03/31/2026
By Nazim Ahmed Belabbaci

The Richard A. Miner School of Computer and Information Sciences invites you to a doctoral dissertation defense by Nazim Belabbaci entitled "Generative and Domain-Informed Modeling of Biomedical Signals Under Data Scarcity and Real-World Noise."

Date: Friday, April 10, 2026
Time: 2 - 4 p.m. (EDT), Noon - 2 p.m. (MST)
Location: Hybrid: Via Zoom and Dandeneau 309

Title: Generative and Domain-Informed Modeling of Biomedical Signals Under Data Scarcity and Real-World Noise

Committee Members

  • Chair: Mohammad Arif Ul Alam, Ph.D., Assistant Professor, Richard A Miner School of Computer and Information Sciences, UMass Lowell
  • Benyuan Liu, Ph.D., Professor, Richard A Miner School of Computer and Information Sciences, UMass Lowell
  • Yan Luo, Ph.D., Professor, Department of Electrical and Computer Engineering, UMass Lowell
  • Hassan Ghasemzadeh, Ph.D., Program Director & Associate Professor, College of Health Solutions, Arizona State University (ASU)


Abstract: Wearable and spectroscopic sensing systems offer strong potential for non-invasive health monitoring and disease diagnosis. However, their practical deployment is limited by the lack of data and the presence of noise and artifacts that differ significantly from controlled acquisition conditions. These factors make it harder to develop robust and generalizable prediction models. This dissertation addresses these challenges across both spectroscopic and wearable sensor data through a combination of generative modeling, domain-informed signal processing, and real-world system design.

We first address data scarcity in spectroscopic analysis by developing a latent diffusion framework for augmenting Fourier-transform infrared (FTIR) spectroscopy data, demonstrating improved diagnostic performance under class imbalance and limited sample sizes. We then study wearable sensing data, evaluating hybrid deep learning approaches (Positionally Encoded Transformers, LSTM-Tabnet) on multi-modal physiological time-series from a Hajj crowd monitoring dataset, achieving competitive performance on fatigue, mood, and activity prediction tasks.

We then build HydroTrack, a multi-wavelength optical wearable system for non-invasive hydration monitoring and conduct an IRB-approved data collection study. By combining domain-informed feature design with learning-based modeling, we show that changes in osmolarity can be predicted from optical signals under controlled conditions, highlighting both the potential and limitations of deploying wearable sensing systems. Building on this controlled setting, we examine the impact of real-world conditions on signal quality and model performance. Data collected under unconstrained conditions, including subject movement and environmental noise, introduces a significant distribution shift relative to controlled measurements. To address this, we propose a diffusion-based denoising strategy guided by Maximum Mean Discrepancy (MMD) to characterize the discrepancy between clean and noisy signal distributions. A denoising diffusion model trained on controlled data is applied at inference time to project noisy signals toward a clean signal manifold, resulting in improved downstream classification performance.

Together, these contributions show that combining generative modeling with domain-informed co-design enables robust learning from biomedical signals under realistic constraints and provides a practical path toward robust and deployable health monitoring systems.