Recorded Norepinephrine Restart Estimates in Intensive Care Databases: Retrospective Study of Episode Matching and Index Times (Preprint)

BACKGROUND Electronic health records contain medication orders and infusion observations with different temporal meanings. Reconstructing a stopping event requires decisions about episode matching and the start of follow-up. OBJECTIVE To assess how event matching and follow-up origin affect norepinephrine stopping classifications and recorded restart estimates in intensive care databases. METHODS In this retrospective methodological study, we audited source-defined labels in MIMIC-IV, eICU and HiRID. We compared reciprocal maximum-overlap and terminal-overlap matching within common eICU stays, decomposed segment and clock contributions, and varied gap thresholds and pair-quality restrictions. Supplementary MIMIC-IV analyses masked pre-landmark stop evidence while keeping patients and follow-up constant. We estimated intervals using clustered bootstrap resampling. RESULTS Excluding single-observation segments and zero-duration overlaps before rematching retained 4,596 eICU stays from 52 hospitals. At the 120-minute threshold, 24-hour order-minus-infusion recorded restart contrasts were −28.18 and −11.23 percentage points under maximum-overlap and terminal-overlap matching. The between-policy change was 16.95 points (95% interval 14.97–18.95). Among 1,207 stays common to four gap thresholds, changes ranged from 2.57 to 24.11 points. In 11,384 MIMIC-IV stays at the 24-hour landmark, 30% episode-hash masking changed label prevalence by −12.47 points and the 28-day death risk-difference contrast by −0.47 points (−1.49 to 0.53). CONCLUSIONS Matching and observation clocks changed recorded restart estimates within the same records. Segmentation defined both the index event and subsequent restart; neither policy was a clinical reference standard. Event definitions should specify segment construction, reciprocal matching and observation origin. These comparisons do not verify cessation or estimate withdrawal treatment effects. Changes in mortality estimates were small under the tested masks.