Table 8.

Summary of deep learning models for predicting AVF functionality and VA quality in HD patients.

Author	Dataset	Model	Method	Outcome
Ota et al. [17]	AVF sounds recording 1 min from 20 patients.	CNN + BiLSTM (Bidirectional Long Short-Term Memory)	Extracted heartbeat-specific arteriovenous fistula sounds.	0.70 to 0.93 of Accuracy, 0.75 to 0.92 of AUC.
Julkaew et al. [18]	Clinical data from 398 HD patients	DeepVAQ-CNN	Used Photoplethysmogram (PPG) to predict VA quality, trained and fine-tuned CNN.	0.92 of Accuracy, 0.96 of Specificity, 0.88 of Precision, 0.84 of F1-score, 0.86 of AUC.
Peralta et al. [19]	13,369 dialysis patients (EuCliD® Database)	AVF-FM (XGBoost)	Predict AVF failure in 3 months.	0.80 of AUC (95% CI 0.79–0.81).
Nguyen et al. [22]	300 qualified and 202 unqualified PPG waveforms	1D-CNN + FCNN	Developed an ML algorithm to assess PPG signal quality for prediction of blood flow volume, using waveform quality criteria.	Transformed NN: 0.94 of Accuracy. 1D-CNN: 0.95 of Accuracy.
Zhou et al. [38]	2565 AVF blood flow sounds from 433 patients	Vision Transformer (ViT)	AVF sounds from 6 locations, pre-processed into Mel-spectrograms and recurrence plots.	ViT: 0.92 of Sensitivity, 0.79 of Specificity, 0.91 of F1-score.
Chung et al. [23]	437 audio recordings from 84 HD patients	CNN, ViT-GRU	AV access bruit recordings converted to Mel-spectrograms. Models trained to predict dysfunction.	CNN: 0.70 of F1 Score, 0.71 of AUC. ViT-GRU: 0.52 of F1-score, 0.60 of AUC.
Park et al. [39]	80 audio files from 40 HD patients.	ResNet50, EfficientNetB5, DenseNet201	Digital AVF sounds recorded, converted to mel spectrograms, and used DCNN models.	ResNet50: 0.99 of AUC. EfficientNetB5: 0.98 of AUC.