Results
A LightGBM model (BS2_LightGBM_V10) built on ten routine laboratory parameters discriminated intracerebral hemorrhage (ICH) from acute ischemic stroke (IS) with an AUC of 0.926 in the test cohort.
- Training data came from 12,213 hospitalized stroke patients (3,251 ICH, 8,962 IS) and 2,893 hypertensive controls at Xijing Hospital, 2013–2023.
- BorderlineSMOTE-2 was applied to address class imbalance, and six supervised machine learning models were trained with ten-fold cross-validation.
- The final model was BS2_LightGBM_V10, using ten laboratory variables selected from three feature selection strategies (RFECV-ADA, Lasso, Boruta).
Results
The ten retained laboratory variables were white blood cell count, neutrophil percentage, cystatin C, uric acid, total protein, potassium, sodium, chloride, fibrinogen degradation product, and D-Dimer.
- All ten are routine clinical chemistry and hematology parameters.
- They were screened with three feature selection strategies: RFECV-ADA, Lasso, and Boruta.
Results
The model also separated each stroke type from hypertensive controls with high AUCs.
- AUC was 0.979 for ICH versus controls in the test cohort.
- AUC was 0.923 for IS versus controls in the test cohort.
Results
In the three-class task (ICH, IS, controls), the model reached an overall accuracy of 79.07%.
- Performance was evaluated by AUC, sensitivity, specificity, and accuracy.
- The three-class accuracy was lower than the pairwise AUCs reported for binary discrimination.
Results
SHAP analysis identified neutrophil percentage, cystatin C, D-Dimer, and total protein as the most influential predictors in the model.
- SHAP was used to interpret feature importance in the final model.
Methods
A temporal validation cohort of patients admitted in 2024 was established to assess the model.
- The cohort comprised 153 ICH and 342 IS patients admitted from January to December 2024.
- The abstract does not report the temporal-cohort performance figures separately from the test-cohort AUCs.
Results
Exploratory Mendelian randomization suggested uric acid has a potential causal association with ICH risk.
- The reported odds ratio for uric acid and ICH was 1.309.
- Mendelian randomization and colocalization were described as exploratory complementary analyses.
Results
Mendelian randomization suggested potential causal associations of uric acid, cystatin C, and potassium with IS risk.
- Reported odds ratios were 1.0008 for uric acid, 1.0011 for cystatin C, and 1.0040 for potassium.
- The effect sizes for IS were very small compared with the ICH–uric acid estimate.
Results
Colocalization analysis identified 18 genes linked to uric acid in the context of ICH.
- This was part of the exploratory genetic analyses complementing the diagnostic model.