In rare neurodegenerative disease the evidence usually exists long before the diagnosis does, sitting in an EEG trace or a clinical note that nobody reads again in that light. We built the detection layer that reads it at population scale, across brainwave data, clinical records and claims.
better early detection from brainwave reports, in pilot
previously invisible populations located and sized by hospital
EEG image processing and clinical text, read together
THE PROBLEM
The signal was recorded years before anyone read it
A rare neurodegenerative diagnosis typically arrives after years of a patient being investigated for something else. That delay is not usually a failure to measure. The early signal is frequently already present, written into an EEG report or a clinical note at a point when nobody had reason to read it as evidence of a rare disease.
At population scale that history is unreadable by hand. So the cohorts stay invisible: the patients exist, the evidence exists, and neither can be pointed to. For a company whose medicines treat those patients, not knowing where they are is the constraint on everything downstream.
THE DETECTION
Reading the trace and the report together
Two modalities, because either one alone is thin. Image processing reads the EEG traces themselves, and text classification reads the reports and clinical notes written around them, extracting the findings and qualifiers a neurologist recorded in prose. Combined against claims history, the layer surfaces patients whose pattern fits a rare neurodegenerative profile before that profile has been named for them.
Rare disease makes this hard in a specific way: positives are scarce, so a model that chases accuracy learns to say no. The work was in holding precision high enough to be worth a clinician’s attention while still recovering patients the prevailing process misses.
THE COHORTS
Where those patients actually are
The same models run as cohort analysis rather than one patient at a time, mapping detected populations across hospital systems in the Atlanta region. That turns a clinical capability into a commercial one: which institutions hold the patients, at what density, and where an early-detection programme would pay for itself.
It answers a question that rare disease usually cannot answer at all, because the denominator was never visible.
THE RESULT
36 percent better, in pilot
The pilot detects early disease from brainwave reports 36 percent better than the prevailing rate. In rare disease that margin is measured in years of diagnostic odyssey, not percentage points on a dashboard.
Cohorts that could not previously be seen can now be located and sized by hospital and region.
Also delivered at UCB
Detection, cohort analysis and the pipeline underneath both.
- EEG image processing over brainwave traces
- Text classification across clinical reports and notes
- Patient and claims data integration for population-scale analysis
- Rare disease early-detection models tuned for scarce positives
- Rare disease cohort analysis across hospital populations
- Regional cohort mapping and sizing for the Atlanta area
Discuss your next system.
Start with a free, one-hour engineering review. Leave with a practical next step.