Briden, M. (CSE) – Representation Learning and Generative Forecasting for Noisy and Limited Clinical Data: Applications in Wound Healing and EEG

The rapid integration of artificial intelligence and machine learning into clinical practice has driven advances in disease classification, segmentation, and clinical decision support. However, the complexities of medical data pose a challenge to widespread adoption. The rarity of medical conditions, ethical considerations, and varying acquisition protocols leads to limited and noisy data. The time-intensive process of labeling data, the high degree of accuracy required in clinical settings, and the ill-defined nature of certain medical conditions further complicate the application and deployment of machine learning models. Likewise, high‐stakes medical decisions demand trustworthy and interpretable predictions. However, prioritizing trust and explainability is rarely a primary objective in most model designs.

This thesis addresses three key challenges in machine learning for healthcare. First, we develop methods for learning under noisy and limited medical data, focusing on representation learning strategies that improve generalization when datasets are small or contain mislabeled samples. Second, we explore the prediction of generative outcomes amid label noise and data scarcity, utilizing parameter-efficient and temporal generative models to forecast disease trajectories. Third, we advance trustworthy and explainable medical artificial intelligence by designing deep architectures that provide interpretable outputs suitable for clinical decision-making.

These challenges are addressed in the context of two complementary medical modalities: wound healing images and electroencephalogram signals. Wound healing tasks focus on predicting healing trajectories while enhancing interpretability through segmentation-based explanations and training large models in light of extreme data noise and scarcity. Electroencephalogram-based tasks emphasize representation learning and explainability for non-invasive mental state classification. These experiments demonstrate the clinical relevance of the proposed approaches and their ability to operate under challenging medical conditions across both imaging and physiological signal domains.

Event Host: Michael Briden, PhD Candidate, Computer Science & Engineering

Advisor: Narges Norouzi