Multiscore Integration of GRACE, ACEF, and TIMI via Machine Learning for Risk Prediction of 6-Month Mortality in Acute Myocardial Infarction

Technical Summary: Multiscore Integration of GRACE, ACEF, and TIMI via Machine Learning for Risk Prediction of 6-Month Mortality in Acute Myocardial Infarction

Problem Statement
Contemporary risk stratification for acute myocardial infarction (AMI) relies heavily on established scores such as the Global Registry of Acute Coronary Events (GRACE), Thrombolysis in Myocardial Infarction (TIMI), and Age-Creatinine-Ejection Fraction (ACEF). While these tools have evolved, current optimization efforts primarily focus on reweighting existing variables rather than reevaluating the fundamental variable set itself. Furthermore, predictors in these scores are often derived from historical clinical contexts that may not fully reflect contemporary patient profiles or treatment paradigms. Although machine learning (ML) has shown promise in enhancing risk prediction, the optimal strategy for integrating variables from multiple classical scores and selecting the most robust feature set for predicting 6-month mortality following percutaneous coronary intervention (PCI) remains uncertain.

Methodology
This single-center retrospective cohort study utilized data from 883 AMI patients undergoing emergency PCI at the First Hospital of Lanzhou University (derivation cohort) and 185 patients from Qinghai Provincial People's Hospital (external validation cohort). The study integrated variables derived from the GRACE, ACEF, and TIMI scores.

The methodology employed a rigorous, dual-stage feature selection and modeling pipeline:

1. Data Preprocessing: Missing values were addressed using Multiple Imputation by Chained Equations (MICE). Class imbalance was considered, though the primary focus was on feature selection rather than synthetic oversampling, as models using LASSO and Boruta without Adaptive Synthetic Sampling (ADASYN) performed comparably or better in specific metrics.
2. Feature Selection: A two-stage strategy was implemented:
   • Stage 1 (LASSO): Least Absolute Shrinkage and Selection Operator (LASSO) regression with L1 regularization was applied to shrink non-significant coefficients to zero, reducing the initial 20 variables to 10.
   • Stage 2 (Boruta): The Boruta algorithm (specifically BorutaShap) was subsequently applied to the LASSO-selected features to perform secondary screening, identifying the most informative features.
3. Algorithm Comparison: Seven ML algorithms were systematically compared using Grid Search with Cross-Validation (GSCV): Random Forest (RF), LightGBM, AdaBoost, XGBoost, Logistic Regression (LR), K-Nearest Neighbors (KNN), and Deep Neural Networks (DNN).
4. Evaluation Metrics: Model performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUROC) and the Area Under the Precision-Recall Curve (AUPRC), alongside calibration curves and Decision Curve Analysis (DCA). Interpretability was enhanced using SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations).

Key Contributions

• Dual-Stage Feature Selection: The study introduces a robust feature selection framework combining LASSO and Boruta, moving beyond single-method selection to identify a parsimonious set of predictors.
• Multiscore Integration: Rather than relying on a single scoring system, the model integrates variables from three distinct historical scores (GRACE, TIMI, ACEF) to capture a broader spectrum of prognostic factors.
• Algorithmic Benchmarking: The study provides a systematic comparison of seven diverse ML algorithms, demonstrating the superiority of ensemble methods (specifically Random Forest) for this specific clinical task.
• Parsimonious Clinical Model: The final model was distilled down to only three variables, balancing high predictive accuracy with clinical simplicity.

Results

• Final Predictors: The optimal model retained three predictors: Killip class, Left Ventricular Ejection Fraction (LVEF), and Creatinine (CREA).
• Internal Validation: The Random Forest model, optimized with LASSO, Boruta, and GSCV, achieved an AUROC of 0.97 (95% CI: 0.93–1.00) and an AUPRC of 0.68 (95% CI: 0.32–0.93) in the internal validation set.
• External Validation: In the external cohort, the model maintained robust performance with an AUROC of 0.89 (95% CI: 0.74–0.96) and an AUPRC of 0.48 (95% CI: 0.24–0.72). The authors note that the slight decrease in AUPRC in the external cohort may be attributed to heterogeneity in baseline characteristics, particularly differences in Killip class distribution between the two institutions.
• Clinical Utility: Calibration analysis showed good agreement between predicted and observed risks in the low-to-intermediate risk range. Decision Curve Analysis confirmed a consistent net clinical benefit across a wide range of threshold probabilities compared to "treat-all" or "treat-none" strategies.
• Tool Development: A web-based interactive tool was developed using the Streamlit framework to facilitate real-time risk estimation based on the three selected variables.

Significance and Claims
The paper claims that an ML-based fusion model, optimized through sequential feature selection and ensemble learning, provides excellent discrimination for 6-month mortality in AMI patients post-PCI, significantly outperforming conventional risk scores. The study emphasizes that model performance is not solely dependent on the number of variables but on the synergistic effects of feature selection, data characteristics, and algorithm choice.

The authors assert that the resulting three-variable model (Killip, LVEF, CREA) achieves a critical balance between parsimony and predictive power, supporting its potential for routine clinical risk stratification. However, the authors maintain a modest stance regarding generalizability, acknowledging limitations such as the single-center nature of the derivation data, potential heterogeneity between institutions, and the lack of prospective, multicenter validation. They conclude that while the model offers a clinically applicable framework, further validation in larger, diverse cohorts is warranted.