Evaluation of Hemodialysis Arteriovenous Bruit by Deep Learning

Abstract

Physical findings of auscultation cannot be quantified at the arteriovenous fistula examination site during daily dialysis treatment. Consequently, minute changes over time cannot be recorded based only on subjective observations. In this study, we sought to supplement the daily arteriovenous fistula consultation for hemodialysis patients by recording the sounds made by the arteriovenous fistula and evaluating the sounds using deep learning methods to provide an objective index. We sampled arteriovenous fistula auscultation sounds (192 kHz, 24 bits) recorded over 1 min from 20 patients. We also extracted arteriovenous fistula sounds for each heartbeat without environmental sound by using a convolutional neural network (CNN) model, which was made by comparing these sound patterns with 5000 environmental sounds. The extracted single-heartbeat arteriovenous fistula sounds were sent to a spectrogram and scored using a CNN learning model with bidirectional long short-term memory, in which the degree of arteriovenous fistula stenosis was assigned to one of five sound types (i.e., normal, hard, high, intermittent, and whistling). After 100 training epochs, the method exhibited an accuracy rate of 70-93%. According to the receiver operating characteristic (ROC) curve, the area under the ROC curves (AUC) was 0.75-0.92. The analysis of arteriovenous fistula sound using deep learning has the potential to be used as an objective index in daily medical care.

Keywords: arteriovenous fistula; artificial intelligence; auscultation; convolutional neural network; deep learning; hemodialysis patient; shunt sound.

Figure 1

Outline of the arteriovenous fistula sound learning model. A series of arteriovenous fistula sounds are recorded in one medical examination. All continuous tones are converted into spectrograms. A single heartbeat is detected using the mechanism of object detection (R-CNN: regions with convolutional neural networks) similar to that used for face detection in digital cameras and smartphones. The spectrogram of one arteriovenous fistula sound is used as input data. A deep learning model is used for learning, and the importance of the sounds obtained during a general medical examination is expressed by multiclass classification.

Figure 2

Learning preprocessing. A spline curve was created to eliminate noise from one auscultation sound. We extracted a number of convex curve ranges from the created spline curve with a duration of 0.5–2.0 s. The result was estimated to include an arteriovenous fistula sound equivalent to one heartbeat. The sound of one beat of the arteriovenous fistula was extracted by a deep learning classifier. Arteriovenous fistula sounds of 10,000 beats were classified into one of five types (i.e., normal sound, hard sound, high sound, intermittent sound, and whistling).

Figure 3

Learning curve for each learning model and each input source. Accuracy is presented in the upper row, whereas the loss is in the lower row. The left side is the transition based on the training data, and the right side is the transition based on the verification data. The horizontal axis indicates the number of times of learning, and accuracy increases as learning progress. Loss represents the difference between the answer of the input data predicted by the model during the learning process (e.g., the degree of firing that is a high tone) and the teacher’s answer to the actual input data (the high tone is the correct answer). It can be observed that the difference between the answer and answer from the learning model obtained during learning has decreased. The CRNN:Bi-GRU model, which had the mel-frequency log spectrogram as input, was the learning model with good accuracy and loss in both training data and verification data.

Figure 4

Final ROC curve obtained using the final stage model. The ROC curve is obtained by changing the threshold and plotting the true positive rate (TPR) at each threshold on the vertical axis and the false positive rate (FPR) on the horizontal axis. The ROC curve is on the blue line. The red dotted line shows the curve for obtaining the FPF at each threshold. It can be observed that, by changing the threshold, the threshold can be adjusted to lower FPR to detect rare diseases.

Figure 4

Final ROC curve obtained using the final stage model. The ROC curve is obtained by changing the threshold and plotting the true positive rate (TPR) at each threshold on the vertical axis and the false positive rate (FPR) on the horizontal axis. The ROC curve is on the blue line. The red dotted line shows the curve for obtaining the FPF at each threshold. It can be observed that, by changing the threshold, the threshold can be adjusted to lower FPR to detect rare diseases.

Figure 5

Arteriovenous fistula analysis sound of two cases. The first case is arteriovenous fistula construction resulting from acute renal failure. The puncture began two weeks later. Immediately after creation, the initial sound had a small hardness component due to the influence of a vasospasm. In the second case, an arteriovenous fistula was constructed after a long history of diabetes. The puncture began three weeks later. In both cases, it was observed that the arteriovenous fistula sound was hard at the beginning, even when the arteriovenous fistula sound developed, and that the ratio of random intermittent and harmonic sounds due to the start of puncturing increased.

Figure 6

Grad-CAM for visualization of characteristic sites. With Grad-CAM, if the image is judged to be a dog using a dog image classifier, the image is colored. The same applies to a cat image classifier. When the characteristic portion of the sound pattern image that felt “high” is visualized, the 250–750 Hz region in the systole is emphasized in the heat map. In the case of a sound pattern image that felt “hard,” the silent region in the diastolic area was emphasized.