Papers

This page contains select research papers, with a focus on graph neural networks and time series signal processing. For a full list of my publications, please see my Google Scholar profile.

Online Graph Learning via Time-Vertex Adaptive Filters: From Theory to Cardiac Fibrillation

We propose AdaCGP, an online algorithm that adaptively estimates graph shift operators from streaming multivariate time series using time-vertex autoregressive models. The method achieves 82% improvement over baseline adaptive VAR models in graph structure recovery and demonstrates clinical utility in tracking ventricular fibrillation dynamics under anti-arrhythmic treatment.

Relational Conformal Prediction for Correlated Time Series

We introduce CoRel, a novel distribution-free conformal prediction method that leverages graph neural networks to quantify uncertainty in correlated time series forecasting by exploiting relational dependencies across sequences.

Learning to Predict Global Atrial Fibrillation Dynamics from Sparse Measurements

This paper introduces a new method to visualise cardiac fibrillation dynamics termed ‘imputation mapping’. Using a graph neural network approach, global atrial fibrillation dynamics are reconstructed from sparse clinical measurements, achieving superior performance in tracking cardiac arrhythmia patterns compared to existing methods.

Improving Diffusion Models for ECG Imputation with an Augmented Template Prior

We introduce PulseDiff, a diffusion-based approach for ECG signal imputation that leverages subject-specific pulse templates as informative priors. Our method outperforms leading deterministic models for short-interval missing data and achieves comparable performance for long-interval data loss, addressing key challenges in mobile health monitoring.

Fair and Skill-Diverse Student Group Formation via Constrained k-Way Graph Partitioning

We introduce a constrained k-way graph partitioning approach for automated student group formation that maximizes skill diversity while ensuring fairness. Using Laplacian eigenmaps for dimensionality reduction of student course marks, our method balances group composition with respect to sensitive attributes while promoting diverse skill sets within each group. Our algorithm is now used routinely by 3 European universities.

Methods of Causal Effect Estimation for High-Dimensional Treatments: A Radiotherapy Simulation Study

We introduce the Causal Adaptive Lasso (CAL) for estimating treatment effects in high-dimensional radiotherapy settings. Our method addresses confounding bias in spatially embedded dose-response relationships, achieving up to four orders of magnitude improvement in MSE compared to standard voxel-based estimators while maintaining minimal bias.