Wearable Sensors-based Real-Time Fluid Balance Monitoring in Dialysis Patients with Edge AI

The monitoring of fluid balance is a critical measure in controlling dialysis patients with chronic kidney disease (CKD) because overhydration and underhydration may cause severe clinical complications. This paper introduces a wearable sensing system that provides real-time monitoring of fluid status in patients during hemodialysis possibly coupled with Edge AI services that allow predicting fluid-status continuously and on an individual patient basis. Multimodal patch with bioimpedance and photoplethysmography (PPG) sensors can record physiological signals, and a lightweight edge-deployed neural network specifically optimized on ARM Cortex-M4F microcontrollers is used to locally process the signals. The system was tested against the data gathered during 30 dialysis patients in 120 sessions and in comparison with regular clinical tools legally permitted to measure. The model returned root mean squared error (RMSE) of 72.4 +/- 8.7 mL and a mean absolute percentage error (MAPE) of 4.6%, considerably better than traditional baselines using linear regression, random forest models and commercial bioimpedance analyzers. Classification of the hydration state into hypovolemia, euvolemia, and hypervolemia resulted in an accuracy of 91.2 percent average F1-score of 0.89. Also, the system exhibited ultra-low power consumption (18.6 ms inference latency and 2.3 mJ/inference) signifying its appropriateness in real-time implementation on power-constrained wearable devices. These findings confirm the potential of the suggested framework to facilitate the proactive management of the fluids in the acute setting, minimize intradialytic complications, and facilitate lifelong at-home follow-ups, facilitating the development of individual and precision nephrology care.