Interpretable machine learning for early neurological deterioration prediction in atrial fibrillation-related stroke

Figures

• 

  Figure 1. Flowchart of included patients.

• 

  Figure 2. Model performance. (A), Solid lines and shades represent receiver operating characteristics curves and its 95% confidence intervals. An asterisk (*) indicates significant difference (P < 0.005) in comparison with logistic regression. (B), Solid lines and shades represent precision-recall curves and its 95% confidence intervals. Only the confidence intervals of the baseline model (logistic regression, “LogReg”) are represented with polka dot pattern in both plots. (C), Detailed performance analysis for the best model (LightGBM) in different discrimination thresholds. Solid lines and shades represent mean values and 95% confidence intervals in each variable. Abbreviations: AUC, area under the curve; CI, confidence interval; LogReg, Logistic regression; SVM, Support vector machine; XGBoost, Extreme gradient boosting; LightGBM, light gradient boosting machine; MLP, Multilayer perceptron.

• 

  Figure 3. Matrix plots of top 23 important features. Bar plot (A) and violin plot (B). In the bar plot, the SHAP value implies the degree of contribution of a specific feature. The higher the SHAP value, the larger the model contribution of a specific feature. In the violin plot, each dot represents one patient and accumulates vertically to depict the density. The color reflects the high and low values of each feature, with the red color indicating a higher value and the blue color indicating a lower value. The X-axis of the graph represents the SHAP value, and a positive SHAP value indicates that it contributes positively to predicting the model, and that the probability of END occurring is high, and vice versa. Abbreviations: NIHSS, National Institute of Health Stroke Scale; mRS, modified Rankin scale; ALP, alkaline phosphatase; SVS, susceptibility vessel sign; ICAS, intracranial atherosclerosis; aPTT, activated partial thromboplastin time; FDP, fibrin degradation product; LA, left atrium; DBP, diastolic blood pressure; AST, aspartate aminotransferase; Hct, hematocrit; LDL, low-density lipoprotein.

• 

  Figure 4. Partial SHAP dependence plot of the four representative features. Values are plotted with a scatter plot and a regression line represented with the orange line of mean and shade of SD. A red diamond represents a cut-off value of the variable. Histograms on the right and top of each plot are distributions of the SHAP and values of variables. Abbreviations: NIHSS, National Institute of Health Stroke Scale; LA, left atrium; LDL, low- density lipoprotein.