Prediction of cardiovascular risk from panoramic radiographs using deep learning

Introduction Panoramic dental radiographs are widely used in clinical practice. While they are known to capture carotid artery calcifications associated with increased risk of cardiovascular disease (CVD), the prognostic information contained in these images is rarely used for cardiovascular risk assessment. Purpose To develop a deep learning–based survival model for predicting cardiovascular risk from panoramic radiographs and to evaluate whether image-derived risk scores improve risk prediction beyond standard demographic factors. Methods A cohort of 28652 patients who underwent panoramic radiography between 2014 and 2024 was linked to follow-up diagnoses from the Israeli national hospital discharge register via the Ministry of Health Big Data Research Platform (TIMNA). A sub-cohort of patients aged 40-79 years, reflecting the clinically relevant age range, was used for risk assessment. Convolutional neural network (CNN) models were trained to predict a continuous image-based risk of four cardiovascular outcomes (cardiac, cerebral, peripheral and any CVD). Model performance was assessed using the concordance index (C-index) in cross-validation (CV) and an independent holdout set. Incremental prognostic value was evaluated by comparing a demographic model including age and sex with an image-augmented model incorporating the CNN-derived risk score, using likelihood ratio tests and changes in C-index. Clinical utility was assessed by decision curve analysis (DCA) at a 5-year horizon. Model interpretability was explored using gradient-based class activation mapping (Grad-CAM). Results The risk assessment cohort included 10656 patients (46% male; mean age 56±11 years), of whom 2811 (26.4%) experienced a CVD outcome during follow-up. The CNN learned an image-based risk score associated with cardiovascular survival. The image-augmented model provided effective risk stratification, with marked separation of predicted survival curves across quartiles of 5-year cardiovascular risk in both sexes (Figure), corresponding to an absolute risk difference of 0.38 (95% CI 0.36-0.41) between lowest and highest quartiles for any CVD outcome. Compared with the demographic model, addition of the image-derived risk score improved model fit (P<0.001) and discrimination, with significant C-index improvements of 2.5-5.5 points (P<0.02) across nearly all outcomes in the CV and holdout sets (Table). DCA demonstrated higher net benefit for the image-augmented model across clinically relevant threshold probabilities. Grad-CAM analysis highlighted anatomically plausible regions associated with increased predicted risk, including carotid artery region. Conclusion Deep learning–derived risk scores from panoramic dental radiographs provide incremental prognostic information beyond age and sex. Routine dental imaging may enable incidental cardiovascular risk assessment, supporting earlier identification of individuals at increased risk.