CBCT Radiomics for Cardiovascular Risk Prediction

Tags: Machine Learning · Medical Imaging · Radiomics · CBCT · Healthcare AI · Python · Scikit-learn

This project investigates whether radiomic features extracted from dental Cone Beam CT (CBCT) scans can be used to predict major cardiovascular and cerebrovascular outcomes. Calcium deposits visible in head and neck CBCT scans are known indicators of systemic vascular disease, and this work explores their predictive value using machine learning.

Problem Motivation

Standard diagnostic tools for cardiovascular disease (e.g., stroke or myocardial infarction) require specialized equipment and clinical access. CBCT scans, however, are routinely collected in dental settings. Leveraging these scans for early risk prediction could enable low-cost, opportunistic screening in populations that otherwise lack access to advanced care.

Dataset

237 anonymized patients from Stony Brook University Hospital
Paired CBCT scans and longitudinal medical histories
~112 radiomic features extracted from calcification regions

Methodology

Radiomic feature extraction from segmented calcification regions
Unsupervised analysis to identify feature correlations
Feature selection via Information Gain (IG) and Recursive Feature Elimination (RFE)
Supervised learning using SVMs, Random Forests, Neural Networks, and ensembles
Binary and multiclass classification across multiple clinical outcomes

Predicted Outcomes

Stroke
Myocardial infarction
Hypertension
Low anticoagulant levels
Future cardiac surgery

Results

Models trained on features selected via Recursive Feature Elimination achieved the strongest performance. Final binary classification accuracies included:

Stroke: 88.23%
Myocardial Infarction: 82.35%
Hypertension: 91.31%
Low Anticoagulants: 89.36%
Cardiac Surgery: 86.54%

A multiclass model predicting all outcomes simultaneously achieved an accuracy of 70.6%.

Repository Structure

Allfeatures_* — Models trained using all extracted radiomic features
InfoGainFeatures_* — Models trained on features selected via mutual information
RFE_* — Models trained on features selected via recursive feature elimination
Recursive_Feature_Elimination_* — Trait-specific RFE procedures
Feature_Selection_&_Initial_Training.ipynb — Feature extraction, IG selection, and baseline modeling
Final_Model_Optimizations.ipynb — Final hyperparameter tuning and evaluation

Impact

This work demonstrates the feasibility of predicting major cardiovascular outcomes using routinely collected dental imaging. It highlights the potential for CBCT-based opportunistic screening pipelines that augment traditional healthcare workflows with data-driven risk assessment.