To Ensure General Reproducibility of Studies, a Common Ground Is Needed.

1. Reproducibility and Transparency

• Problem: Results based on unpublished code or only roughly defined criteria are challenging to verify or reproduce, undermining good scientific practice.
• Solution: A large-scale harmonised dataset allows independent researchers to replicate studies, confirm findings, and validate methodologies.

2. Generalisability

• Problem: Studies conducted on small or localised datasets may not generalise to diverse ICU populations with varying demographics, conditions, and treatments. Studies conducted on large, but still geographically constricted datasets, suffer the similar fates.
• Solution: A large-scale harmonised dataset includes a wide range of patients from different locations, ensuring that research outcomes are applicable across different settings and populations.

3. Collaboration

• Problem: Fragmented data sources hinder cross-laboratory collaboration, as researchers cannot directly and easily compare or combine their findings.
• Solution: A common dataset fosters collaboration across institutions and disciplines by serving as a shared resource, reducing duplication of effort and enabling synergistic progress.

To Build a Common Ground, a Common Language Is Needed.

SNOMED, LOINC, RxNorm and similar vocabularies are great for ensuring that similar concepts are treated similarly. At the same time, due to their complexity and granularity a lot of practical issues arise.

If one determines a monitor signal to reflect Heart rate.beat-to-beat by EKG (LOINC code: 76282-3), is this the same as Heart rate (SNOMED ID: 364075005)?

For practical purposes — e.g. the attending physician stands at the bedside and looks at the monitor to determine the vital state of the ICU patient — there is no real difference. Unfortunately, one cannot directly translate between these vocabularies. Thus, manual harmonisation is needed.

A Common Language Is Only Possible if One Gets Back to Basics

Pre-calculated scores and values within EHR systems entail the risks of biases. Without specific knowledge about which parameters from which systems are taken into account for the calculation, there will be slight differences between institutions (unpublished but referenced data points notwithstanding).

To combat this, reprodICU builds upon the concept of

Everything that can be derived, should be derived.

Instead of taking, for example, a pre-calculated SOFA score, the score is calculated after harmonisation for all individual ICU stays exactly the same way. Sepsis is determined for every patient exactly the same way. Comorbidity indices are calculated for every patient exactly the same way. And so on…