Automatic quantification of pressure waveform damping during cardiac catheterisation using artificial intelligence

Abstract Background Undetected abnormalities in the aortic pressure waveform caused by damping of the signal can lead to complications during coronary angiography. Most commonly caused by over engagement of the catheter into the coronary ostium, the severity of pressure damping can be visually quatified by experience operators: it starts with the loss of the dicrotic notch, progresses to varying degrees of waveform "ventricularisation" and finally reaches a severe form with loss of the typical coronary waveform. The clinical consequences of undetected pressure damping are proportional to its severity, and include underestimation of stenosis severity during physiological assessment, vessel injury and, catastrophically, ventricular fibrillation. An automated approach to quantify pressure damping severity would improve patient safety by eliminating the need for constant operator attention during the procedure. Purpose To develop an index to automatically quantify the degree of pressure waveform damping using artificial intelligence. Methods 761 aortic pressure recordings obtained during pressure wire assessments and ranging from normal to severely damped, were manually processed into 1561 shorter recordings of 6-8 beats each. Recordings were then randomly allocated into separate training and validation datasets in a 70:30 ratio. The training and validation datasets were labelled by two separate groups of experts: 5 labelled the training dataset and 4 labelled the validation dataset. All experts were practicing interventional cardiologists who had undertaken doctoral and post-doctoral research in coronary physiology. 14490 individual expert discriminations were used to train the neural network. A purpose-built platform was used to label pressure waveforms. A neural network was trained using the training dataset. The validation dataset was then assessed by the neural network. The score output was transformed into an index which ranged from 0 (normal waveform) to 1 (severely damped). The correlation between expert consensus and the AI derived index was compared using Spearman’s Rho and numerical agreement assessed with Bland and Altman analysis. Results The AI derived index was able to continuously discriminate progressive degrees of pressure damping and demonstrated excellent correlation with the expert consensus score (Spearman’s rho 0.89) and numerical agreement on Bland and Altman analysis (bias +0.05 and limits of agreement ±0.21). Individual expert’s scores showed a range of agreement with the expert consensus score (Spearman’s rho 0.75-0.94, mean 0.86). The AI index agreed with expert consensus more closely than each expert agreed with each other. Conclusion An AI derived index can provide an automated, real-time and operator-independent quantification of the degree of aortic pressure waveform damping, improving precision of physiological assessment and potentially increasing patient safety during coronary angiography.