A K-nearest neighbors machine-learning model predicted syncope recurrence after patent foramen ovale closure with an AUC of 0.993, sensitivity of 1.000, and specificity of 0.986.
Cohort (n=284)
Can machine learning models accurately predict syncope recurrence after percutaneous closure of patent foramen ovale?
A K-nearest neighbors machine learning model demonstrated excellent internal discrimination for predicting syncope recurrence after PFO closure, identifying syncope inducements, episode burden, and plateletcrit as key predictors.
Estimación del efecto: AUC 0.993
Patent foramen ovale (PFO) has been increasingly associated with otherwise unexplained syncope, but predictors of syncope recurrence after percutaneous closure remain unclear. This study aimed to develop and internally evaluate machine-learning models for predicting post-closure syncope recurrence. We retrospectively analyzed 284 patients with PFO and otherwise unexplained syncope who underwent transcatheter closure between January 2017 and December 2023. Syncope recurrence was modeled as a binary outcome during available follow-up; follow-up duration and time to recurrence were not incorporated into the primary model. Clinical, echocardiographic, laboratory, and procedural variables were extracted. Feature selection was performed using Boruta and least absolute shrinkage and selection operator (LASSO) regression. Ten machine-learning algorithms were trained and internally evaluated using discrimination, calibration, decision-curve analysis, and repeated stratified resampling. Syncope recurred in 41 patients (14.44%). LASSO retained ten predictors, including syncope inducements, preoperative episode frequency, plateletcrit, occluder type, D-Dimer, blood pressure status, syncope attack duration, age, platelet distribution width, and diabetes. Among candidate models, K-nearest neighbors showed the highest apparent performance in the internal hold-out test set, with an AUC of 0.993, sensitivity of 1.000, specificity of 0.986, accuracy of 0.988, precision of 0.929, and F1 score of 0.963. Repeated resampling showed performance variability, suggesting that the near-ceiling single-split performance may be optimistic. SHAP analysis identified syncope inducements, episode burden, and plateletcrit as the dominant contributors to predicted recurrence risk. A K-nearest neighbors model showed excellent apparent internal discrimination for observed syncope recurrence after PFO closure. However, because recurrence was analyzed as a binary endpoint without incorporating follow-up time, temporal and external validation were not performed, and recurrence events were limited, these findings should be considered hypothesis-generating. Prospective studies with standardized follow-up, time-to-event analysis, and external validation are required before clinical implementation.
Wang et al. (Mon,) conducted a cohort in Patent foramen ovale (PFO) and otherwise unexplained syncope (n=284). Machine learning prediction models was evaluated on Syncope recurrence (AUC 0.993). A K-nearest neighbors machine-learning model predicted syncope recurrence after patent foramen ovale closure with an AUC of 0.993, sensitivity of 1.000, and specificity of 0.986.