Figure 1

Study design and analysis workflow. FDR, false discovery rate; HRUS, high resolution ultrasound; IS, ischemic stroke; ICC, intraclass correlation coefficient; ROI, region of interest."

Figure 2

HRUS plaque image processing. (A) shows the selected longitudinal end diastolic image at the maximum plaque area. (B) shows the manually delineated ROI overlaid on the HRUS image. (C) shows the resulting plaque mask. HRUS, high resolution ultrasound; ROI, region of interest."

Table 1

Clinical characteristics of patients without IS, patients with IS in the primary cohort, and patients in the additional IS cohort"

Table 2

Overlap-weighting balance diagnostics for the eight clinical covariates in the primary cohort"

Figure 3

Absolute standardized mean differences before and after overlap weighting. The dashed reference line marks an absolute SMD of 0.10."

Figure 4

Propensity score distributions before and after overlap weighting in the primary cohort. Panel A shows unweighted distributions and panel B shows weighted distributions; blue and red curves represent patients without and with IS."

Table 3

Radiomics features meeting the prespecified reliability and robustness criteria"

Figure 5

Overlap-weighted associations of seven robust radiomics features with ischemic stroke. Points show odds ratios per standard deviation, horizontal lines show 95% confidence intervals, and the dashed line marks an odds ratio of 1. W denotes wavelet features."

Figure 6

LASSO penalty selection by fivefold cross validation. Points show mean AUC and error bars show one standard error. Dashed lines mark the penalty with the highest mean AUC (gray, λ = 1.0000) and the penalty selected by the one standard error rule (red, λ = 4.6416)."

Figure 7

LASSO penalty selection and model size across 1,000 bootstrap samples. Panel A shows λ selected by fivefold cross-validation using the one-standard-error rule; panel B shows the number of selected features, with the median marked by a dashed line."

Figure 8

Selection frequencies of the 20 most frequently selected radiomics features across 1,000 bootstrap samples. Bars indicate the proportion of samples selecting each feature; the dashed line marks 70%."

Table 4

Full pipeline bootstrap internal validation with 1,000 iterations"

Figure 9

Bootstrap optimism in model performance. (A-C) show optimism in AUC, Brier score, and calibration slope across 1,000 bootstrap samples. Dashed lines mark mean optimism; corrected estimates appear below each panel."

Figure 10

Balance of 16 clinical covariates in the expanded model. Points show absolute SMDs before and after overlap weighting; the dashed line marks 0.10."

Table 5

Robust features in the primary and expanded clinical models"

Figure 11

Overlap-weighted associations of seven robust radiomics features in the primary and expanded clinical models. Points show odds ratios per standard deviation, horizontal lines show 95% confidence intervals, and the dashed line marks an odds ratio of 1."

Figure 12

Exploratory associations of the seven robust radiomics features with IS by lipid-lowering medication use. Points show overlap-weighted ORs per SD with 95% CIs; the dashed line marks an OR of 1."

Figure 13

Differences in robust radiomics features between the two IS case series. Bars show SMDs comparing the 83 additional IS cases with the 361 IS cases in the primary cohort; positive values indicate higher feature values in the additional cohort."