Unsupervised Machine Learning Analysis of Invasive Indices of Functional Significance of Coronary Artery Disease: Focus on a Novel Diastolic Pressure Ratio Index
Presenter
Giuseppe Biondi-Zoccai, M.D., Sapienza – Università di Roma, Latina, LT, Italy
Francesco Versaci, M.D.1, Micaela Conte, M.D.2, Marcel van't Veer, M.D.3, Sébastien Lalancette4, Keith G. Oldroyd, M.B.Ch.B., M.D. (Hons)5, Simone Calcagno, M.D.1 and Giuseppe Biondi-Zoccai, M.D.6, (1)Santa Maria Goretti Hospital, Latina, Italy, (2)Clinic Saint Jean, Bruxelles, Belgium, (3)Catharina Hospital, Eindhoven, Netherlands, (4)Opsens Medical, Quebec, QC, Canada, (5)Golden Jubilee National Hospital, Glasgow, United Kingdom, (6)Sapienza – Università di Roma, Latina, LT, Italy
Keywords: Coronary and Imaging & Physiology
Background:
Several invasive indices of the functional significance, including fractional flow reserve (FFR), instantaneous wave-free ratio (iFR), and distal coronary pressure (Pd)/aortic pressure (Pa) ratio (Pd/Pa). We aimed at characterizing the features of a novel non-hyperemic pressure ratio: Opsens diastolic pressure ratio (dPR).
Methods:
Details on patients in whom several functional indices were measured for lesions with angiographically moderate severity were retrospectively collected. In particular, each lesion was appraised with FFR, iFR, Pd/Pa, and dPR. Unsupervised learning analysis was carried out using principal component analysis, hierarchical clustering and k-means clustering using Euclidean distances.
Results:
A total of 525 lesions from 479 patients were included, with a median iFR of 0.92 (1st quartile 0.87, 3rd quartile 0.95), and a median Opsens dPR of 0.92 (0.87; 0.95). Unsupervised learning was performed by appraising correlation with dendrograms and heat map plots of features, confirming the strong association between Opsens dPR and iFR. Principal component analysis was performed, showing that most (97.2%) variability in Opsens dPR, FFR, iFR and Pd/Pa was accounted by the first 2 components. Hierarchical clustering was performed based on such components, identifying several potential clusters, despite the evidence of 3 main ones. Notably, correlation between iFR and Opsens dPR was strong irrespective of the chosen cluster.
Conclusions
:
Unsupervised machine learning analysis confirms the strong association between Opsens dPR and iFR.
