INR was inversely associated and Hct positively associated with stroke status in 29,831 adult ICU patients from MIMIC-IV and eICU-CRD, and an exploratory GBM model for ICU mortality among patients with stroke had modest discrimination that declined in external validation; the authors say these observational associations require cautious interpretation.
Key Findings
Results
Higher INR was associated with lower odds of stroke among adult ICU patients.
The retrospective cohort included 29,831 adult ICU patients from MIMIC-IV and eICU-CRD.
After full adjustment, each 1-unit increase in INR was associated with lower stroke odds, with ORs of 0.105-0.395 across cohorts.
Stroke proportions decreased across increasing INR quartiles.
INR was defined as the first measurement within 24 h of ICU admission, and stroke status was identified using predefined ICD-9 and ICD-10 codes.
Results
Higher hematocrit was associated with higher odds of stroke among adult ICU patients.
After full adjustment, each 1-percentage-point increase in Hct was associated with higher stroke odds, with ORs of 1.076-1.103.
Stroke proportions increased across increasing Hct quartiles.
Hct was defined as the first measurement within 24 h of ICU admission.
Associations were evaluated using multivariable logistic regression, quartile analyses, subgroup analyses and sensitivity analyses.
Results
Restricted cubic spline analyses supported overall associations of INR and Hct with stroke but not consistent nonlinearity after full adjustment.
RCS analyses were used to assess dose-response relationships.
Evidence for nonlinear relationships was not consistent across cohorts after full adjustment.
Results
Multiple-imputation sensitivity analyses gave results directionally consistent with the primary complete-case analyses.
Primary analyses used complete-case cohorts.
Multiple imputation was applied as a sensitivity analysis for missing data.
Results
An exploratory gradient boosting machine model predicting ICU mortality among patients with stroke showed modest discrimination that decreased from training to external validation.
Six prognostic variables were identified by Cox regression in a post hoc exploratory secondary analysis among patients with stroke.
Training AUC was 0.717 (95% CI, 0.669-0.766).
Internal validation AUC was 0.670 (95% CI, 0.603-0.737).
External validation AUC was 0.630 (95% CI, 0.557-0.703).
Conclusions
The authors conclude that the INR and Hct associations with stroke are observational and require cautious interpretation, with the mortality model offering a complementary prognostic perspective.
INR was inversely and Hct positively associated with stroke status.
The ICU mortality analysis was post hoc and exploratory.
What This Means
This study looked at records of nearly 30,000 adults in intensive care units from two large U.S. databases. It asked whether two routine blood measures, INR (a measure of how quickly blood clots) and hematocrit (the percentage of blood made up of red blood cells), were linked to whether a patient had a stroke. Patients with higher INR were less likely to have a stroke diagnosis, while patients with higher hematocrit were more likely to have one. These patterns held after adjusting for other factors and in additional checks, including a method for handling missing data.
The researchers also built a machine learning model to predict death in the ICU among patients with stroke. The model performed modestly, and its accuracy dropped when tested on data from a different database (AUC 0.717 in training, 0.670 in internal validation, 0.630 in external validation). This suggests such a model would need more work before it could be considered reliable.
Because this was a retrospective, observational study, it cannot show that INR or hematocrit cause or prevent stroke. The blood values were measured at ICU admission, and the mortality analysis was exploratory. This research suggests these simple lab values may be associated with stroke status in critically ill patients, but the findings should be interpreted cautiously.
Han J, Zhang L, Lei C. (2026). Associations of International Normalized Ratio and Hematocrit With Stroke Status and Machine Learning Mortality Prediction in Critical Care Databases.. Journal of visualized experiments : JoVE. https://doi.org/10.3791/73411