Machine learning for abdominal aortic calcification assessment from bone density machine-derived lateral spine images

Area of Science:

• Cardiovascular Disease Research
• Artificial Intelligence in Medical Imaging
• Osteoporosis Screening

Background:

• Lateral spine images from bone density machines can reveal abdominal aortic calcification (AAC).
• AAC scoring by specialists assesses cardiovascular disease (CVD) risk but is time-consuming and requires extensive training.
• Automating AAC scoring could streamline CVD risk assessment.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) algorithm for automated AAC-24 scoring.
• To assess the association between machine-learning AAC (ML-AAC-24) scores and future Major Adverse Cardiovascular Events (MACE).

Main Methods:

• A CNN algorithm was trained and tested on 5012 lateral spine images with specialist AAC scores.
• Model performance was validated in a cohort of 8565 older adults.
• Cox proportional hazards models analyzed the link between ML-AAC-24 scores and MACE (death, heart attack, stroke).

Main Results:

• The ML-AAC-24 scores showed strong agreement (ICC=0.84) and 80% classification accuracy compared to specialists.
• Higher ML-AAC-24 scores correlated with increased MACE incidence (low 7.9%, moderate 14.5%, high 21.2%).
• After adjustments, moderate and high ML-AAC-24 scores significantly predicted MACE (HR 1.54 and 2.06, respectively).

Conclusions:

• Automated ML-AAC-24 scoring demonstrates high agreement with expert assessment.
• ML-AAC-24 scores effectively predict cardiovascular event risk in a real-world population.
• This AI-driven approach offers a scalable solution for identifying individuals at high CVD risk using existing imaging data.