DialySafe: A Capacitance-Based Sensor to Detect Infection in Patients Undergoing Peritoneal Dialysis

Background: Peritoneal dialysis (PD) offers greater autonomy and lower costs than hemodialysis, yet is underutilized in the US, with only 10% of eligible ESRD patients on PD. Furthermore, peritonitis occurs in 26% of PD patients annually with 2/3 resulting in hospitalization. Early detection of infection can reduce hospitalization risk by over 60%, but most patients rely on visual inspection of effluent for turbidity or cloudiness which is subjective and insensitive. An objective real-time detection method could reduce complications and improve patient outcomes. Methods: A non-invasive capacitance-based sensor was developed that clamps onto standard PD tubing and detects dielectric changes associated with cellular content or other anomalous debris in the effluent. The signal is digitized and processed using trained machine learning models to assess infection risk (Figure 1). Benchtop validation was conducted using serial dilutions of E. coli and K562 cells in dialysate to simulate peritonitis. Capacitance measurements were collected at multiple frequencies and used to train three machine learning classifiers: Random Forest (RF) and Support Vector Machines (SVM) with RBF and polynomial kernels (Table 1). Results: The sensor showed frequency-dependent trends correlating with cell concentration. All models demonstrated strong performance, confirming the feasibility of capacitance-based infection detection in PD effluent. Conclusion: This capacitance-based sensor provides a non-invasive, real-time method to detect early signs of peritonitis in PD patients. Its high accuracy and ease of integration make it a promising tool for remote monitoring and earlier intervention especially for pediatric and underserved populations at higher risk of infection-related complications. Funding: Other NIH Support - R25 Grant ($2,500.00)Capacitance sensor detects effluent changes from infection; signal is analyzed by ML to classify healthy vs. infected PD fluid. Sensor Machine Learning Model Performance - Model Type Accuracy Sensitivity Specificity F1-Score Random Forest 98.8% 100% 97.6% 98.8% SVM (RBF Kernel) 98.3% 96.7% 100% 98.3% SVM (Poly Kernel) 96.7% 100% 93.3% 96.8%