Machine learning improved CAD diagnosis sensitivity by 2.1%, specificity by 4.3%, accuracy by 1.3%, and AUC by 2.5%, and increased prognostic AUC by 5.9%.
Does machine learning improve the prediction of obstructive CAD and cardiovascular events compared to non-machine learning methods in patients undergoing SPECT or PET MPI?
Machine learning integration with nuclear cardiology imaging improves diagnostic and prognostic accuracy for coronary artery disease compared to conventional interpretation methods.
Tasa de eventos absoluta: 0% vs 0%
Abstract Background Nuclear cardiology studies, including single photon emission computed tomography (SPECT) and positron emission tomography (PET), use noninvasive radiotracer imaging to assess myocardial perfusion (MPI), blood flow, and heart function at rest and stress. This generates imaging variables that need to be integrated with clinical data to predict the presence of obstructive coronary artery disease (CAD) and risk of cardiovascular events. Artificial intelligence methods, including machine learning (ML), are well-suited for integrating large amounts of clinical and imaging data. Purpose This systematic review compared the performance of ML to conventional non-ML methods for predicting the presence of obstructive CAD and risk of cardiovascular events in patients undergoing SPECT or PET MPI. Methods Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, we performed a systematic literature search using MEDLINE, EMBASE, and Cochrane databases from inception to October 2024. Eligible studies included primary research articles comparing the performance of ML and non-ML methods for predicting obstructive CAD and risk of cardiovascular events in adult patients undergoing SPECT and PET MPI. Invasive coronary angiography (ICA) was used as the gold standard for obstructive CAD diagnosis. Studies on prognosis included any cardiovascular event outcome, such as major adverse cardiovascular events, revascularization, and mortality. ML and non-ML methods were compared using key performance metrics including sensitivity, specificity, accuracy, and area under the receiver operating curve (AUC). Their mean differences and standard deviations were reported. The study is registered with PROSPERO. Results Of 2,976 articles identified, 36 were included in our quantitative review, including 22 studies on diagnosis and 14 on prognosis. In total, 28 studies used SPECT and 8 used PET. ML methods included neural networks (n=24) and gradient/boosted ensemble algorithms (n=12). Non-ML methods included expert interpretation (n=17), semi-quantitative (n=16), and logistic regression (n=3). For diagnosis, 10,349 patients were included, where 5,813 (56.2%) had obstructive CAD on ICA (performed within 6 months of MPI). Compared to non-ML methods, ML improved diagnostic sensitivity by 2.1% (9.7%), specificity by 4.3% (10.1%), accuracy by 1.3% (8.6%), and AUC by 2.5% (5.4%) for CAD (Figure 1). For prognosis, 59,727 patients were included, where 9,079 (15.2%) had a predefined cardiovascular event (follow-up duration range 90 days to 6 years). Similarly, ML improved cardiovascular outcome prediction compared to non-ML methods with an increase in the AUC by 5.9% (3.9%) (Figure 2). Conclusion ML integration with nuclear cardiology imaging can improve disease diagnosis and outcome prediction in CAD. Further clinical and prospective validation of ML methods in nuclear cardiology are needed.ML vs non-ML AUC CAD diagnosis ML vs non-ML AUC CAD prognosis
Dai et al. (Sat,) reported a other. Machine learning improved CAD diagnosis sensitivity by 2.1%, specificity by 4.3%, accuracy by 1.3%, and AUC by 2.5%, and increased prognostic AUC by 5.9%.