Artificial Intelligence for Lymphedema Detection and Staging Using Clinical Microwave Sensor Data

Lymphedema is a chronic and progressive condition characterized by abnormal accumulation of lymphatic fluid, frequently resulting from oncological interventions such as lymph node removal or radiotherapy. Traditional diagnostic techniques, though effective, often involve invasive procedures or expensive imaging systems, limiting their use for early screening and continuous monitoring. This thesis presents a novel, non invasive diagnostic approach combining microwave sensor data with machine learning techniques to detect, localize, and assess the severity of lymphedema. Using a MiniVNA Tiny vector network analyzer connected to a split ring resonator sensor, frequency dependent return loss measurements were collected from both healthy individuals and lymphedema patients. These signals were then filtered and processed to extract dielectric features sensitive to water content variations in tissue. A comprehensive pipeline was developed, including outlier rejection, feature engineering, and classification using a suite of models. The AdaBoost classifier achieved the best performance in binary classification of lymphedema presence, reaching an F1 Score of 92% and sensitivity of 95%. Gradient Boosting achieved 84% accuracy in identifying the affected limb. K-means clustering on handcrafted signal features successfully mapped patient samples onto a four cluster structure corresponding to Anderson’s clinical staging of the disease. Additionally, a Graph Neural Network (GNN) architecture representing each patient as a graph of limb measurement positions allowed interpretable classification, with attention mechanisms highlighting the most diagnostically relevant sensing sites. This approach matched handcrafted heuristics in 65.38% of cases and offers a data driven framework for explainable diagnostics. Altogether, this study demonstrates the feasibility of microwave sensing enhanced by AI to support early detection and monitoring of lymphedema with high accuracy, scalability, and clinical interpretability.