Mumtahina Ahmed

Mumtahina Ahmed

Auto-generated profile, based on publicly available information

Mumtahina Ahmed

Department of CSE, Bangladesh University of Business and Technology, Dhaka, Bangladesh

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
25 October 2025 in Array
Verified
Journal Article
Open access
AKAzam Khan
Azam Khan
ARArifur Rahman
Arifur Rahman
FMFarhad Uddin Mahmud
Farhad Uddin Mahmud
Kanchon Kumar BishnuKanchon Kumar Bishnu
Kanchon Kumar Bishnu
MAMumtahina Ahmed
Mumtahina Ahmed
Zeyar AungZeyar Aung
Zeyar Aung

27 November 2025 in IEEE Open Journal of the Computer Society
Verified
Journal Article
Open access
MNMasuduzzaman Niloy
Masuduzzaman Niloy
TITanzimul Islam
Tanzimul Islam
SUShafiq Ullah
Shafiq Ullah
JAJobayar Alom
Jobayar Alom
MAMumtahina Ahmed
Mumtahina Ahmed
JHJakir Hossen
Jakir Hossen
Accurate and generalizable ECG classification remains a critical challenge in automated cardiac diagnostics, particularly when dealing with multi-lead signals from heterogeneous sources. In this paper, we propose MuRe-LAT, a Multi-Resolution Lead-Aware Transformer architecture with domain-adaptive normalization designed for robust arrhythmia detection. The model integrates a convolutional stem for local feature extraction, a lead attention module to adaptively weigh channel contributions, dual-scale transformer encoders for capturing both short-term and long-term dependencies, and a domain-specific batch normalization mechanism to mitigate dataset shift. To enhance generalization, MuRe-LAT is pretrained using a masked signal modeling objective. We evaluate the model on two benchmark datasets MIT-BIH Arrhythmia and PTB-XL spanning both ambulatory and clinical ECG recordings. On MIT-BIH, MuRe-LAT achieves an accuracy of 94.0%, a macro-F1 of 88.4, an AUROC of 96.2, and a PR-AUC of 91.6; on PTB-XL, it reaches 88.7% accuracy, 78.1 macro-F1, and 90.5 AUROC. The model also demonstrates strong cross-domain generalization, achieving up to 91.7% accuracy when transferred from PTB-XL to MIT-BIH without retraining. Compared to strong baselines including ResNet-1D, BiLSTM, ECG-BERT, and CardioFormer, MuRe-LAT outperforms or closely matches state-of-the-art results in over 90% of evaluation settings. With only 12.3M parameters and an inference time of 8.4 ms per 10-second ECG window, MuRe-LAT offers a compelling balance between accuracy and efficiency, making it suitable for deployment in real-time clinical and wearable applications.

2 February 2026 in PLOS ONE
Verified
Journal Article
Open access
MAMD AL AMIN
MD AL AMIN
MAMohammad Shafat Ahsan
Mohammad Shafat Ahsan
JMJannatul Maua
Jannatul Maua
MAMumtahina Ahmed
Mumtahina Ahmed
Kamruddin NurKamruddin Nur
Kamruddin Nur
This paper proposes a novel Physics-Informed Self-Supervised Diagnosis (PI-SSD) framework for rotating machinery fault detection, combining physical modeling, self-supervised representation learning, and uncertainty-aware classification. The architecture integrates a multi-resolution convolutional encoder, a windowed Transformer for temporal context modeling, and a latent neural ordinary differential equation (ODE) module that embeds mechanical priors, such as Jeffcott rotor dynamics, directly into the learning process. A masked segment reconstruction objective enables self-supervised pretraining using unlabeled healthy signals, while an evidential classifier head produces fault probabilities with calibrated uncertainty. We evaluate PI-SSD on two publicly available datasets, the NASA PHM’09 Gearbox dataset and the Aalto Rotor dataset, covering 6 fault types and over 5,500 multichannel vibration recordings. Compared to seven strong baselines, PI-SSD achieves the highest Macro-F1 score (0.91) and lowest Expected Calibration Error (ECE = 0.022) on the NASA dataset, while maintaining strong domain transfer performance on Aalto (Macro-F1 = 0.81, PR-MSE = 0.067) without fine-tuning. Ablation studies confirm the contribution of each component, and physics consistency analysis demonstrates low violation rates under varying speeds. These results highlight the potential of embedding physics knowledge into self-supervised neural systems for robust, interpretable, and transferable fault diagnosis in industrial applications.

