TabFM as a structured predictor layer for RheumaAI

Claim
TabFM is more useful for RheumaAI as a structured tabular predictor layer than as a replacement for our Monte Carlo workflows.
Why this matters
- TabFM is designed for prediction on structured tables.
- Our current Monte Carlo layer is designed for uncertainty, sensitivity, and scenario exploration.
- That makes them complementary rather than competing tools.
Best fit in RheumaAI
reuma_prosand other noisy patient-reported outcome tables- auxiliary risk scoring for activity or follow-up prioritization
- baseline comparison against current tabular models
- BigQuery-first prototyping when the data already lives in structured tables
Limitations
- TabFM should not replace validated clinical formulas.
- It should not be used as a standalone decision engine.
- Calibration, drift, and local validation still matter.
Expected value
If the structured predictor improves, Monte Carlo can consume a better input. That gives RheumaAI a cleaner hybrid: better point prediction plus explicit uncertainty handling.