A Study on the Generation of Spectrogram for the Detection of Venous Needle Dislodgement by Image Recognition Using Machine Learning

The safety and quality of life for the approximately 350,000 dialysis patients in Japan are critically affected by the risk of unintentional needle dislodgement during hemodialysis, which represents a significant cause of medical accidents. This study introduces an innovative approach to detect needle dislodgement, focusing on the nonhemolytic detection of indwelling needle dislodgement. By analyzing the impedance change through capacitive coupling between the patient and a detection circuit, we developed a prototype system that generates spectrograms based on frequency and phase characteristics from 1 kHz to 1 MHz. These spectrograms capture the changes in electrical characteristics associated with needle dislodgement, with alterations in color indicating changes in gain and phase difference. Initial experiments demonstrated the feasibility of using spectrogram analysis to visually distinguish normal from venous needle dislodgement (VND) conditions. Our results emphasize the potential for applying machine learning algorithms to these spectrograms for accurate and automated detection of VND. This could significantly improve patient safety by enabling the early detection of dislodgement events under various conditions, including partial needle contact or leakage at the puncture site. Future research will focus on expanding the dataset of normal and VND state spectrograms for machine learning training, aiming to validate the effectiveness of this technology in clinical settings.