News & Events

Teams

Title: Modeling electronic health records: interpretable sequential approaches for enhanced multimodal embeddings and future clinical outcome prediction with time- aware irregular intervals handling

Abstract: With the rapid adoption of Electronic Health Records (EHRs), vast multimodal healthcare data o:er new opportunities for predictive modeling and clinical decision support. However, challenges such as irregular time intervals, missing values, and heterogeneous modalities necessitate specialized deep-learning approaches. To address these challenges in Alzheimer's disease (AD) progression prediction, we developed PPAD, a deep learning framework that integrates longitudinal and cross-sectional patient data. PPAD leverages recurrent neural networks (RNNs) for sequential data processing and a Multilayer Perceptron (MLP) for demographic integration, introducing an age-aware temporal fusion mechanism to enhance prediction accuracy. Building upon PPAD, we designed TA-RNN, an attention-based time-aware RNN that incorporates a time embedding layer and a dual-level attention mechanism. TA-RNN enhances predictive performance and interpretability in EHR modeling, demonstrating e:ectiveness in AD progression prediction and in-hospital mortality assessment. Expanding on these e:orts, we developed CAAT-EHR, a Transformer- based generative embedding model designed to optimize raw multimodal EHR data for downstream predictive tasks. By leveraging self- and cross-attention mechanisms, CAAT- EHR generates task-agnostic longitudinal embeddings, significantly outperforming models trained on raw EHR data for across multiple clinical prediction tasks, including AD progression modeling. The advancements in PPAD, TA-RNN, and CAAT-EHR demonstrate the potential of deep learning to enhance clinical outcome predictions, providing interpretable, time-aware, and transferable solutions for real-world healthcare applications.

Bio: Mohammad Al Olaimat is a Ph.D. candidate in Computer Science at the University of North Texas, set to graduate in Spring 2025. His research focuses on developing deep learning models for electronic health records (EHRs), leveraging multimodal data for disease progression modeling and outcome prediction. He has presented his work at top-tier conferences, including ISMB and AAIC, and has published in Bioinformatics. Mohammad was also a teaching fellow, instructing in a course on bioinformatics algorithms. His expertise spans computational biology, machine learning, and AI-driven healthcare solutions.

Teams

Title: AI in Healthcare and Aging Research

Abstract: Artificial intelligence (AI) profoundly impacts healthcare, enhancing our ability to manage complex diseases such as cancer, cardiovascular ailments, and knee osteoarthritis through advanced medical imaging and personalized treatment plans. This talk will showcase my contributions to medical imaging for knee osteoarthritis, precision oncology, and AI-enhanced stroke prediction. These projects are especially crucial as we age, increasing our vulnerability to diseases and reducing our quality of life—issues like mobility impairments from knee osteoarthritis are prime examples. I will discuss specific AI technologies I have implemented, including deep learning models for image analysis and machine learning algorithms for predicting disease progression. Additionally, I will explore the potential for emotional well-being technologies, such as facial expression recognition, to monitor and enhance the life quality of older adults. Looking forward, I am excited about the opportunity to collaborate with USF’s interdisciplinary teams to further develop technologies that support and enrich the lives of those facing the challenges of aging.

Bio: Dr. Rizwan Qureshi is a seasoned researcher in the field of Artificial Intelligence (AI) with a specialized focus on its applications in healthcare. He received his holds a PhD from the City University of Hong Kong, Hong Kong, in 2021, focusing on AI and Machine Learning, for lung cancer research. Dr. Qureshi has worked on se Dr. Qureshi's work integrates cutting- edge AI technologies, including deep learning and machine learning algorithms, to analyze complex diseases and develop personalized treatment protocols. His contributions to medical imaging are particularly significant in precision lung cancer, for drug response prediction, and addressing degenerative diseases like knee osteoarthritis. Dr. Qureshi has also worked on classical computer vision problems, such as, facial emotion recognition, and object detection.

Dr. Qureshi is also a passionate educator, eager to contribute to the academic community by teaching both undergraduate and graduate courses in computer science. His future aspirations include fostering collaborative research initiatives at the intersection of AI and healthcare, aiming to create innovative solutions that enhance care for vulnerable older adults and those with disabilities.

Teams

Title: Fine-Tuning Large Language Models for Genomic Intelligence.

Abstract: The intersection of artificial intelligence and genomics is transforming our ability to decode the functional impacts of genetic variations, yet several computational challenges persist, including scalability, disease specificity, and adaptability to heterogeneous genomic data. This presentation introduces two AI-driven frameworks, DYNA and LINGO, which leverage machine learning and natural language processing (NLP) innovations to address these limitations in genomic analysis. DYNA leverages a Siamese neural network to fine- tune genomic foundation models for precise, disease-specific predictions, treating the genome as a structured language where genetic variants function like syntax or grammar rules. It adapts pre-trained embeddings to domain-specific tasks, achieving state-of-the-art performance in variant eFect prediction and generalizing to unseen variants. This demonstrates how task-specific fine-tuning transforms pre-trained models for specialized biomedical challenges, similar to NLP applications like semantic similarity or translation. LINGO reimagines genomic data processing by extending NLP foundation models to genomic sequences. Leveraging a novel prefix-based tokenization strategy and adaptive rank sampling for parameter-eFicient fine-tuning, LINGO achieves unparalleled scalability and performance across genome annotation tasks. Together, DYNA and LINGO illustrate how AI and machine learning methodologies can redefine computational genomics. These frameworks oFer a glimpse into the future of AI as a transformative tool in biomedicine, bridging machine learning with precision genomics. The advancements also pave the way for secure and privacy-preserving genomic models, integrating federated learning, homomorphic encryption, and adversarial robustness to enable collaborative analysis without compromising sensitive patient data. This ensures trustworthy and resilient AI- driven genomic intelligence, bridging innovation with ethical deployment in precision medicine.

Bio: Huixin Zhan, Ph.D., is a postdoctoral scientist in the Department of Computational Biomedicine at Cedars-Sinai Medical Center. She earned her Ph.D. in Computer Science from Texas Tech University, specializing in AI, privacy-preserving, secure, and robust machine learning. Huixin has authored over 30 papers in top-tier journals and conferences, including Nature Machine Intelligence, Journal of Artificial Intelligence Research, and AAAI.

Her work has been widely recognized, including selection as an AAAI Student Abstract Finalist. She also serves as a reviewer for AAAI, IJCAI, KDD, Journal of Artificial Intelligence Research, and Communications Medicine. Her research focuses on integrating large language models and graph-based methods to advance precision medicine, with an emphasis on addressing real-world genomic and clinical challenges through innovative, secure, and robust AI-driven solutions.