Lower Extremity Bypass Surveillance and Peak Systolic Velocities Value Prediction Using Recurrent Neural Networks.

Area of Science:

• Vascular Surgery
• Biomedical Engineering
• Machine Learning

Background:

• Routine duplex ultrasound surveillance is crucial for monitoring lower extremity bypass grafts.
• Current methods lack a systematic approach for analyzing peak systolic velocities (PSVs) to predict graft status.

Purpose of the Study:

• To explore the use of recurrent neural networks (RNNs) for predicting future PSVs and identifying bypass graft occlusion.
• To develop and compare RNN models for predicting stenosis and occlusion based on historical PSV data.

Main Methods:

• Developed sequence-to-sequence RNN models, including BiGRU and BiLSTM, to forecast PSVs.
• Utilized 5-fold cross-validation to evaluate model performance based on one to three prior PSV sets.
• Assessed the impact of increasing duplex ultrasound exam data on prediction accuracy.

Main Results:

• The BiGRU model demonstrated superior performance over BiLSTM when using two or more PSV sets.
• Prediction accuracy improved, and error rates decreased with the inclusion of more historical PSV data.
• The study highlights the potential of RNNs in predicting graft occlusion and stenosis.

Conclusions:

• Recurrent neural networks show promise for enhancing lower extremity bypass graft surveillance.
• Integrating PSVs with clinical data can further improve predictive capabilities for graft health.
• This approach offers a systematic method for analyzing duplex ultrasound data to detect early signs of graft failure.