Artificial Intelligence Chest X-Ray Opportunistic Screening Model for Coronary Artery Calcium Deposition: A Multi-Objective Model With Multimodal Data Fusion

Abstract

Objective: To create an opportunistic screening model to predict coronary calcium burden and associated cardiovascular risk using only commonly available frontal chest x-rays (CXR) and patient demographics.

Patients and methods: We proposed a novel multitask learning framework and trained a model using 2121 patients with paired gated computed tomography scans and CXR images internally (Mayo Clinic) from January 1, 2012, to December 31, 2022, with coronary artery calcification (CAC) scores (0, 1-99, and 100+) as ground truths. Results from the internal training were validated on multiple external datasets (Emory University Healthcare and Taipei Veterans General Hospital-from January 1, 2012, to December 31, 2022) with significant racial and ethnic differences.

Results: Classification performance between 0, 1-99, and 100+ CAC scores performed moderately on both the internal test and external datasets, reaching average f1-scores of 0.71±0.04 for Mayo, 0.65±0.02 for Emory University Healthcare, and 0.70±0.06 for Taipei Veterans General Hospital. For the clinically relevant risk identification, the performance of our model on the internal and 2 external datasets reached area under the receiver operating curves of 0.86±0.02, 0.77±0.03, and 0.82±0.03 for 0 versus 400+, respectively. For 0 versus 100+, we achieved area under the receiver operating curve of 0.83±0.03, 0.71±0.02, and 0.78±0.01, respectively. Prospective evaluation across 3 Mayo Clinic sites is on par with the external validations and reports only minimal temporal drift.

Conclusion: Open-source fusion artificial intelligence-CXR model performed better than existing state-of-the-art models for predicting CAC scores only on internal cohort, with robust performance on external datasets. This proposed model may be useful as a robust, first-pass opportunistic screening method for cardiovascular risk from regular CXR.

Figure 1

Multitask model. A ResNeXt101 backbone was used to classify MACE event and CAC category in parallel with weighted loss. CAC, coronary artery calcification; MACE, major adverse cardiovascular events

Figure 2

Top row: MTL fusion model receiver operating characteristic (ROC) curve for discrimination of CAC category on the Mayo Internal hold-out test set, external EUH, and external VGHTPE. Shaded regions display 95% confidence interval; bottom row: MTL fusion model receiver operating characteristic curve for discrimination of MACE on the Mayo Internal hold-out test set. CAC, coronary artery calcification; EUH, Emory University Healthcare; MACE, major adverse cardiovascular events; MTL, multitask learning; VGHTPE, Taipei Veterans General Hospital.

Figure 3

MTL fusion model receiver operating characteristic (ROC) curve for discrimination of high and low CAC group on the Mayo Internal hold-out test set and EUH and VGHTPE external datasets. Shaded regions display 95% confidence interval. CAC, coronary artery calcification; EUH, Emory University Healthcare; VGHTPE, Taipei Veterans General Hospital.

Figure 4

AUCROC achieved by the MTL model on prospective data. Each row represents different tasks and columns show the performance for a different time point. Shaded regions display 95% confidence intervals. AUCROC, area under the receiver operating characteristic curve; MTL, multitask learning