Background: Machine Learning (ML) constitutes a revolutionary path to elucidate complex patterns from data in order to optimize prediction and can be applied at different levels of data integration. Cardiovascular imaging classification represents a well suited purpose for ML implementation. Although ML has been increasingly applied to diagnostic classification in non-invasive cardiac imaging, implementation in prognostic data is currently lacking. A sequential approach in hybrid PET/CT for suspected CAD (i.e. initial CTA with selection for further stress myocardial perfusion PET within the same imaging session) has demonstrated diagnostic and prognostic efficacy. However, traditional statistical analyses of these data may be unable to exploit the full potential of extensive structured clinical, CTA and PET data. Therefore, we aimed to implement a stepwise ML workflow considering clinical and hybrid PET/CT data for the identification of patients suspected with CAD who developed MI or death in a long-term follow up registry.
Methods: Data from 951 symptomatic patients with suspected CAD that underwent sequential 15O-water PET/CTA and completed an average follow-up of 6 years were analyzed. Clinical data on demographics and risk factors were extracted from the electronic patient records (sex, age, smoking, diabetes, hypertension, dyslipidemia, fam history, chest complaints and dyspnea). CTA images were evaluated segmentally in terms of: system dominance, segment anatomy, the presence of an atherosclerotic plaque, % stenosis and plaque calcification. Stress PET perfusion data were evaluated regionally (LAD, LCx and RCA) in absolute terms (ml/g/min). Prior feature selection, modeling was conducted utilizing 10-fold cross validated boosted random forests (RF) ensembles to process structured clinical, PET and CTA data. Predictive performance for the development of MI or death was evaluated through AUCs and accuracy.
Results: There were 525 women and 426 men with a mean age of 61±9 years. 24 MI and 49 cardiac death events were documented during follow-up (range: 1 month – 9.6 years), while 109 patients underwent early revascularization. Boosted RF ensembles predictive performance was discrete for clinical data (AUC = 0.65, Acc = 90%) and moderate for clinical + quantitative PET data (AUC = 0.69, Acc = 92.5%), while there was significant performance improvement (p=0.005) when integrating clinical + quantitative PET + CTA data (AUC = 0.82, Acc = 95.4%) in the identification of patients who experienced MI or death during follow up independently from early revascularization.
Conclusion: Stepwise ML implementations for the integration of clinical and hybrid sequential cardiac PET/CT data can improve the identification of symptomatic patients with suspected CAD at who will develop MI or death in long-term follow up. This supports the added prognostic value of ML in cardiac hybrid imaging. Further research into the real-world clinical value of such estimations is warranted.