Fig. 2: SHAP summary plot for the Platt-calibrated XGB model using medication use, age, clinical history features. | npj Digital Medicine

Fig. 2: SHAP summary plot for the Platt-calibrated XGB model using medication use, age, clinical history features.

From: Interpretable machine learning models for stroke risk prediction in patients with newly diagnosed atrial fibrillation

Fig. 2: SHAP summary plot for the Platt-calibrated XGB model using medication use, age, clinical history features.

The SHAP (SHapley Additive exPlanations) summary plot illustrates the relative contribution of each predictor to the 1-year stroke risk prediction in the XGB model. Each dot represents a patient; color indicates the feature value (red = high, blue = low), and horizontal position reflects the SHAP value (positive values increase stroke risk, negative values decrease it). Prior ischemic stroke was the most influential predictor, consistently increasing stroke risk, followed by age, atrial arrhythmia drugs (AAD), and hypertension (HTN). Some features (e.g., age) exhibited bidirectional effects, highlighting context-dependent interactions captured by the XGB model. This interpretability enables clinicians to understand individualized risk profiles, supporting precision decision-making. AGE age at diagnosis, AAD anti-arrhythmic drugs, HTN hypertension, APA antiplatelet agents, DM diabetes mellitus, MLD mild liver disease, PVD peripheral vascular disease, PUDB peptic ulcer disease (excluding bleeding), VHD valvular heart disease, XGB extreme gradient boosting.

Back to article page