.

Refer Fig. 16 shows the real-time remote monitoring by proposed model over a 5G uRLLC-enabled network, the QFF-ML Net model achieved an accuracy of 88.25%, with precision, recall, and F1-scores of 0.93, 0.88, and 0.91 for the Normal class, and 0.81, 0.88, and 0.84 for the Abnormal class, respectively. The ROC curve achieve AUC of 0.949, demonstrating strong discriminative ability. These results highlight the effectiveness and reliability of QFF-ML Net in real-time telemedicine applications that assisted uRLLC enabled 5G Networks .

Fig. 16.

Real-time remote diagnosis on EEG signal.

Result and discussion

The proposed QFF-ML Net evaluated on benchmark EEG datasets (Bonn and CHB-MIT) and real-time transmission over a 5G uRLLC network. On the Bonn dataset, it achieved 92.3% accuracy (precision 0.92, recall 0.93, F1-score 0.92), and on CHB-MIT, 91.0% accuracy (precision, recall, F1-score all 0.91), outperforming conventional classifiers such as RF, SVC,GB, and LR. For real-time monitoring (Fig. 16), the model achieved 88.25% accuracy, with the Normal class showing precision 0.93, recall 0.88, F1-score 0.91, and the Abnormal class precision 0.81, recall 0.88, F1-score 0.84. An AUC of 0.949, confirming strong discriminative ability and reliability for remote EEG classification in telemedicine applications (Fig. 17, Tables 6, 7).

Fig. 17.

Performance analysis of the QFF-ML Net model in CHB-MIT Scalp EEG Database.

Table 6.

Performance comparison of QFF-ML Net on Bonn, CHB-MIT EEG datasets.

Model	Dataset	Accuracy	Precision	Recall	F1-score
QFF-ML Net (Proposed)	Bonn	0.923	0.92	0.93	0.92
QFF-ML Net (Proposed)	CHB-MIT	0.910	0.91	0.91	0.91
QFF-ML Net (Proposed)	Remote 5G data	0.8825	0.93	0.88	0.91
Random Forest	Bonn	0.840	0.85	0.82	0.83
Support Vector Classifier (SVC)	Bonn	0.825	0.85	0.79	0.80
Gradient Boosting	Bonn	0.839	0.84	0.82	0.83
Logistic Regression	Bonn	0.554	0.54	0.54	0.54

Table 7.

5-fold cross-validation performance of the proposed QFF-ML Net on EEG dataset.

Fold	Accuracy	AUC
1	0.9122	0.9354
2	0.9270	0.9550
3	0.9261	0.9576
4	0.9326	0.9589
5	0.9213	0.9546
Mean ± Std	0.9238 ± 0.0068	0.9523 ± 0.0086

The proposed QFF-ML Net is validated using 5-fold cross-validation.The model achieved a mean accuracy of 92.38% () and a mean AUC of 95.23% () indicating high classification reliability. The low standard deviation across folds confirms consistent generalization.

Statistical significance and time complexity analysis

This section evaluates the proposed QFF-ML Net in terms of both statistical significance and computational efficiency. The statistical significance matrix (Fig. 18a) indicates that QFF-ML Net achieves a statistically significant improvement over most classical models (p < 0.05) such as Logistic Regression, SVM, Decision Tree, KNN, Naïve Bayes, and AdaBoost. No significant difference is observed when compared with Random Forest and Gradient Boosting, suggesting comparable performance to these ensemble methods.

Fig. 18.

Comprehensive evaluation of the proposed QFF-ML Net compared to classical machine learning models: (a) Statistical significance across models (p-values), (b) Training time comparison, and (c) Accuracy–time trade-off visualization.

As shown in Fig.18b, QFF-ML Net requires slightly higher computational time (30.63 s) due to its hybrid quantum-classical architecture, yet provides robust and consistent accuracy. The scatter plot in Fig. 18c further demonstrates that while QFF-ML Net incurs higher training cost, it delivers the best accuracy among all models, confirming its strong generalization capability. Overall, the analysis establishes that QFF-ML Net achieves a favorable balance between accuracy and computational efficiency, offering statistically validated superiority over conventional classifiers.

