Development and validation of a blood biomarker-based model for differentiating stroke etiology in acute large vessel occlusion

Figure 1. Variable importance analysis using the Boruta algorithm for LVOS etiology prediction. (a) Boxplot showing the relative importance of laboratory parameters and composite biomarkers based on Boruta algorithm analysis. Green boxes represent confirmed important variables, yellow boxes represent tentatively important variables, and red boxes represent rejected variables. (b) Time series plot showing the convergence of importance scores over 100 iterations of the Boruta algorithm for confirmed (green), tentative (yellow), and rejected (red) variables.

Figure 2. LASSO regression analysis for variable selection in the prediction model. (a) LASSO coefficient profiles of the candidate variables plotted against the log(λ) values. Each colored line represents a variable’s coefficient path. The vertical dashed line represents the optimal λ value selected through cross-validation. (b) Cross-validation error curve showing the binomial deviance (±1 SE) against log(λ). (c) Bar plot showing the standardized coefficients of the six variables selected by LASSO regression at the optimal λ value. Variables are ordered by the absolute magnitude of their coefficients, with sex showing the strongest association.

Figure 3. Nomogram and web-based calculator for predicting cardioembolism in acute LVOS patients. (a) Nomogram for estimating the probability of cardioembolism. Points are assigned for each variable by drawing a vertical line from the variable value to the “Points” line. The sum of points plotted on the “Total Points” line corresponds to the predicted probability of cardioembolism on the “Risk of Y” line. (b) Screenshot of the web-based interactive calculator (available online at https://gaoww.shinyapps.io/dynnomapp/). The tool provides real-time probability estimates with 95% confidence intervals based on input laboratory values. The interface allows for rapid clinical assessment through slider-based input and immediate visual feedback.

Figure 4. Receiver operating characteristic curves showing the discriminative performance of predictive variables and the integrated model. (a) ROC curves for individual predictive variables. (b) ROC curves comparing model performance in training and internal validation cohorts.

Figure 5. Calibration plots and decision curve analyses for model evaluation in training and validation cohorts. (a) Calibration plot for the training cohort (n = 291) showing excellent agreement between predicted and observed probabilities of cardioembolism. (b) Calibration plot for the internal validation cohort (n = 124). Although showing slight deviation at extreme probabilities, the model maintains good calibration across most of the probability range. (c) Decision curve analysis for the training cohort demonstrating the net benefit of the prediction model (blue line) compared to the strategies of treating all patients as CE (red line) or none as CE (green line). The model shows clinical utility across threshold probabilities of 20%–75%. (d) Decision curve analysis for the internal validation cohort confirming the model’s clinical utility, with consistent net benefit patterns observed in the training cohort.