Deep Learning for Cardiovascular Risk Assessment: Proxy Features from Carotid Sonography as Predictors of Arterial Damage

TL;DR

This work tackles scalable cardiovascular risk assessment by exploiting carotid ultrasound videos and using hypertension as a proxy for visual arterial damage. It finetunes VideoMAE on 31,019 ultrasound videos from the Gutenberg Health Study to classify individuals as hypertensive or not, treating the model output as a proxy marker of vascular damage. The study demonstrates that ultrasound-derived visual features correlate with clinical risk factors, laboratory biomarkers, comorbidities, and future cardiovascular events, supporting a novel digital biomarker for CV risk. It highlights the potential for non-invasive, video-based risk stratification at the individual level and points to future directions in explainable AI and multi-proxy risk scoring to improve cardiovascular prevention.

Abstract

In this study, hypertension is utilized as an indicator of individual vascular damage. This damage can be identified through machine learning techniques, providing an early risk marker for potential major cardiovascular events and offering valuable insights into the overall arterial condition of individual patients. To this end, the VideoMAE deep learning model, originally developed for video classification, was adapted by finetuning for application in the domain of ultrasound imaging. The model was trained and tested using a dataset comprising over 31,000 carotid sonography videos sourced from the Gutenberg Health Study (15,010 participants), one of the largest prospective population health studies. This adaptation facilitates the classification of individuals as hypertensive or non-hypertensive (75.7% validation accuracy), functioning as a proxy for detecting visual arterial damage. We demonstrate that our machine learning model effectively captures visual features that provide valuable insights into an individual's overall cardiovascular health.

Figures (6)

• Figure 1: Video before (top) and after (bottom) preprocessing. We remove the user interface and the heartline to avoid biases.
• Figure 2: Distribution of individuals' gender and hypertension diagnosis according to age. The mean age of the validation population was found to be $55.5\pm11.2$ years, with $46.9\%$ of the female population and $55.9\%$ of the male population diagnosed with hypertension. The validation dataset revealed a total of $49.2\%$ women.
• Figure 3: Confusion matrix for the trained model on validation data. Counts are given at sample level. During an evaluation, our pipeline samples uniformly 9891 samples from 6231 unique videos from 2847 individuals in the validation dataset.
• Figure 4: Statistical comparison of the classification results on the validation data with respect to different clinical parameters. For all variables the individuals classified as “high visual arterial damage” (high VD) show a significant higher likelihood or median value than individuals with low visual damage (low VD). Individuals with high visual damage and a positive hypertension diagnosis exhibit the worst cardiovascular health condition (in terms of the presented indicators).
• Figure 5: The total event counts in the validation dataset with respect to the model’s classifications and the use of antihypertensive agents. The upper section considers individuals taking antihypertensive agents, while the lower section focuses on those not using such medications. In both groups, the role of hypertension as a major risk factor for various cardiovascular events is emphasized by the observed incidents within the GHS cohort. Among individuals taking antihypertensive agents, the majority of events are observed in those with a positive diagnosis of hypertension and high visual arterial damage. However, in cases of untreated hypertension, individuals without hypertension but exhibiting high visual damage account for a significant proportion of all events, regardless of the type of event.