Comparator Loss: An Ordinal Contrastive Loss to Derive a Severity Score for Speech-based Health Monitoring
A novel ordinal contrastive loss turns speech recordings into disease severity scores…
A team of researchers (Webber, Watts, Wihlborg, et al.) from the University of Edinburgh and NHS Lothian have introduced a novel approach for monitoring neurodegenerative diseases (NDDs) through speech analysis. Their paper, accepted at Odyssey 2026, proposes the comparator loss — an ordinal contrastive loss function that trains a model to output a continuous severity score. Unlike traditional classification methods that simply discriminate patients from healthy controls, this system learns an ordering relation from easily obtained labels (e.g., diagnosis, clinical scores, or even the chronological order of recordings).
The model's key innovation is its ability to incorporate disparate health metrics, making it highly data-efficient for small clinical datasets. In experiments, the system successfully distinguished NDD subjects from controls and produced severity scores that correlated with unseen expert annotations, such as the ALSFRS-R scale for ALS patients and assessments from speech-language therapists. By using only lightweight annotation (e.g., “this recording is later than that one”), the approach opens the door to scalable, remote health monitoring tools for conditions like ALS, Parkinson’s, and Alzheimer’s.
- Proposes a novel 'comparator loss' — an ordinal contrastive loss that enforces severity score ordering from minimal supervision.
- Trained on lightly annotated data (e.g., chronological order of recordings), yet correlates with clinical scales like ALSFRS-R.
- Accepted to Odyssey 2026, demonstrating potential for scalable, remote monitoring of neurodegenerative diseases via speech.
Why It Matters
Enables early, low-cost detection and tracking of neurodegenerative disease progression from everyday speech.