Objective: We aimed to improve the diagnostic accuracy of myocardial perfusion SPECT (MPS) by integrating clinical data and quantitative image features with machine learning (ML) algorithms.
Methods: 1,181 rest (201)Tl/stress (99m)Tc-sestamibi dual-isotope MPS studies [713 consecutive cases with correlating invasive coronary angiography (ICA) and suspected coronary artery disease (CAD) and 468 with low likelihood (LLk) of CAD <5%] were considered. Cases with stenosis <70% by ICA and LLk of CAD were considered normal. Total stress perfusion deficit (TPD) for supine/prone data, stress/rest perfusion change, and transient ischemic dilatation were derived by automated perfusion quantification software and were combined with age, sex, and post-electrocardiogram CAD probability by a boosted ensemble ML algorithm (LogitBoost). The diagnostic accuracy of the model for prediction of obstructive CAD ≥70% was compared to standard prone/supine quantification and to visual analysis by two experienced readers utilizing all imaging, quantitative, and clinical data. Tenfold stratified cross-validation was performed.
Results: The diagnostic accuracy of ML (87.3% ± 2.1%) was similar to Expert 1 (86.0% ± 2.1%), but superior to combined supine/prone TPD (82.8% ± 2.2%) and Expert 2 (82.1% ± 2.2%) (P < .01). The receiver operator characteristic areas under curve for ML algorithm (0.94 ± 0.01) were higher than those for TPD and both visual readers (P < .001). The sensitivity of ML algorithm (78.9% ± 4.2%) was similar to TPD (75.6% ± 4.4%) and Expert 1 (76.3% ± 4.3%), but higher than that of Expert 2 (71.1% ± 4.6%), (P < .01). The specificity of ML algorithm (92.1% ± 2.2%) was similar to Expert 1 (91.4% ± 2.2%) and Expert 2 (88.3% ± 2.5%), but higher than TPD (86.8% ± 2.6%), (P < .01).
Conclusion: ML significantly improves diagnostic performance of MPS by computational integration of quantitative perfusion and clinical data to the level rivaling expert analysis.