Georgia Institute of Technology
Interpretable Deep Transfer Learning for Non-Invasive Assessment of Knee Osteoarthritis Via Acoustic Emissions
Abstract
dc:description.abstractKnee osteoarthritis (OA) is a prevalent and disabling joint disease that is often assessed using imaging, symptom reports, and clinical examination. Although these approaches are valuable, they may be expensive, subjective, or insufficiently sensitive to early functional changes. Knee acoustic emissions (KAEs), also referred to as joint acoustic emissions (JAEs), provide a non-invasive and low-cost sensing modality that captures mechanical information generated during knee movement. This dissertation investigates the utility of KAEs for osteoarthritis assessment from two complementary perspectives: understanding the physiological factors that influence these signals and developing an interpretable deep learning framework for OA classification. The first part of this work examines the independent and interacting effects of body mass index (BMI) and OA on JAEs recorded during different movement tasks and from different sensor locations. The results show that BMI has a stronger effect during weight-bearing movements such as sit-to-stand, particularly in infrapatellar regions, whereas OA-related differences are more clearly expressed during seated flexion-extension, especially in medial regions. These findings suggest that movement type and sensor placement substantially affect the interpretability of knee acoustic signals and should be carefully considered in both analysis and model design. Importantly, the results indicate that JAEs remain informative even in individuals with higher BMI, provided that acquisition conditions and feature interpretation are appropriately controlled. The second part of the thesis presents an end-to-end deep learning pipeline for classifying knee OA from KAE-derived spectrograms. A ResNet-18 model with transfer learning was used to learn discriminative representations directly from acoustic data, avoiding reliance on handcrafted features. The study further evaluated spectrogram design choices and introduced a signal-envelope weighting strategy to emphasize acoustically meaningful regions of the input representation. The proposed framework achieved strong classification performance in a clinically representative cohort, demonstrating the feasibility of deep learning-based OA detection using KAEs. To improve trustworthiness and clinical interpretability, explainable artificial intelligence methods were applied to analyze model behavior. In particular, FullGrad-based saliency analysis was used to identify time-frequency regions that contributed most strongly to predictions. The resulting explanations showed that the model relied on structured and clinically plausible acoustic patterns rather than spurious artifacts, supporting the validity of the learned decision process. Overall, this dissertation demonstrates that knee acoustic sensing is a promising approach for OA assessment and that both physiological confounding factors and model interpretability must be addressed for reliable deployment. By combining statistical analysis of BMI and OA effects with an interpretable deep learning framework, this work contributes both practical and methodological insights toward the development of non-invasive, point-of-care tools for knee health monitoring.
Degree
thesis:*- Name thesis:degree_name
- Electrical and Computer Engineering, MS
- Department thesis:degree_department
- Electrical and Computer Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Emirdagi, Ahmet Rasim
- Advisor dc:contributor.advisor
-
- Inan, Omer T.
- Committee members dc:contributor.committeemember
-
- Ewart, David
- Davenport, Mark
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81635