Predicting short-term outcomes in atrial-fibrillation-related stroke using machine learning

Figure 1. Flowchart of study subjects.

Figure 2. Receiver operating characteristics curves and precision–recall curves illustrating model performance. Solid lines and shades represent mean curves and 95% confidence interval areas, respectively. For the baseline model (logistic regression, “LogReg”), confidence intervals are represented with a polka dot pattern. An asterisk (*) indicates significantly higher AUROC than the baseline model (p < 0.05, Benjamini–Hochberg corrected).

Figure 3. AUROC and precision–recall curves representing performance on different subgroup cohorts for the prediction of unfavorable functional outcomes. The (n = A, B%) notation for each subgroup indicates the number of samples in the test set (A) and prevalence rate of the outcome (B) of the subgroup. Box plots are plotted with whiskers of 1.5 times the interquartile ranges. AUPRC, area under the precision–recall curve; AUROC, area under the receiver operating characteristics curve; DM, diabetic mellitus; HTN, hypertension; PAF, paroxysmal atrial fibrillation (AF); PeAF, persistent atrial fibrillation (AF).

Figure 4. Importance of selected variables for prediction of unfavorable functional outcomes (A,B) and mortality (C,D). (A,C) Individual influences of every value and overall contributions to the model prediction of the top 10 variables are represented as a dot on the right and bar on the left, respectively. In the plot on the right, red dots indicate high values in continuous/ordinal variables. Positive and negative SHAp values indicate positive contributions resulting in higher prediction scores and negative contributions resulting in lower prediction scores, respectively. (B,D) Partial SHAP dependence plots for four representative variables. Histograms on the right and the top axes of each plot indicate SHAP distributions and variable values, respectively. The original labels of the numeric codes are as follows: 1, single corticosubcortical; 2, cortical; 3, subcortical (≥ 15 mm); 4, subcortical (< 15 mm); 5, small scattered lesion in one vascular territory; 6, confluent and an additional lesion in one vascular territory; and 7, multiple lesions in multiple vascular territories. BMI, body mass index; CrCl, creatinine clearance; NIHSS, National Institute of Health Stroke Scale; DWI,…