Probabilistic Modelling

FAIR: Interpretable Latent Severity Scoring for Neurological Disorders

A latent composite severity score with transparent item-level mapping for the objective assessment of neurological disorders

A clearer severity score for Friedreich ataxia that weights each symptom by how informative it really is, and links to daily function.

The problem

Clinical scales add item scores into simple sums, ignoring that items matter differently at different disease stages.

The approach

A Bayesian graded-response Item Response Theory model calibrated on five clinical instruments (105 participants, 467 visits) to derive the FAIR score, which then supervises sensor-based models.

The outcome

r = 0.97 with held-out clinical totals, a stronger link to daily-living function than the standard scale, and sensor-based estimates of r = 0.78 for home monitoring.

In detail

Developed an interpretable severity-scoring framework for progressive neurological disorders, demonstrated in Friedreich ataxia. The work uses Bayesian Item Response Theory to combine information from multiple clinical and functional scales into a single latent score while retaining transparent item-level mappings. The goal is to provide a more principled representation of disease severity than simple summed clinical scores and support objective longitudinal assessment.

Built with

Biomedical AI
Interpretable ML
Item Response Theory