A retrospective evaluation of temporal aggregation and creatinine reuse in critical care records
Year: 2026
Type: preprint
Abstract: Background Research using intensive care records often represents treatment and physiology in fixed time bins and reuses earlier laboratory measurements. Whether these choices preserve the intended measurements requires empirical evaluation. We examined crystalloid exposure reconstruction, recorded-panel availability and agreement between reused and subsequently measured creatinine. Methods This retrospective methodological study used MIMIC-IV 3.1 records from adults with coded diabetes, hypertension and early physiological stress. A fixed 6–12-hour crystalloid dose was reconstructed from four-hour grids with four phase offsets and compared with recorded-event interval overlap. Physiological-panel eligibility required timely documentation and was evaluated under creatinine age limits of 0–24 hours. At observed repeat-test times, the new creatinine was withheld and compared with the latest previously available value. Uncertainty used 2000 patient-cluster bootstrap resamples. A separate eICU 2.0 analysis assessed population-specific interface coverage. Results The master cohort contained 12175 stays in 10183 people. Among 10717 stays observed through 15 hours, fixed-target dose-category disagreement ranged from 5.66% (95% interval 5.21–6.12) to 16.01% (15.30–16.67) across grid phases. Four consecutive eligible panels were available in 146 stays (1.20%) without creatinine carry-forward, 6456 (53.03%) with a 12-hour limit and 8755 (71.91%) with a 24-hour limit. In 35081 paired repeat tests, mean absolute creatinine differences across prior-value age groups of 0–4, > 4–8, > 8–12 and > 12–24 hours were 0.155, 0.169, 0.177 and 0.188 mg/dL. Corresponding proportions differing by at least 0.3 mg/dL were 16.43%, 21.03%, 21.78% and 24.10%; severe creatinine-stage proxy disagreement was nonmonotonic at 3.12%, 3.85%, 3.63% and 3.11%. Conclusions Four-hour aggregation did not reliably preserve a fixed six-hour exposure. Reusing creatinine expanded operational panel availability but showed imperfect agreement at clinically selected retesting times. These complementary analyses do not establish accuracy in unmeasured intervals or a safe universal carry-forward limit. Measurement audits should accompany temporal dataset construction before treatment-response modelling. Clinical trial registration Not applicable. This study did not assign an intervention.
Source: Research Square
Author: Tianyi Yu
Institution: King's College London
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235168978