Comparative analysis

The experimental results on the Bonn dataset show that ensemble models such as Random Forest (Accuracy: 0.840, F1: 0.83) and Gradient Boosting (Accuracy: 0.839, F1: 0.83) achieved the best overall performance, followed by SVC (Accuracy: 0.825, F1: 0.80). In contrast, Logistic Regression performed significantly worse, with an accuracy and F1-score of only 0.554, highlighting the limitations of classical models in remote telemedicine applications refer Fig. 20.The grouped bar chart compares QFF-ML Net with classical ML models across Accuracy, Precision, Recall, and F1-score. QFF-ML Net bars are highlighted in bright colors, showing consistently higher performance, especially on real-time 5G data, demonstrating its superiority over traditional classifiers (Fig. 19a).

Fig. 20.

Comparative performance of QFF-ML Net against classical machine learning models in 5G-assisted telemedicine.

Fig. 19.

Comparison: (a) performance metrics and (b) ROC curves of QFF-ML Net against classical ML models.

The comparative analysis highlights that recent EEG-based models employ diverse architectures and feature extraction techniques. Shoka et al.⁵ achieved moderate accuracy with CNN on CHB-MIT, while Berrich et al. ⁴ improved performance using CNN/DNN with SVM on Bonn EEG. Huang et al. ⁷ reported strong results with STFFDA on both Bonn and CHB-MIT datasets. Xiaoshuai et al. ⁹ further enhanced classification using CNN-Bi-LSTM with DWT features, attaining the highest accuracy. In contrast, the proposed QFF-ML Net demonstrates consistently robust performance across Bonn, CHB-MIT, and real-time remote EEG data, and uniquely supports remote diagnosis, thereby addressing practical deployment in telemedicine (Table 8).

Table 8.

Comparison of recent EEG-based models, feature fusion techniques, and real-time remote diagnosis.

Author/Year	Model	Dataset	Acc	P	R	F1	Feature	Remote Diagnosis
Shoka et al. 5, 2023	CNN	CHB-MIT	0.791	0.733	0.916	N/A	Encrypted	Yes
Berrich et al. 4, 2025	CNN-SVM DNN-SVM	Bonn EEG	0.960	N/A	N/A	N/A	PCA	No
Huang et al. 7, 2025	STFFDA	Bonn EEG CHB-MIT	0.951	0.952	0.952	0.952	N/A	No
Xiaoshuai et al. 9, 2025	CNN-Bi-LSTM	Bonn EEG CHB-MIT	0.962 0.984	0.975 0.991	0.951 0.978	0.962 0.984	DWT	No
This Work (2025)	QFF-ML Net	Bonn EEG CHB-MIT Remote EEG	0.923 0.910 0.883	0.92 0.91 0.93	0.93 0.91 0.88	0.92 0.91 0.91	QFF	Yes

The training curves indicate that both CNN and LSTM models progressively improve their accuracy across epochs, with the CNN achieving slightly faster convergence and lower overall loss compared to the standalone LSTM. The hybrid CNN–LSTM model, however, demonstrates a more balanced behavior by achieving higher and more stable validation accuracy, while maintaining a minimal gap between training and validation loss. This indicates strong generalization capability and reduced risk of overfitting. The complementary strengths of CNN’s spatial feature extraction and LSTM’s temporal pattern learning contribute to improved learning efficiency and predictive robustness (Figs. 21, 22).

Fig. 21.

Training and validation accuracy and loss comparison of CNN and LSTM models.

Fig. 22.

Training and validation accuracy and loss of the hybrid CNN–LSTM model.

