Associazione Nazionale Medici Cardiologi Ospedalieri

CONGRESS ABSTRACT

CONGRESS ABSTRACT

PREDICTIVE MACHINE LEARNING MODEL FOR MECHANICAL DILATATION IN TRANSVENOUS LEAD EXTRACTION PROCEDURES

De Lucia Raffaele Pisa(Pisa) – UO Cardiologia 2, Ospedale di Cisanello, Azienda Ospedaliero Universitaria Pisa, Pisa | Micheli Alessio Pisa(Pisa) – Università di Pisa, Dipartimento di Informatica, Pisa | Parlato Alessandro Pisa(Pisa) – Scuola di Specializzazione di Cardiologia, Università di Pisa, Pisa

BACKGROUND
Transvenous lead extraction (TLE) remains a procedure that requires a high level of expertise, with a doubled risk of death and clinical failure when performed in low-volume centers compared to high-volume ones.

PURPOSE
The aim of this study was to create a machine learning (ML)-based risk stratification system for predicting the need for mechanical dilatation in patients undergoing TLE due to infection.

METHODS
We designed a ML-based risk stratification system trained with data from our registry to predict the need for mechanical dilatation in patients undergoing TLE for infection. An extensive evaluation of 5 different ML models (k-nearest neighbors, support vector machine, decision tree, and decision tree ensembles, such as random forest and gradient boosting machine) was conducted to identify a classifier with the highest potential to correctly predict previously unseen patients.

RESULTS
Data to train the model was extracted from our 25-year registry of patients undergoing TLE (June 1998 – March 2023), for a total of 491 patients (77.8% male; age 69.7 ± 12.8 years) and 938 leads (ICD 21.2%; pacing 78.8%; indwelling time 61 ± 60 months) removed with success in 100% of cases. Each patient was represented by a set of 21 attributes (14 clinical, 7 device-related). Manual traction (MT) was used in 27.5% of cases, and mechanical dilatation (MD) was employed in the remaining 72.5% of cases. 5-fold nested cross validation was used to estimate performances: in turn, 393 patients were used for training and model selection, and 98 patients were used for independent testing. According to the evaluation, Gradient Boosting Machine performed best, achieving test accuracy of 89% (+/- 2% std. dev.), test sensitivity of 95% (+/- 3% std. dev.), test specificity of 73% (+/- 8% std. dev.), test AUROC of 92% (+/- 1% std. dev.). A further interpretability analysis on the best performing decision tree was conducted, showing remarkable adherence between the internal decisions taken by the model to make predictions and the current clinical practice for TLE.

CONCLUSION
ML models have enhanced the prediction of MD requirements in TLE procedures. This data may support the referral of patients to specialized and high-volume centers, thereby improving preoperative risk assessment.