Prediction of revascularization after myocardial perfusion SPECT by machine learning in a large population

Abstract

Objective We aimed to investigate if early revascularization in patients with suspected coronary artery disease (CAD) can be effectively predicted by integrating clinical data and quantitative image features derived from perfusion SPECT (MPS) by machine learning (ML) approach. Methods 713 rest 201Thallium/stress 99mTechnetium MPS studies with correlating invasive angiography (372 revascularization events (275 PCI / 97 CABG) within 90 days after MPS (91% within 30 days) were considered. Transient ischemic dilation (TID), stress combined supine/prone total perfusion deficit (TPD), quantitative rest and stress TPD, exercise ejection fraction, and end-systolic volume along with clinical parameters including patient gender, history of hypertension and diabetes mellitus, ST-depression on baseline ECG, ECG and clinical response during stress, and post-ECG probability by boosted ensemble ML algorithm (LogitBoost) to predict revascularization events. These features were selected using an automated feature selection algorithm from all available clinical and quantitative data (33 parameters). 10-fold cross-validation was utilized to train and test the prediction model. The prediction of revascularization by ML algorithm was compared to standalone measures of perfusion and visual analysis by two experienced readers utilizing all imaging, quantitative, and clinical data. Results The sensitivity of machine learning (73.6±4.3%) for prediction of revascularization was similar to one reader (73.9±4.6%) and standalone measures of perfusion (75.5±4.5%). The specificity of machine learning (74.7±4.2%) was also better than both expert readers (67.2±4.9% and 66.0±5.0%, P < 0.05), but was similar to ischemic TPD (68.3±4.9%, P < 0.05). The Receiver-Operator-Characteristics areas-under-curve for machine learning (0.81±0.02) was similar to reader 1 (0.81±0.02) but superior to reader 2 (0.72±0.02, P < 0.01) and standalone measure of perfusion (0.77±0.02, P < 0.01). Conclusion ML approach is comparable or better than experienced reader in prediction of the early revascularization after MPS and is significantly better than standalone measures of perfusion derived from MPS.

ASNC imaging guidelines for nuclear cardiology procedures

“…53,54 Unlike multivariate regression modeling, machine learning algorithms are not fitted models, and thus are not affected by collinearity between variables. Furthermore, they can be improved in an ongoing basis incorporating accumulative observations after clinical implementation.…”

“…55 Furthermore, machine learning applications, integrating clinical, ECG, exercise, hemodynamic, defect quantification, and ancillary imaging data provide a patient-specific estimate of likelihood of early revascularization and allcause mortality, thus aiding in individualized decisionmaking in a way the human brain cannot do. 53,56 Machine learning algorithms are a natural complement to nuclear cardiology analyses packages and structured reporting software, from which multi-faceted data can be derived to generate risk estimates factored in DSTs and patient-centered decision guidance. …”

Artificial Intelligence and Machine Learning in Cardiology - A Change of Paradigm

“…Although the performance of the new system is reliable for predicting mortality and cardiac events in 30 days, the authors themselves acknowledge that clinical decisions are dependent on factors that still can not be fully incorporated into the machines, one of which is the physicians experience. 7 In nuclear cardiology Arsanjani et al 8 evaluated the use of the Machine Learning tool to predict myocardial revascularization from myocardial perfusion scintigraphy data, finding an accuracy comparable or even superior to that of experienced examiners in the interpretation of the scintigraphic examination. Garcia et al, 9 in an excellent review on the subject, point out that clinical decision support and artificial intelligence systems serve as warnings for the cognitive bias of clinicians and reduce intra and interobserver variability, allowing to interpret the exams faster and with greater accuracy, as observed in studies in which the diagnostic interpretation of the examination by the computer is similar to that of the experts.…”

Normal limits for transient ischemic dilation with 99mTc myocardial perfusion SPECT protocols

“…This final computed probability can be then discretized to several categories of risk or probability of disease, rather than just assigning a normal/abnormal finding. Indeed, such holistic approach has been shown to be very promising in diagnostic 14 and prognostic 15,16 applications demonstrating a significant overall improvement of diagnostic accuracy or risk reclassification. …”