Acoustic feature selection and clinical information integration for machine learning based dialysis shunt function prediction
- 1Department of Nephrology, Graduate School of Medicine, University of Yamanashi
- 2Department of Nephrology and Rheumatology, Kofu Municipal Hospital
Abstract
Recent studies have suggested that analysis of hemodialysis shunt sounds can predict hemodynamics, although optimal acoustic features remain unclear. This study compared engineered features (time- and frequency-domain statistics) with perceptual features such as Mel-scale transformations for predicting blood flow volume (FV) and resistive index (RI). We examined whether integrating clinical parameter variables could improve prediction accuracy. Shunt sounds from 107 recordings (48 patients with chronic kidney disease who underwent arteriovenous fistula creation surgery) were analyzed. Forty-seven acoustic features were extracted and classified into engineered or perceptual pools. After patient-level data splitting (80% training, 20% test), five features per pool were selected using bootstrap-based stability selection. Machine learning models were developed to classify brachial FV (threshold: 400 mL/min) and RI (threshold: 0.6), and model performance was evaluated using bootstrap validation. Both acoustic feature sets demonstrated stable FV classification performance (AUC: 0.866 [engineered], 0.864 [perceptual]). Identical acoustic features were selected in acoustic-only and combined models, and the inclusion of clinical parameters did not improve test-set performance. RI classification performed worse than FV classification. Overall, perceptual features yielded stable FV predictions comparable to those from engineered features, supporting their utility for acoustic-based shunt monitoring. Further validation in larger cohorts is warranted.
Related
- Quantum-classical hybrid learning framework for Parkinson’s disease prediction using acoustic and clinical features
- Validation Study of ASA Physical Status Prediction Model Developed Through Holistic Integration of Lung Images, Electrocardiogram Data and Clinical Data
- Externally validated machine learning models for prediction of treatment selection and clinical outcomes in posterior communicating artery aneurysms
- Outcome-specific value of longitudinal clinical information and model complexity for dynamic prediction of type 2 diabetes-related outcomes: a pooled landmarking study
- Cross-corpus consensus feature selection for language-invariant Parkinson’s disease detection: a dual-stream acoustic and spectrogram-attention framework
BibTeX
@article{yoshida2026acoustic,
title = {{Acoustic feature selection and clinical information integration for machine learning based dialysis shunt function prediction}},
author = {Shun Yoshida and Daiichiro Akiyama and Kie Ohkoshi and Kohei Yamamura and Keiichi Osano and Miho Shikata and Toshihisa Ishii and Makiko Konishi and Kazuya Takahashi and Ayumu Nakashima},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-62029-w},
url = {https://doi.org/10.1038/s41598-026-62029-w},
note = {CC BY},
}
RIS
TY - JOUR
TI - Acoustic feature selection and clinical information integration for machine learning based dialysis shunt function prediction
AU - Shun Yoshida
AU - Daiichiro Akiyama
AU - Kie Ohkoshi
AU - Kohei Yamamura
AU - Keiichi Osano
AU - Miho Shikata
AU - Toshihisa Ishii
AU - Makiko Konishi
AU - Kazuya Takahashi
AU - Ayumu Nakashima
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-62029-w
UR - https://doi.org/10.1038/s41598-026-62029-w
AB - Recent studies have suggested that analysis of hemodialysis shunt sounds can predict hemodynamics, although optimal acoustic features remain unclear. This study compared engineered features (time- and frequency-domain statistics) with perceptual features such as Mel-scale transformations for predicting blood flow volume (FV) and resistive index (RI). We examined whether integrating clinical parameter variables could improve prediction accuracy. Shunt sounds from 107 recordings (48 patients with chronic kidney disease who underwent arteriovenous fistula creation surgery) were analyzed. Forty-seven acoustic features were extracted and classified into engineered or perceptual pools. After patient-level data splitting (80% training, 20% test), five features per pool were selected using bootstrap-based stability selection. Machine learning models were developed to classify brachial FV (threshold: 400 mL/min) and RI (threshold: 0.6), and model performance was evaluated using bootstrap validation. Both acoustic feature sets demonstrated stable FV classification performance (AUC: 0.866 [engineered], 0.864 [perceptual]). Identical acoustic features were selected in acoustic-only and combined models, and the inclusion of clinical parameters did not improve test-set performance. RI classification performed worse than FV classification. Overall, perceptual features yielded stable FV predictions comparable to those from engineered features, supporting their utility for acoustic-based shunt monitoring. Further validation in larger cohorts is warranted.
SN - 2045-2322
ER -