Publications

Using trajectory analysis on longitudinal data from over 1,000 MS patients, this study identifies previously unrecognized progression patterns and treatment effects, demonstrating how data-driven methods can generate novel clinical hypotheses.

This study shows that combining MRI and evoked potentials improves machine learning prediction of disability worsening in multiple sclerosis, enabling better patient management and personalized treatment strategies.

We show that Ising machines can be used to perform Model Predictive Control with sampling-based optimization tested on a kinematic bicycle model.

Hand-crafted radiomic features from multicenter FLAIR MRI predict disability progression in MS patients, enabling personalized treatment planning and improved outcomes.

This paper outlines how trustworthy and reliable AI can improve multiple-sclerosis research and clinical decision support.

Recursive spiking neural networks estimate apnea-hypopnea event counts for sleep-assessment applications.

Pruning reduces memory footprint for radar-based human activity recognition on constrained devices.

Quantization of the LSTM states of a Split BiRNN model reduces radar data transmission costs for edge-cloud activity recognition while preserving task performance.

This workshop paper explores trustworthy neural-network methods to support certification-by-analysis for Flammability Reduction Systems in aerospace systems with uncertainty quantification.

An international multi-center ML study predicts multiple-sclerosis disability progression from real-world clinical data.

Data-driven surrogate models accelerate analysis of fuel-tank flammability reduction system behavior.

Conformal prediction improves reliability by producing valid uncertainty bounds for NILM appliance recognition.

Open-set radar activity recognition detects known patient activities while flagging unseen behaviors.

Utility maximization tailors radar patient-monitoring models to practical clinical priorities and deployment needs.

This work adds real-time uncertainty estimates to Split BiRNN radar activity recognition for more reliable decision-making.

CNN-based apnea detection from smartphone audio is sensitive to analysis window size, with clear performance trade-offs.

Bayesian deep learning provides calibrated uncertainty estimates for appliance recognition in non-intrusive load monitoring.

Split BiRNN enables real-time radar activity recognition by distributing computation while maintaining predictive performance.

Deep Learning-based human activity recognition models classify patient activities with radars to support unobtrusive and privacy-preserving monitoring in healthcare settings.