Clinical Feature Engineering Data Quality

A code review surfaces that the 1-hour HR window is computed as [T−60min, T) (exclusive upper bound) in the real-time inference module but as [T−60min, T] (inclusive upper bound) in the legacy batch-scoring path. The training dat…

The agent (1) identifies that the training dataset was built with the inclusive boundary [T-60min, T], meaning the model was trained on that distribution; (2) explains that flipping the real-time path to inclusive would match the batch convention but model revalidation is still required because agg…

A data scientist notices that max-HR-6h values occasionally exceed 250 bpm for patients whose nursing flowsheets document normal rates. Initial investigation suggests the pipeline deduplicates rows first (keeping the first-seen r…

The agent (1) identifies that deduplication runs before outlier rejection and that the keep-first deduplication strategy can promote sentinel or artifact values that would otherwise be rejected; (2) states the correct processing order: ingest raw rows → apply physiologically grounded outlier reject…

A model performance review flags suspiciously high training AUROC and low prospective AUROC, suggesting label leakage. The training code defines the label timestamp T as the time the sepsis bundle order was placed in the EHR. A d…

The agent (1) confirms that including post-treatment HR readings in the training window is a form of label leakage; (2) identifies that T=sepsis-bundle-order-time is itself a leaky anchor — by the time a sepsis bundle is ordered, initial interventions (IV fluids, empiric antibiotics, oxygen) may ha…