Although these deep learning architectures deliver strong performance, they typically demand significant computational resources, memory footprint, and continuous parameter optimization, making them less suitable for deployment at the network edge. To address these limitations, the proposed QFF-ML Net integrates lightweight quantum-inspired feature fusion, enabling efficient inference using compressed discriminative features. This substantially reduces model complexity, energy consumption, and latency, while preserving competitive accuracy. As a result, QFF-ML Net is more suitable for real-time, edge-aware environments where hardware constraints, energy efficiency, and rapid response are critical (Table 9).

Table 9.

Performance metrics with 95% confidence intervals for all models.

Model	Accuracy	95% CI (Accuracy)	F1-Score	ROC-AUC
Logistic Regression	0.9257 ± 0.0069	[0.9155, 0.9351]	0.9266 ± 0.0074	0.9510 ± 0.0066
Random Forest	0.9319 ± 0.0055	[0.9231, 0.9383]	0.9331 ± 0.0059	0.9629 ± 0.0085
Gradient Boosting	0.9324 ± 0.0055	[0.9231, 0.9368]	0.9334 ± 0.0057	0.9629 ± 0.0068
SVM	0.9236 ± 0.0070	[0.9123, 0.9304]	0.9248 ± 0.0079	0.9543 ± 0.0086
Decision Tree	0.9168 ± 0.0128	[0.9003, 0.9349]	0.9172 ± 0.0136	0.9053 ± 0.0153
KNN	0.8547 ± 0.0201	[0.8341, 0.8841]	0.8625 ± 0.0205	0.9168 ± 0.0205
Naïve Bayes	0.9233 ± 0.0070	[0.9141, 0.9336]	0.9241 ± 0.0076	0.9522 ± 0.0074
AdaBoost	0.9192 ± 0.0053	[0.9124, 0.9261]	0.9213 ± 0.0056	0.9588 ± 0.0070
QFF-ML Net	0.9319 ± 0.0062	[0.9216, 0.9386]	0.9330 ± 0.0067	0.9606 ± 0.0073

The proposed QFF-ML Net demonstrates statistically significant performance improvements with compared to most baseline models. The model also exhibits narrow 95% confidence intervals across all evaluation metrics, confirming its robustness, stability, and reliability in EEG-based classification under 5G-assisted telemedicine environments.

Conclusion

The proposed QFF-ML Net model performs better across both benchmark and real-time EEG datasets. On the Bonn dataset, it achieved an accuracy of 92.3% highlighting its strong discriminative and On the large-scale CHB-MIT dataset, the model maintained robust generalization with 91% accuracy and balanced precision, recall, and F1-score of 0.91 each. Furthermore it deployed for real-time remote monitoring over a 5G uRLLC-enabled network, QFF-ML Net achieved reliable performance while requiring low computational resources, demonstrating its suitability for practical, scalable telemedicine applications. These results confirm that QFF-ML Net is both effective and efficient for real-world EEG analysis of epilepsy seizure detection and remote healthcare monitoring.

• Bonn EEG Dataset: Available for direct download from the Universitat Pompeu Fabra (UPF) repository via the ”Bonn EEG time series download page”³³. This dataset comprises five subsets (A–E), each containing 100 single-channel EEG recordings (ASCII text, 4,096 samples per file, sampled at 173.61 Hz)^33,34.

• Preprocessed CHB-MIT Scalp EEG Database: Accessible via IEEE Dataport³¹.

Author contributions

These authors contributed equally to this work

Funding

Open access funding provided by Vellore Institute of Technology. The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The datasets analyzed in this study are publicly available. The Bonn EEG dataset can be directly downloaded from the Universitat Pompeu Fabra (UPF) repository via the Bonn EEG time series download page³⁴. It consists of five subsets (A–E), each containing 100 single-channel EEG recordings in ASCII text format, with 4,096 samples per file at a sampling rate of 173.61 Hz³⁴. In addition, the preprocessed CHB-MIT Scalp EEG database is accessible through IEEE Dataport³¹.

Declarations

Competing interests

The authors declare no competing interests.