RS24 - Artificial Intelligence Application for the Interpretation of Pulse Volume Recordings

Objectives: Artificial Intelligence (AI) and Deep Learning models are increasingly becoming utilized in medicine for the optimization of processes. Most notably, many of these models have been deployed in the field of radiology for the identification of suspicious findings. We developed an AI/Deep Learning model capable of interpreting Pulse-Volume Recording (PVR) studies.

Methods: PVR studies from a single-institution vascular practice were obtained and interpreted de novo. The data and interpretation were then used to train a convoluted neural network (CNN) to grade and classify PVR waveforms. In parallel, Gradient Boosting was used to develop a protocol to incorporate waveform analysis in the interpretation of segmental pressures and indices for the purposes of identifying segmental disease in the high thigh, low thigh, calf and ankle. These protocols were the combined in series to generate a model capable of interpreting the study as a whole. The model was then tested for accuracy, sensitivity and specificity on a separate set of studies that were not used in the training of the model.

Results: A total of 1193 studies were included to develop the model. For the purposes of training and validating, 930 of these studies were used. The remaining 263 studies were used for testing. The overall accuracy of the model was 92%, with a sensitivity of 81% and a specificity of 96%. The accuracy in detection of aortoiliac or iliofemoral disease was 96% with a sensitivity of 72% and a specificity of 100%. The accuracy in detection of femoral disease was 92% with a sensitivity of 80% and a specificity of 95%. The accuracy in detection of popliteal disease was 89% with a sensitivity of 93% and a specificity of 88%. The accuracy in detection of tibial disease was 90% with a sensitivity of 64% and a specificity of 95%.

Conclusions: AI has the potential to significantly improve and streamline processes within the field of medicine. We demonstrate the feasibility of utilizing AI for the interpretation of PVR studies. The strategies for developing this protocol can be readily applied for the purposes of duplex ultrasound interpretation. Further study will include optimization of the model to enhance accuracy and sensitivity; however, as with all AI medical applications, final interpretations and decisions must always be made under direct physician supervision.