Learning Normal Representations for Blood Biomarkers

The real contribution isn't scale (2 billion measurements) but the statistical framework for distinguishing signal from noise in sparse, noisy time series without triggering false positives. That's a calibration problem, not a data problem.

This connects directly to the tabular foundation models distillation work from May: both tackle the gap between what works in research and what actually deploys in regulated healthcare settings. Where that story showed how to compress models for speed and fairness guarantees, this one shows how to compress reference ranges from population-level to patient-level without overfitting. The Kalman filter paper on learned memory attenuation also shares the core insight: classical statistical systems (fixed reference ranges, fixed forgetting factors) break when real-world conditions vary per individual, and learned adaptation fixes that without abandoning interpretability.

If this approach ships in a major EHR system or clinical lab network within 18 months and reduces unnecessary follow-up orders by >10% while maintaining sensitivity on true positives, it confirms the model generalizes beyond research cohorts. If adoption stalls or sensitivity drops in production, the overfitting risk was real despite the paper's claims.