Prediction of coronary artery disease in positron emission tomography using machine learning algorithms with clinical data and calcium score

Abstract

\n \n \n Type of funding sources: None.\n \n \n \n Patients with clinical suspicion of obstructive coronary artery disease (OCAD) are frequently referred for myocardial perfusion imaging (MPI). However, many are at low-to-moderate risk, and have normal MPI. A predictive model may help to identify patients who may not need MPI.\n \n \n \n To develop novel prediction models for OCAD using machine learning and to compare their performance to a clinical score and to coronary artery calcium score (CACS) to identify patients without OCAD.\n \n \n \n All consecutive patients undergoing 82Rb MPI positron emission tomography (PET) from 2016 to 2020 at our centre and without prior OCAD were enrolled. As clinical data, we recorded age, gender, body mass index, cardiovascular risk factors, chest pain type and dyspnoea. CACS was measured on low-dose computed tomography scans in Agatston units (AU). OCAD was defined as a summed stress score ≥4 in PET. The patient sample was randomly split into a 70% derivation sample and a 30% validation sample. Using 22 different machine learning algorithms, we developed models based on clinical data and based on clinical data + CACS to predict OCAD in the derivation sample, with manual optimization and multiple cross-validation. Then, models were evaluated in the validation sample to select the highest area under the receiver operating characteristic curve (AUC, with 95% confidence intervals). Finally, the best models were compared to the pre-test probability of the European Society of Cardiology (PTP-ESC) and to CACS alone in the complete sample for AUC and for patients predicted as free of OCAD, using cut-offs at 95% sensitivity and Fisher’s exact test.\n \n \n \n We included 1426 patients (1000 for derivation and 426 for validation). Mean age was 64.3 years (standard deviation 11.3), median CACS 58 AU (interquartile range 363), 780 patients were male (54.7%), and 303 had OCAD on PET (21.2%). In model development, extreme gradient boosting (XGBoost) emerged as the best algorithm to predict OCAD with clinical data (AUC 0.74 [0.68-0.80]), and with clinical data + CACS (AUC 0.84 [0.80-0.88], P < 0.001 vs. XGBoost clinical, P < 0.001 vs. logistic regression). In the complete sample, PTP-ESC had an AUC of 0.67 (0.64-0.70), XGBoost clinical 0.75 (0.72-0.78), CACS 0.81 (0.78-0.84), and XGBoost clinical + CACS 0.86 (0.84-0.88), with all P < 0.001 (see Figure). PTP-ESC identified 17% of total patients as OCAD-free with a negative predictive value (NPV) of 93%, XGBoost clinical 24% as OCAD-free with NPV 96%, CACS 33% as OCAD-free with NPV 97%, and XGBoost clinical + CACS 41% as OCAD-free with NPV 97%, with all P < 0.001.\n \n \n \n In patients referred for MPI with PET, machine learning using XGBoost can generate powerful predictive models based on clinical data and CACS to identify patients free of OCAD, better than with PTP-ESC or CACS. Such models may be used as gatekeepers before MPI to reduce radiation burden for patients and costs for the health system.\n Abstract Figure. Prediction of OCAD in complete sample\n