MIMIC-IV: Interpreting storetime < charttime for vital signs and urine output

[rladudwns559-cmyk]

Hello MIMIC team and community,

I am an undergraduate student at Hanyang University ERICA, conducting an academic machine learning project using MIMIC-IV v3.1.

Our project predicts 28-day mortality using clinical measurements available within the first 24 hours after ICU admission.

While examining the CHARTEVENTS and OUTPUTEVENTS tables, we observed records where storetime < charttime.

I have read the official MIMIC-IV Core Concepts documentation and understand that this ordering can occur for observations such as GCS. However, I am unsure how to interpret it for other vital signs and urine output records.

My questions are:

1. Are there known recording or validation processes that explain storetime < charttime for heart rate, blood pressure, or oxygen saturation?
2. Can the same timestamp ordering occur normally for urine output measurements in OUTPUTEVENTS?
3. Does storetime reliably represent when the final recorded value first became available, or could later modifications be reflected in the released data?
4. For a prediction model at ICU admission +24 hours, is it reasonable to require both charttime and storetime to be no later than the prediction time? Are there more appropriate approaches?

In our development-data sensitivity analysis, excluding timestamp-inverted records changed some model predictions. However, we found no direct evidence of information from after the prediction time being used.

We are not assuming these records are errors. We would like to understand the timestamp semantics and choose a defensible temporal preprocessing strategy.

Any guidance or references would be greatly appreciated.

Thank you!