Differentially private and explainable machine learning for vasovagal syncope detection: a feasibility study of homomorphic encryption for secure inference

Abstract

Vasovagal syncope (VVS) is a common cause of transient loss of consciousness and presents diagnostic challenges. The head-up tilt test (HUTT) is the standard method for diagnosis; however, existing approaches often show limited sensitivity, and adoption of Artificial Intelligence (AI)-based tools is constrained by concerns related to privacy and interpretability. This study investigates a proof-of-concept privacy-preserving framework for VVS detection using HUTT-derived physiological signals. A dataset of 137 participants (54 VVS-positive, 83 VVS-negative) was analysed. 54 features were extracted from electrocardiogram (ECG) and blood pressure signals recorded during the HUTT protocol. Logistic regression was trained using differentially private stochastic gradient descent (DP-SGD) with per-example gradient clipping and Gaussian noise injection. For secure inference, fully homomorphic encryption (FHE) using the Cheon–Kim–Kim–Song (CKKS) scheme (TenSEAL) was applied, with a degree-3 polynomial approximation of the sigmoid function for encrypted prediction. Model interpretability was evaluated using LIME and SHAP analyses. A stratified 5-fold cross-validation framework with within-fold pre-processing was implemented to prevent data leakage. Baseline models included logistic regression, random forest, and artificial neural networks (ANN). The DP-SGD model with feature importance selection achieved an accuracy of 0.833 ± 0.035 and an area under the receiver operating characteristic curve (ROC-AUC) of 0.888 ± 0.019 under a formal privacy budget ( ε ≈ 16.5, δ = 10⁻ 5 ). In contrast, FHE-based inference (applied only at the prediction stage) demonstrated substantially reduced performance (ROC-AUC ≈ 0.53) due to polynomial approximation constraints within the CKKS encrypted inference pipeline. Interpretability analyses consistently identified tilt-phase haemodynamic features (SBP_T, DBP_T) and autonomic indices (LFHF_RRI_T, LFHF_SBP_T) as key predictors. This proof-of-concept study demonstrates that differential privacy can be integrated into a VVS detection pipeline with minimal utility loss, while FHE-based inference remains constrained by approximation error. All reported performance metrics represent internal, cross-validated estimates from a single-centre dataset and should not be interpreted as evidence of generalizable clinical performance. Findings are limited by the small single-centre dataset and absence of external validation, and omission of subgroup analysis due to data minimization and sample size constraints. Future work should include multi-centre evaluation and optimized homomorphic inference architectures.