Prediction of myocardial ischaemia in position emission tomography with clinical data, coronary artery calcium score and machine learning algorithms. (25th November 2020)

• Record Type: Journal Article
• Authors: Clerc, O.F; Caobelli, F; Haaf, P; Zellweger, M.J
• Is Part Of: European heart journal. Volume 41:(2020)Supplement 2
• Publication Date: 2020-11-25
• Subjects: Imaging: Coronary Artery Disease; Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/ehjci/ehaa946.0320
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background: With the present evolution of demographics and medical costs, algorithms able to predict the likelihood of coronary artery disease (CAD) may be valuable gatekeepers before advanced cardiac testing, but their value is unknown. Purpose: We assessed whether clinical data, with or without coronary artery calcium score (CACS), combined with machine learning algorithms can predict ischaemia in position emission tomography (PET). Methods: All consecutive patients undergoing 82Rb myocardial perfusion PET with CACS from 2016 to 2019 at our hospital were enrolled. Patients with known CAD were excluded. As clinical data, we used demographics, classical risk factors, chest pain type, dyspnoea and electrocardiogram with pathological Q waves or abnormal repolarization. CACS was measured from low-dose computed tomography scans in Agatston units (AU). Ischaemia was defined as a summed difference score ≥4 in PET. As prediction models, we compared the classical logistic regression, regularized regression (elastic net), decision trees, random forests and a boosting algorithm. The patient cohort was split into a 80% training cohort and a 20% test cohort. Models were carefully optimized with 10-fold cross-validation on the training cohort and applied to the test cohort to assess area under the curve (AUC) for ischaemia with 95% confidence intervals in receiver operating characteristic (ROC) analysis. Results: We included 927 patients, of which 743 were in the training cohort and 184 in the test cohort. Mean age was 64.3±11.1 years, median CACS 55 AU (interquartile range 358) and 179 patients had ischaemia (19.3%). In the test cohort, we found the following AUC (see Figure): logistic regression with clinical data 0.62 (0.52–0.72, light blue) and CACS 0.73 (0.63–0.82, dark blue) with P=0.012, elastic net regression with clinical data 0.63 (0.54–0.73, light violet) and CACS 0.81 (0.75–0.88, dark violet) with P<0.001, decision trees with clinical data 0.5 (0.5–0.5, light green) and CACS 0.76 (0.68–0.84, dark green) with P<0.001, random forests with clinical data 0.59 (0.48–0.70, light orange) and CACS 0.76 (0.69–0.84, dark orange) with P=0.001, and boosting algorithm with clinical data 0.68 (0.58–0.77, light red) and CACS 0.84 (0.77–0.90, dark red) with P<0.001. Thus, adding CACS substantially improved each prediction model. Compared with classical logistic regression, prediction using clinical data tended to improve with the boosting algorithm (P=0.09). For prediction including CACS, both elastic net regression and the boosting algorithm were superior to logistic regression (P=0.007 and P<0.001). Conclusion: Prediction of myocardial ischaemia is limited with clinical data only, but can be greatly improved with the use of CACS and machine learning with elastic net regression or a boosting algorithm, reaching a good performance. This approach could be further developed as a gatekeeper to select appropriate patients for further cardiac imaging. Funding Acknowledgement: Type of funding source: None