19 March 2026 in IEEE Open Journal of the Computer Society
Verified
Journal Article
Open access
Jungpil SHINJungpil SHIN
Jungpil SHIN
JMJannatul Maua
Jannatul Maua
MNMir Nafiul Nagib
Mir Nafiul Nagib
AAAhsan Ahmed
Ahsan Ahmed
MAMumtahina Ahmed
Mumtahina Ahmed
MMM. F. Mridha
M. F. Mridha
Early and accurate diagnosis of Parkinson's Disease (PD) is essential for timely intervention and management. While prior studies have used unimodal signals like voice, gait, or handwriting separately, real-world screening requires integrating diverse biomedical data. This work proposes XAMT++, a Cross-Attention Multi-Modal Transformer that jointly processes voice (Oxford Telemonitoring), gait (Daphnet FoG), and handwriting (Spiral Drawings) data for improved PD diagnosis. XAMT++ includes (i) modality-specific encoders, (ii) a cross-attention fusion module, and (iii) a hierarchical transformer backbone. It is trained with a hybrid supervised and contrastive alignment loss to remain robust even when some modalities are missing. On the three public datasets, XAMT++ achieves 0.87 classification accuracy, 0.88 F1 score, 0.89 AUROC, and 0.90 AUPRC, outperforming unimodal and standard fusion baselines. Notably, it maintains strong early-stage detection performance, achieving 0.60 accuracy within 0–30 days post-diagnosis, with further improvements as the disease progresses. Inter-modal alignment metrics confirm high representation consistency, with a cosine similarity of 0.87 and a mutual information of 0.89. These results demonstrate XAMT++ as a practical and effective multi-modal diagnostic framework for early PD detection.

27 November 2025 in IEEE Open Journal of the Computer Society
Verified
Journal Article
Open access
MAMD AL AMIN
MD AL AMIN
MAMohammad Shafat Ahsan
Mohammad Shafat Ahsan
RIRahomotul Islam
Rahomotul Islam
MAMumtahina Ahmed
Mumtahina Ahmed
MMM. F. Mridha
M. F. Mridha
JHJakir Hossen
Jakir Hossen
Cloud-based systems have become prime targets for increasingly sophisticated cyberattacks due to their distributed, multi-tenant, and dynamic nature. Traditional intrusion detection systems often fail to generalize across diverse cloud environments, particularly when dealing with heterogeneous data sources such as network traffic and user behavior. In this paper, we propose a novel Dual-Encoder Adversarial Learning framework designed to detect cyber intrusions in cloud systems through a unified representation of multi-domain features. The architecture integrates private encoders for domain-specific signal extraction, a shared encoder for domain-invariant representation learning, and adversarial training implemented via a Gradient Reversal Layer to enforce cross-domain generalization. A reconstruction decoder is also employed to retain semantic fidelity in latent embeddings. Extensive experiments conducted on two benchmark datasets, UNSW-NB15 and the Cybersecurity Intrusion Detection dataset, demonstrate that the proposed model consistently outperforms state-of-the-art baselines in terms of accuracy (0.917), precision (0.903), recall (0.910), F1-score (0.906), and AUC (0.947). The model also exhibits strong robustness under noisy conditions and achieves competitive efficiency in training time and memory usage. These results validate the effectiveness of the proposed architecture for efficient and robust intrusion detection in cloud environments.

11 December 2025 in IEEE Open Journal of the Computer Society
Verified
Journal Article
Open access
RIRahomotul Islam
Rahomotul Islam
NNNafiza Humauara Hasan Neha
Nafiza Humauara Hasan Neha
MAMumtahina Ahmed
Mumtahina Ahmed
JMJannatul Maua
Jannatul Maua
NCNeha Chaudhary
Neha Chaudhary
SRSatyabrata Roy
Satyabrata Roy
Industrial systems require effective anomaly detection and predictive maintenance strategies to prevent unexpected failures and optimize operational efficiency. However, traditional approaches often treat anomaly detection and failure prediction as independent tasks, limiting their ability to leverage operational patterns for early fault diagnosis. This paper proposes a novel two-stage hybrid deep learning framework that integrates unsupervised anomaly detection with supervised failure prediction through cross-domain transfer learning. The framework first employs a variational autoencoder (VAE) to extract latent anomaly embeddings from HVAC system operational data without requiring labeled anomalies. These embeddings are transferred and combined with sensor readings from a predictive maintenance dataset to enrich the failure prediction model. Experimental evaluations on two publicly available datasets demonstrate that the proposed hybrid model achieves superior performance compared to conventional methods, with notable improvements in accuracy (90.2%), precision (0.89), recall (0.90), F1-score (0.90), and AUC-ROC (0.873). Additionally, the approach exhibits strong cross-domain generalization ability, highlighting its potential applicability across different industrial environments. This work introduces a scalable and flexible anomaly-driven transfer learning strategy that addresses the scarcity of labeled failure data and enhances the robustness of predictive maintenance systems. The findings suggest that integrating unsupervised operational insights into supervised learning frameworks is a promising direction for advancing intelligent industrial monitoring.