WPI - Computer Science, PhD Proposal Defense Wafaa Almuhamm " Minimal Supervision Multimodal Predictive Models for Osteoarthritis Pain " "

Wafaa Almuhammadi

PhD Candidate 

WPI - Computer Science Department 

Friday, December 20, 2024,

Time:  10:30 a.m. – 12:30 p.m.

Location: Fuller Labs  311

Committee members:

Advisor: Prof. Emmanuel Agu, WPI – Computer Science Dept.

Prof. Kyumin Lee, WPI – Computer Science Dept.

Prof. Miad Faezipour, Purdue University – Regenstrief Center for Healthcare Engineering.

Prof. Ahmed El-Sayed, University of Bridgeport – Electrical Engineering Dept.

Abstract:

Knee Osteoarthritis (KOA) affects over 364 million people globally and imposes significant economic and healthcare burdens. Accurate pain assessment is critical for effective KOA management but remains challenging due to reliance on subjective self-reports that are often inconsistent and prone to bias. Additionally, the scarcity of high-quality annotated pain data due to the time-consuming and expensive nature of data collection further complicates the development of machine-learning-based pain prediction models. 

Building on our completed work that explored Spatio-Temporal Gait Parameter (STGP) features to assess OA, this dissertation proposes minimal supervision multimodal OA pain prediction models that leverage large amounts of unlabeled or weakly labeled data to address critical data challenges. The proposed research integrates IMU data from wearable sensors to deliver objective, personalized, real-time pain assessments. Critical challenges are tackled in three research thrusts: 

1. Inter- and intra-subject variability in Physical Activity (PA) Data. Self-Supervised Momentum Contrastive Learning is proposed to extract robust and generalizable representations of PA patterns from unlabeled data.
2. Imbalance and Insufficiency in OA Pain Datasets. A framework that combines Few-Shot Learning to handle limited labeled examples, Style Transfer-based Methods to augment data diversity, Domain Adaptation to leverage auxiliary datasets, and Meta-Learning to optimize performance on imbalanced datasets, is proposed.
3. Heterogeneity and temporal complexity in longitudinal multimodal data. A framework that addresses temporal complexity of patient data using Graph Neural Networks with Hierarchical Contrastive Learning to capture both multimodal relationships and temporal dependencies, enabling personalized, dynamic pain predictions, is proposed.

Our contributions will enhance continuous OA pain monitoring, clinical decision-making, and tailored interventions, improving KOA management and patients' quality of life.