Using Ubiquitous Mobile Sensing and Temporal Sensor-Relation Graph Neural Network to Predict Fluid Intake of End Stage Kidney Patients

Mingyue Tang, Guimin Dong, Jamie Zoellner, Brendan Bowman, Emaad Abel-Rahman, Mehdi Boukhechba

Abstract: End-Stage Kidney Disease (ESKD) patients on hemodialysis suf-fer from kidney failure, with the inability to remove excess fluid causing fluid overload. This can cause many morbidities, and is one of the most insidious and common risk factors for mortality in ESKD patients. Existing solutions for fluid intake monitoring such as self-report and weight gain monitoring are burdensome, non-continuous, and usually administered in clinics only. It is then critical to develop a ubiquitous fluid intake monitoring system to help ESKD patients better control their fluid consumption. In this study, we propose to leverage smartwatch sensor data (e.g., Photoplethysmography (PPG), Gyroscope, etc.) combined with a temporal sensor relation graph neural network (TSR-GNN) to predict fluid intake given past sensing data between two dialysis sessions. Our empirical experiments highlight promising findings about the feasibility of using ubiquitous sensing to predict fluid intake, and demonstrate that the proposed model TSR-GNN outperforms the selected baseline models in both accuracy and robustness. Additionally, an in-depth analysis of model interpretability by attention weights and GNNExplainer variant is conducted to better under-stand the inter-sensor interactions and sensor contributions to the fluid intake prediction results.