Background: Vasovagal syncope (VVS) is the most common form of syncope, accounting for more than half of all syncope cases. At least 35% of people between the ages of 35 and 60 have experienced VVS at least once in their lives. The Head-Upright Tilt Test (HUTT) is commonly used to elicit VVS while monitoring clinical signs and changes in heart rate (HR) and blood pressure (BP). The current protocol is both uncomfortable and time-consuming with suboptimal diagnostic yield. This study aims to develop and evaluate a machine learning (ML) pipeline capable of providing an early prediction of whether a patient with suspected VVS will experience syncope during HUTT, while also identifying features that contribute to a better pathophysiological understanding of VVS.

Methods: The study included 434 adult patients with suspected VVS who underwent HUTT. A ML pipeline was developed for the early prediction of VVS, classifying patients into syncope and no syncope groups. Based on the results of the preliminary study, raw continuous BP, HR and Electroencephalogram (EEG) data were separated into 3-minute epochs before and after the tilt. Linear interpolation was selected as the most effective method to fill in missing data points. The Python package ‘tsfresh’ was used for feature extraction and the Boruta algorithm for feature selection. Three classification models were tested: Random Forest (RF), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost). Model performance was evaluated using 5-fold cross-validation with AUC as the primary metric. In addition, a SHapley Additive exPlanations (SHAP) analysis was performed to assess the importance of features, which provided insight into the features that contributed most to the model's predictions. These features were further analyzed to determine the significant difference between both the syncope and no syncope groups.

Results: The best performing model was the RF model with 61% AUC, 70% sensitivity, and 45% specificity. All selected features contributed to the classification of the syncope and no syncope groups across all folds. Five features were selected repeatedly during cross-validation, including three stroke volume index (SVI) features and two HR features. The two most frequently selected features were the after SVI minimum and the before HR partial autocorrelation features. In particular, the after SVI minimum feature was identified as a valuable feature in every fold. Both features showed a significant difference, suggesting that higher values for these features were associated with a greater likelihood of not experiencing syncope during HUTT.

Conclusion: This study introduced a novel approach to the early prediction of VVS during HUTT. By developing an automated ML pipeline, we demonstrated that with hemodynamic features from the 3 minutes before and after the tilt can provide predictive insight into the occurrence of syncope 20 to 30 minutes later, although with limited sensitivity and specificity. Future research should focus on methods for handling artifacts and outliers and improving model robustness and accuracy. Ultimately, early detection of syncope has the potential to make HUTT more efficient, reduce patient discomfort, and avoid unnecessary testing.