Objective:

To investigate the predictive value of routine laboratory biomarkers, including the Systemic Immune-Inflammation Index (SII), for post-ICH Remote Punctate Ischemic Lesions (RPIL) and to benchmark various machine learning algorithms for optimal predictive modeling.

Approach:

• Study Design: Retrospective cohort study analyzing data from 12,327 patients with ICH, with 6,134 included after exclusions. The dataset was randomly split into training (n = 4,294) and validation (n = 1,840) sets.
• Feature Selection: LASSO regression identified six key predictors: Age, History of Diabetes, SII, D-Dimer, Glucose, and Fibrinogen.
• Machine Learning Benchmarking: Fifteen ML algorithms were benchmarked, with XGBoost achieving the highest discrimination (AUC = 0.799), outperforming Neural Network (AUC = 0.798) and Logistic Regression (AUC = 0.770).
• Model Interpretability: Feature importance analysis and a clinical nomogram were developed for risk stratification, demonstrating excellent calibration (Mean Absolute Error = 0.034) and clinical net benefit.

Key Findings:

• XGBoost model outperformed other algorithms in predicting RPIL, achieving an AUC of 0.799.
• Key predictors included systemic inflammation and hypercoagulability markers.
• The derived nomogram showed excellent calibration and clinical net benefit in Decision Curve Analysis.

Interpretation:

The study presents findings on the predictive value of systemic inflammation and hypercoagulability markers for post-ICH ischemia, with the XGBoost model providing a precise predictive tool.

Limitations:

• The study is based on single-center data, which may limit the generalizability of the findings.
• The retrospective nature of the study may introduce biases in data collection and outcome assessment.

Conclusion:

Future multicenter external validation is necessary to confirm the generalizability and clinical applicability of the predictive models developed.