Interpretable AI Enhances Multicenter PET Biomarker Diagnosis of Coronary Artery Disease
Overview
An AI model integrating 10 PET myocardial perfusion imaging parameters demonstrated superior diagnostic accuracy for coronary artery disease (CAD) compared to conventional clinical scores and individual PET biomarkers. Validated across multiple centers with 1278 external patients, the model achieved an AUC of 0.83, outperforming ischemic total perfusion deficit, myocardial flow reserve, stress myocardial blood flow, and coronary artery calcium scores.
Background
Positron emission tomography (PET) myocardial perfusion imaging combined with computed tomography (CT) provides comprehensive data including perfusion, myocardial blood flow (MBF), functional metrics, and coronary artery calcium (CAC) for assessing suspected coronary artery disease (CAD). Despite the availability of these markers, clinical interpretation often does not fully integrate them systematically. Artificial intelligence offers an opportunity to synthesize these diverse parameters into a single, interpretable diagnostic tool. This study developed and externally validated a deep learning and machine learning AI model using standard PET imaging measurements to improve CAD diagnosis across multiple centers.
Data Highlights
Parameter
CAD Patients (Median or %)
Non-CAD Patients (Median or %)
Significance (p-value)
Ischemic Total Perfusion Deficit (TPD)
Higher
Lower
<0.001
Stress TPD
Higher
Lower
<0.001
Transient Ischemic Dilation Ratio (TID)
Higher
Lower
0.022 - 0.001
Stress Myocardial Blood Flow (MBF)
Lower
Higher
<0.001
Myocardial Flow Reserve (MFR)
Lower
Higher
<0.001
Stress Ejection Fraction
Lower
Higher
0.011 - <0.001
Coronary Artery Calcium (CAC) > 400
49.5% (Training), 38.7% (Testing)
Lower percentages
Not specified
AI Model AUC
0.83 (95% CI 0.81–0.85)
NA
Superior to clinical score and individual biomarkers
Key Findings
The AI model integrating 10 PET MPI parameters achieved an AUC of 0.83 in external validation, outperforming clinical scores (AUC 0.80), ischemic TPD (0.79), MFR (0.75), stress MBF (0.75), and CAC (0.69).
Patients with CAD showed significantly higher ischemic and stress TPD, TID ratio, and lower stress MBF, MFR, and stress ejection fraction compared to non-CAD patients.
The AI model demonstrated significantly higher sensitivity than quantitative thresholds (e.g., MFR < 2, stress MBF < 1.8, ischemic TPD ≥ 5%) at matched specificity.
The model effectively identified more high-risk patients classified by Duke 5 and Duke 6 criteria than individual quantitative measurements.
Use of tracer-specific normal limits and kinetic models ensured compatibility across different PET tracers and sites, supporting model generalizability.
Comparison between XGBoost and logistic regression models showed similar diagnostic performance, confirming robustness of the AI approach.
Clinical Implications
This AI model offers a transparent and interpretable tool that integrates multiple PET/CT biomarkers to improve CAD diagnosis beyond conventional clinical assessment. Its multicenter validation supports applicability across diverse clinical settings, potentially aiding clinicians in risk stratification and decision-making. Automated integration of perfusion, flow, functional, and calcium data may streamline workflow and enhance diagnostic confidence.
Conclusion
The interpretable AI model leveraging standard PET myocardial perfusion parameters significantly improves diagnostic accuracy for coronary artery disease across multiple centers. This approach demonstrates the potential for AI to augment clinical PET MPI interpretation by systematically integrating diverse imaging biomarkers.
References
Zhang, W. (2025) -- Assessment of Interpretable Artificial Intelligence for Diagnosing Coronary Artery Disease Using PET Biomarkers Across Multiple Centers
by Wenhao Zhang, Jacek Kwiecinski, Aakash Shanbhag, Robert J. H. Miller, Shiva Mostafavi, Giselle Ramirez, Jirong Yi, Donghee Han, Damini Dey, Dominika Grodecka, Kajetan Grodecki, Mark Lemley, Paul Kavanagh, Joanna X. Liang, Jianhang Zhou, Valerie Builoff, Jon Hainer, Sylvain Carre, Leanne Barrett, Andrew J. Einstein, Stacey Knight, Steve Mason, Viet T. Le, Wanda Acampa, Samuel Wopperer, Panithaya Chareonthaitawee, Daniel S. Berman, Marcelo F. Di Carli, Piotr J. Slomka
Post–cardiac arrest care is increasingly structured around survivorship, with standardized patient-reported assessments supporting multidisciplinary follow-up.
