There are eight scatter plots, each representing a different clinical feature's contribution to the model outputs of a final genetic diagnosis. Each dot in the scatter plots represents an individual patient. The x-axis indicates the feature value, and the y-axis indicates the SHAP value, where positive values increase and negative values decrease the predicted probability of a final genetic diagnosis. The dashed horizontal line denotes SHAP equals 0. Point colors represent the pre-test clinical classification as indicated on the color bar. Panel A shows the SHAP dependence plot for chest or abdominal pain. The presence of chest or abdominal pain is associated with lower SHAP values, indicating a reduced contribution to model outputs predicting a final genetic diagnosis across multiple clinical groups. Panel B shows the SHAP dependence plot for family history. Greater family history scores are associated with progressively increased SHAP values, indicating positive contributions to the model outputs across groups. Panel C shows the SHAP dependence plot for autoantibodies. Autoantibody positivity shows heterogeneous SHAP contributions without a consistent directional effect across clinical subgroups. Panel D shows the SHAP dependence plot for lymphocyte counts. Lower lymphocyte counts are associated with higher SHAP values, whereas higher lymphocyte counts contribute negatively to model outputs. Panel E shows the SHAP dependence plot for skin rash. The presence of skin rash contributes positively to model outputs, particularly among non-ID/AID and AID patients. Panel F shows the SHAP dependence plot for enterocolitis. Enterocolitis shows variable contribution across patients, with generally modest effects on model prediction. Panel G shows the SHAP dependence plot for serum IgG levels. Lower serum IgG levels tend to contribute positively to model outputs, whereas higher IgG levels show reduced or negative contributions. Panel H shows the SHAP dependence plot for female sex. Female sex shows relatively small and heterogeneous contributions to model outputs across groups.
SHAP dependence analysis in relation to a final genetic diagnosis. Panels show SHAP dependence plots for variables contributing to the model outputs of a final genetic diagnosis. Each dot represents an individual patient; the x-axis indicates the feature value (binary features are coded as 0/1, where applicable), and the y-axis indicates the SHAP value (positive values increase, and negative values decrease, the predicted probability of a final genetic diagnosis). The dashed horizontal line denotes SHAP = 0. Point colors represent the pre-test clinical classification (AID, ID, and non-ID/AID) as indicated on the color bar. (A) Chest/abdominal pain. The presence of chest/abdominal pain was associated with lower SHAP values, indicating reduced contribution to model outputs predicting a final genetic diagnosis across multiple clinical groups. (B) Family history. Greater family history scores are associated with progressively increased SHAP values, indicating positive contributions to the model outputs across groups. (C) Autoantibodies. Autoantibody positivity showed heterogeneous SHAP contributions without a consistent directional effect across clinical subgroups. (D) Lymphocyte counts. Lower lymphocyte counts were associated with higher SHAP values, whereas higher lymphocyte counts contributed negatively to model outputs. (E) Skin rash. The presence of skin rash contributed positively to model outputs, particularly among non-ID/AID and AID patients. (F) Enterocolitis. Enterocolitis showed variable contribution across patients, with generally modest effects on model prediction. (G) Serum IgG. Lower serum IgG levels tended to contribute positively to model outputs, whereas higher IgG levels showed reduced or negative contributions. (H) Female sex. Female sex showed relatively small and heterogeneous contributions to model outputs across groups.