Arteriovenous fistula (AVF) occlusion is a problem faced by all hemodialysis patients. However, current clinical methods for assessing the degree of AVF occlusion primarily rely on auscultation and palpation, which are subjective and relatively inaccurate methods. In this study, we developed a portable, noninvasive audio recording device for collecting vascular blood flow sounds and collected data from three patients who underwent two or more percutaneous transluminal angioplasty (PTA) procedures within 8 months. The data collection period was 4 months. By short-time Fourier transform, we extracted 25 signal features and 6 acoustic features from the recorded blood flow sounds. Moreover, we developed the time-adaptive ensemble learning algorithm (TAELA) to create an AI-based regression model for estimating the degree of AVF occlusion. This model can assist physicians in predicting the optimal timing for PTA procedures. Experimental results indicated that the TAELA outperformed adaptive and gradient boosting regression algorithms in terms of the coefficient of determination (R2) and that the TAELA model did not exhibit overfitting. The developed audio recording device and TAELA enabled the effective and accurate identification of the degree of AVF occlusion in three hemodialysis patients. The overall system can serve as a portable, noninvasive, and user-friendly clinical diagnostic tool for assisting physicians in optimizing the interval between PTA procedures, thereby reducing surgical frequency for hemodialysis patients while ensuring that they do not experience complications associated with delayed dialysis