Identification of Patients with Recent Cerebrovascular Symptoms by Combined Clinical and Carotid Plaque’s Calcium Imaging Data Using Interpretable Machine Learning

This repository contains the code to reproduce the experiments of "Identification of Patients with Recent Cerebrovascular Symptoms by Combined Clinical and Carotid Plaque’s Calcium Imaging Data Using Interpretable Machine Learning", currently under review at JACC: Cardiovascular Imaging.

Abstract

Carotid plaque composition plays a significant role in the association with cerebrovascular events, but the role of plaque’s calcium configuration is not very well understood. Using CT angiography examinations of patients with bilateral plaque we derived an interpretable model that identifies patients with recent stroke with reasonable diagnostic accuracy (area-under-curve: 0.71 [95% CI 0.57 – 0.81]). Analysis of predictions revealed that the evidence of calcification, in particular the positive rim sign on the right side, the older age and hyperlipidemia had major impact on identifying symptomatic patients. We also observed that intimal or superficial calcifications on the right carotid plaque were less common in symptomatic patients. To summarize, integrating clinical and demographic factors with data of plaque’s calcium configuration allows identifying symptomatic patients with reasonable diagnostic accuracy.

Acknowledgments

• The code for comparing ROC AUCs using DeLong test is taken from the Yandex School of Data Analysis github repository.

• The code for producing the observed vs. predicted plot is based on the verhulst package.