An artificial intelligence model integrating culprit lesion diagnosis and risk assessment for acute coronary syndrome

Figures

• 

  Fig. 1. Study design concept. This study first identified the optimal machine learning (ML) model for detecting culprit lesions associated with acute coronary syndrome (ACS) in the derivation cohorts and selected the random forest (RF) model as the most effective, then evaluated the model's performance to stratify the risk of future ACS events in two independent prospective cohorts with chronic coronary syndrome. IVUS, intravascular ultrasound; HR, hazard ratio; 95% Cl, 95% confidence interval; NRI, net reclassification improvement.

• 

  Fig. 2. Segmentation, model selection, and importance ranking in the derivation cohort. Panel A illustrates the dataset partitioning for the derivation cohort. Panel B compares the performance of stenosis severity and five machine learning algorithms in detecting culprit lesions in acute coronary syndrome, with the random forest (RF) algorithm showing the best performance. Panel C ranks the importance of each feature in the RF model. CCTA, coronary computed tomography angiography; IVUS, intravascular ultrasound; LASSO Logistic, LASSO logistic regression; DT, decision tree; SVM, support vector machine; XGBoost, extreme gradient boosting; CT-FFR, CT-derived fractional flow reserve; ΔCT-FFR, the difference between proximal and distal CT-FFR; LDP, low-density plaque; DC, dense calcium.

• 

  Fig. 3. Interpretability of the random forest model for acute coronary syndrome (ACS) prediction using SHAP (SHapley Additive exPlanations) analysis. Panel A: Beeswarm plot of feature importance. Beeswarm plot summarizing feature importance across the cohorts. Each dot represents a patient, colored by normalized feature value (red: higher, blue: lower). Horizontal position indicates the SHAP value's magnitude/direction (right: increased ACS risk). Key drivers include stenosis severity and the difference between proximal and distal CT-derived fractional flow reserve (ΔCT-FFR). Panel B: Waterfall plot for a representative positive case. Force plot (individual explanation) for a high-risk case predicted as ACS. Positive contributions (red arrows) from stenosis severity (+0.22 SHAP value) and ΔCT-FFR (+0.07) dominate. Panel C: Waterfall plot for a representative negative case. Force plot for a low-risk case predicted as CCS. Protective factors (blue arrows) include ΔCT-FFR (−0.04) and stenosis severity (−0.04). LDP, low-density plaque; DC, dense calcium; CCS, chronic coronary syndrome.

• 

  Fig. 4. Risk stratification performance of the random forest model in cohorts 3 and 4. Panels A and B: Kaplan–Meier curves. Using a threshold of 0.20 generated by the random forest (RF) model, the RF model demonstrated significant differences in the temporal distribution of survival rates between high-risk and low-risk groups in both cohorts 3 and 4. Panels C and D: Time-dependent area under the receiver operating characteristic curve (AUC) curves. These panels illustrate the predictive performance of both the stenosis severity and RF models at various time points in cohorts 3 and 4. Panels E and F: Bar charts illustrating the concordance index (C-index). In Cohort 4, the RF model showed a higher C-index than the stenosis severity model. Panels G and H: Time-dependent net reclassification improvement (NRI) curves. In both cohorts 3 and 4, the RF model demonstrated superior performance in distinguishing between risk classes compared to the model based solely on stenosis severity. HR, hazard ratio; 95% Cl, 95% confidence interval.