# Assessing arterial stiffness using characteristics of Korotkoff sounds

## TL;DR

This study explores using Korotkoff sounds during blood pressure measurements to assess arterial stiffness, a marker of vascular aging, through machine learning techniques.

## Contribution

The novel use of Korotkoff sound features and machine learning to infer arterial stiffness and age-related vascular differences is presented.

## Key findings

- Features like center of mass and peak frequency from Korotkoff sounds showed significant differences between age groups.
- Deep learning models achieved high classification accuracy (up to 93.7%) for age-related vascular differences.
- Korotkoff sound features had a modest association with brachial-ankle pulse wave velocity (baPWV).

## Abstract

Arterial stiffness is a recognized marker of vascular ageing and is associated with adverse cardiovascular outcomes. However, routine assessment of pulse wave velocity (PWV) remains limited in many clinical and home settings. This study investigated the feasibility of extracting arterial stiffness-related information from Korotkoff sounds recorded during cuff-based blood pressure measurement using feature analysis and machine learning.

Korotkoff sounds were collected from 123 young (25.9 ± 2.2 years) participants and 112 older (67.5 ± 6.7 years) participants using a custom-developed device as a proof-of-concept for age-related vascular differences. In addition, 81 hospital participants with measured brachial-ankle PWV (baPWV) were enrolled and grouped according to baPWV to further evaluate clinical feasibility. Time- and frequency-domain features were extracted, and both traditional feature-based models and deep learning approaches were applied for classification.

Extracted features including center of mass, skewness, and peak frequency showed significant differences between the age-stratified groups. Two deep learning models achieved classification accuracies of 89.3% and 93.7%, respectively, outperforming traditional feature-based analysis. In the baPWV-defined classification task, model performance was moderate (accuracy 87.5% and 81.3%). In the baPWV-measured cohort, Korotkoff sound-derived features showed a statistically significant but modest association with measured baPWV.

Korotkoff sounds contain measurable information related to vascular ageing and arterial stiffness, and machine learning can leverage these signals for group discrimination. Given that the primary comparison used age as a surrogate label and clinical outcomes were not assessed, the present data do not establish incremental value for cardiovascular risk stratification beyond age and blood pressure. Larger studies with standardized PWV measurements, ideally carotid-femoral PWV (cfPWV), and prospective validation are required before prognostic or risk-stratification claims can be made.