Patients at risk of recurrent syncope after patent foramen ovale (PFO) closure could soon be identified using machine learning, according to new research that developed a highly accurate predictive model based on clinical and procedural factors.
Understanding Patent Foramen Ovale and Syncope
PFO, a small opening between the upper chambers of the heart that fails to close after birth, has increasingly been linked to otherwise unexplained syncope.
While transcatheter PFO closure can reduce recurrent fainting episodes in many patients, clinicians have lacked reliable tools to identify who remains at risk of recurrence following the procedure.
A new study sought to address this gap by developing machine-learning models capable of predicting syncope recurrence after PFO closure.
Machine Learning Identifies Patients at Risk After PFO Closure
Researchers retrospectively analysed 284 patients with PFO and unexplained syncope who underwent transcatheter closure between 2017–2023.
During follow-up, syncope recurred in 41 patients, representing 14.4% of the study population. The investigators evaluated a broad range of clinical, echocardiographic, laboratory, and procedural variables to identify factors associated with recurrence.
Using feature-selection techniques, the researchers narrowed the model to 10 key predictors. These included syncope triggers, frequency of episodes before treatment, plateletcrit, occluder type, D-dimer levels, blood pressure status, duration of syncope attacks, age, platelet distribution width, and diabetes status.
K-Nearest Neighbours Model Delivered Strongest Performance
The team trained and tested 10 different machine-learning algorithms. Among them, a K-nearest neighbours (KNN) model demonstrated the strongest apparent performance in the internal test dataset.
The model achieved an area under the curve (AUC) of 0.993, alongside a sensitivity of 100%, specificity of 98.6%, and overall accuracy of 98.8%. These findings suggest that machine learning may be capable of identifying patients at particularly high risk of recurrent syncope following PFO closure.
Further analysis using SHapley Additive exPlanations (SHAP) highlighted syncope inducements, preoperative episode burden, and plateletcrit as the most influential contributors to recurrence predictions.
Important Limitations Remain
Despite the promising results, the authors stressed that the model remains preliminary.
Syncope recurrence was analysed as a simple binary outcome rather than incorporating follow-up duration or time-to-event data. In addition, the study included a relatively small number of recurrence events and lacked external validation.
The researchers also observed variability in model performance during repeated resampling analyses, suggesting that the exceptionally high accuracy seen in the initial test set may be overly optimistic.
As a result, the findings should be considered hypothesis-generating rather than ready for clinical implementation.
Larger prospective studies with standardised follow-up, time-to-event analyses, and independent external validation will be needed before machine-learning tools can be routinely used to guide care for patients undergoing PFO closure.
Reference
Wang X et al. Prediction of syncope recurrence after percutaneous closure of patent foramen ovale based on machine learning. Sci Rep. 2026;DOI:10.1038/s41598-026-54910-5.
Featured image: sorapop on Adobe Stock
