Leveraging Large Language Models for Climate Modeling in High-Performance Computing
Dr. Hoda El-Sayed, Department of Computer Science, Bowie State University (MDREN Participant).
Fahmina Nur Salma, Bowie State University.
This paper explores the potential of integrating Large Language Models (LLMs) into High-Performance Computing (HPC) systems, specifically focusing on climate modeling as an application area. By leveraging LLMs’ capabilities in natural language processing, knowledge representation, and problem-solving, we aim to enhance the efficiency and effectiveness of climate related simulations and data analysis within HPC environments. Our proposed framework involves identifying suitable LLM architectures for the domain, designing an HPC architecture that seamlessly incorporates LLMs, and developing techniques to optimize LLM performance within HPC workflows. Additionally, we implement a statistical optimization framework to fine-tune LLM parameters, maximizing resource efficiency and predictive accuracy.
Exploring the efficacy of Vision Transformer models with Explainable AI in early detection of Dementia
Dr. Hoda El-Sayed, Department of Computer Science, Bowie State University (MDREN Participant).
Tolulope Oshuntoye, Bowie State University.
This research investigates the potential of Vision Transformer (ViT) models enhanced with Explainable AI (XAI) techniques to improve the early detection and classification of dementia using medical imaging data, specifically MRI and PET scans. While traditional convolutional neural networks (CNNs) have been widely utilized in medical image analysis, ViTs offer a promising alternative due to their ability to capture global context and long-range dependencies in images. However, the complexity of transformer models often leads to a “black box” effect, where the decision-making process is opaque. To address this, the research will integrate XAI methods, aiming to make the model’s predictions more interpretable and actionable for clinicians. The study will evaluate whether ViT models, combined with XAI techniques, can outperform CNNs in detecting early signs of dementia, while also providing meaningful insights into the underlying decision processes.
An LLM-Driven End-to-End Framework for Robust Speaker Diarization in Dynamic Environments
Godswill Melford
Dr. Hoda El-Sayed, Department of Computer Science, Bowie State University (MDREN Participant).
This work proposes an LLM‑driven, end‑to‑end framework for robust speaker diarization in dynamic conversational environments, grounded in insights from recent advances in multimodal and speech‑centric large language models. Traditional diarization and recognition pipelines rely on cascaded modules that suffer from error propagation, limited handling of overlapping speech, and weak integration across tasks. The first reviewed system, SpeakerLM, unifies diarization, speaker recognition, and ASR within a single multimodal LLM architecture capable of reasoning jointly over acoustic and linguistic cues. Through large‑scale pretraining, multi‑speaker simulation, and fine‑tuning on diverse real‑world datasets, SpeakerLM achieves state‑of‑the‑art performance in both transcription accuracy and speaker attribution across challenging conditions. The second reviewed approach introduces a unified Speech LLM for multilingual diarization and recognition, reformulating the task with structured prompts and sliding‑window audio inputs to eliminate dependencies on VAD and clustering. Despite using a smaller backbone, this model delivers substantial improvements in tcpWER, tcpCER, and diarization error rate, demonstrating strong multilingual and multi‑speaker reasoning. Together, these studies highlight the emerging potential of end‑to‑end LLM‑based frameworks to surpass traditional pipelines, offering more accurate, scalable, and context‑aware speaker diarization in real conversational settings.
Machine Learning Using Fully Homomorphic Encryption for Medical Health records
Dr. Hoda El-Sayed, Department of Computer Science, Bowie State University (MDREN Participant).
Integrating machine learning (ML) into healthcare systems has revolutionized
diagnostic capabilities and personalized treatment plans. However, using sensitive
medical data introduces significant privacy concerns, necessitating secure computational
frameworks. Privacy-preserving machine learning (PPML) addresses these challenges by
enabling data processing without exposing sensitive information. This study explores the
application of Fully Homomorphic Encryption (FHE) with machine learning models for
classifying medical dialogues containing potentially sensitive personally identifiable
information (PII). Benchmarking FHE-based inference models against the traditional
plaintext approaches. This study demonstrates that FHE can achieve comparable ensuring
robust data privacy. However, enhanced security comes at the cost of increased
computational time and memory usage. The study highlights a critical trade-off between
maintaining high-speed inference and preserving privacy, providing valuable insights into
medical data. These results guide decision-makers in evaluating the deployment of secure,
FHE-based solutions in production environments. Overall, this study explores the
integration of privacy-preserving techniques with machine learning and outlines future
research directions.
Deep Learning-Based Automatic Delineation of P and T Waves in Electrocardiogram Signals Using Multi-Stage Architecture with Attention Mechanisms
Dr. Hoda El-Sayed, Department of Computer Science, Bowie State University (MDREN Participant).
The accurate and automated identification of P and T waves in electrocardiogram (ECG) signals remains a critical challenge in cardiovascular diagnostic technologies. Traditional signal processing techniques, including wavelet transforms and nonlinear filtering, struggle with variability in ECG signal morphology, noise contamination, and limited generalizability across diverse patient populations. While recent deep learning approaches have shown promise, they often fail to adequately address the subtle characteristics of P and T waves compared to the dominant QRS complex, resulting in suboptimal detection rates and high false positive/negative rates. This study proposes a novel multi-stage deep learning framework that integrates adaptive preprocessing, dual-path signal processing, and attention-based neural networks for robust P and T wave delineation. The proposed architecture consists of four key components: (1) an adaptive preprocessing module that learns patient-specific noise characteristics and signal enhancement strategies, (2) a dual-path processing system that simultaneously analyzes raw and enhanced signals to preserve critical wave features, (3) a multi-scale convolutional neural network with self-attention mechanisms specifically designed to focus on subtle P and T wave characteristics, and (4) a transformer encoder that captures long range temporal dependencies between cardiac events. Additionally, the framework incorporates uncertainty quantification to provide confidence scores for each detection, enabling clinicians to identify cases requiring manual review. The methodology is evaluated using the MIT-BIH Arrhythmia Database and compared against state-of-the-art approaches including wavelet-based methods, Gaussian mixture models, and existing deep learning architectures. The proposed approach addresses fundamental limitations in current methods by adapting to individual signal characteristics, handling various noise types, and maintaining high accuracy across different patient populations and cardiac conditions. Preliminary theoretical analysis suggests potential improvements of 2-5% in P and T wave detection accuracy compared to existing methods, with particular strength in noisy conditions and abnormal morphologies. This research contributes to the field of automated ECG analysis by providing a comprehensive framework that balances accuracy, computational efficiency, and clinical interpretability. The resulting system has potential applications in continuous cardiac monitoring, early arrhythmia detection, and automated ECG screening in resource-limited settings, ultimately supporting improved cardiovascular care and patient outcomes.