Pilot Study of Using Machine Learning to Detect Atherosclerotic Renal Artery Stenosis From Spectral Doppler Waveforms

Abstract

Introduction

We investigated whether machine learning (ML) could be used to determine atherosclerotic renal artery stenosis (ARAS) using spectral Doppler waveforms in renal duplex ultrasound (DUS).

Methods

Patients with unilateral ARAS (contralateral normal kidney) confirmed by angiogram and requiring renal artery stent placement were retrospectively identified from January 2000 to January 2022. The exclusion criteria were unavailable preoperative renal DUS images, concomitant fibromuscular dysplasia, more than 1 renal artery on either side, or a previously placed renal artery stent with in-stent restenosis. Two hundred patients were selected; the affected kidney was used as the positive case and the contralateral kidney was used as the control. The spectral waveforms were reconstructed by manually tracing the outer envelope using WebPlot Digitizer. The graphical coordinates were then converted into 1-dimensional velocity signals. Signals were labeled as ARAS and normal and then randomly divided into training (80%) and testing (20%) datasets. A 1-dimensional convolutional neural network (CNN) was trained to classify the signals and detect ARAS. An Adam optimizer with a learning rate of 0.001 and a cross-entropy loss function were utilized. Five-fold cross-validation was applied, and the model was trained for 1000 epochs.

Results

A total of 396 signals were used from 198 patients after excluding 2 patients because of inadequate signal extraction (median age = 72 years, females = 51.0%). The overall accuracy of the trained model was 0.95 with a precision of 0.94. The area under the receiver operating characteristic curve was 0.97.

Conclusion

ML has been successfully employed to detect ARAS using arterial spectral Doppler waveforms in DUS.