Enhancing transvenous lead extraction risk prediction: Integrating imaging biomarkers into machine learning models

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Machine learning (ML) models are increasingly used for risk prediction in medical procedures.
• Transvenous lead extraction (TLE) is a complex procedure with associated risks, including major adverse events (MAEs) and long procedural times.

Purpose of the Study:

• To evaluate if incorporating imaging data into an ML model enhances its predictive accuracy for MAEs and lengthy TLE procedures.
• To test the hypothesis that specific radiographic features from chest X-rays can improve risk prediction for TLE.

Main Methods:

• A deep-learning convolutional neural network was developed to automatically detect features from pre-TLE chest radiographs (CXRs).
• Key features analyzed included lead angulation, coil percentage in the superior vena cava (SVC), and overlapping leads in the SVC.
• The ML model's performance was assessed with and without the integration of these imaging biomarkers.

Main Results:

• The deep-learning model achieved high accuracy in detecting cardiac structures and lead features on CXRs.
• Integrating imaging biomarkers significantly improved the prediction of MAEs, with enhanced balanced accuracy, sensitivity, specificity, and AUC.
• Prediction of lengthy procedures (≥100 minutes) also showed significant improvement across all performance metrics.

Conclusions:

• Risk prediction tools that integrate imaging biomarkers substantially improve the ability of ML models to forecast MAEs and long procedural times in TLE.
• Utilizing CXR-derived imaging biomarkers represents a valuable advancement in optimizing patient safety and procedural efficiency for TLE.