Research
NeuroVoice — can a machine hear Parkinson's disease?
Neurodegenerative diseases leave an audible fingerprint in speech — and the right representations can pick it up. I benchmarked ten self-supervised speech foundation models (WavLM, HuBERT, wav2vec 2.0, TRILLsson) against traditional acoustic features for classifying dysarthria across early- and advanced-stage Parkinson's disease, Huntington's disease, and multiple system atrophy. The best model distinguishes patients from healthy controls from a single spontaneous monologue — only a microphone required, no clinical equipment.
Model selection ran on patient-stratified nested cross-validation scores — guarding against data leakage — and the final evaluation on a held-out test set, where the best configuration, TRILLsson embeddings with logistic regression, reached a Macro F1 of 0.88 on the spontaneous monologue task.
Interesting part was interpreting the embedding dimensions themselves: SHAP, linear probes, and correlation analysis surfaced individual TRILLsson dimensions weakly but significantly tied to known acoustic markers — formant-based measures of vowel articulation above all, with others tracking the long-term average spectrum and pause duration. The models seem to rely on distributed acoustic information that partly overlaps with what clinicians already listen for. Voice-based screening won't replace a neurologist — but it could flag who should see one sooner.
Code on GitHub Read the thesis (PDF)
Hospital-integrated deep learning for real-time glioma analysis with advanced diffusion MRI
Deep learning that runs where it matters — inside the hospital. Selected for presentation at ECR 2026: a real-time system for multi-shell diffusion MRI analysis of gliomas, with a secure PACS-connected workflow and embedded models for molecular prediction and automatic segmentation.
ECR 2026 session