Automated lead detection and chest X-ray features significantly improve ability of machine learning models to predict risk of major adverse events related to transvenous lead extraction

Vishal S. Mehta, YingLiang Ma, Nadeev Wijesuriya, Felicity DeVere, Mark K. Elliott, Steven A. Niederer, Reza Razavi, Christopher A. Rinaldi

Research output: Contribution to journalAbstract


Background: Multiple risk models have been proposed to predict risk of adverse events during and following transvenous lead extraction (TLE).  
Objective: To evaluate whether integrating imaging data and a novel method to determine SVC location, lead location and lead type on a plain anterior-posterior (AP) chest x-ray (CXR) to existing machine learning (ML-) models, increases the ML-model's ability to predict major adverse events (MAE). MAE was defined as procedure-related major complications and procedure-related deaths.  
Methods: We hypothesised that the following features extracted from a plain AP CXR performed prior to TLE would be predictive of MAE: i) lead angulation ii) the percentage of coil inside the superior vena cava (SVC), iii) the number of overlapping leads in the SVC. To determine the approximate location of the SVC on a plain AP CXR, a 3D average heart model from CT scans of a subset of patients was created and overlaid with the plain CXR image. A deep-learning neural network was used to detect the location of the heart and the location of the SVC. Alongside this, we trained a U-Net model to automatically detect the leads and coils on an AP CXR and training data were created from the manual segmentation of the lead and coils in the CXR images.  
Results: 1053 cases that underwent TLE with AP CXRs were included in the final analysis. 24 (2.3%) of these cases experienced a MAE. The following features significantly improved risk prediction: i) >50% of coil within the SVC, ii) 3 overlapping leads in the SVC, iii) acute LV lead angulation. Using the imaging features increased balanced accuracy to predict a MAE from 0.74 to 0.91. Recall in unbalanced data was 0.88 (the new model could detect 88% of cases with MAE). 10-fold cross validation using XGBoost demonstrated a mean accuracy of 95.7% (standard deviation 0.47%).

Conclusion: Integration of imaging biomarkers from a plain AP CXR into risk prediction tools increases the ability of ML models to predict risk of MAE following TLE. Using these novel techniques in more complex images with richer data such as CT or fluoroscopy may provide further improvements.
Original languageEnglish
Article numberCI-452770-2
Pages (from-to)S86
Number of pages1
JournalHeart Rhythm
Issue number5 Supplement
Publication statusPublished - 19 May 2023

Cite this