Figure 3
From: Towards interpretable, medically grounded, EMR-based risk prediction models
Impact of features of pre- and postoperative leaks risk predictions Subplots show how a given feature affects the risk prediction of the model. X-axes depict the values of the feature, y-axes the associated SHAP values. Missing values are represented on the far left of the x-axis. A positive SHAP value indicates that the value of the feature increases the overall risk, a negative SHAP values lowers risk of postoperative leaks. The larger the magnitude of the SHAP value the more it affects the risk prediction. Red vertical lines indicate inflection points were the value of the feature switches from having a negative to positive (or vice versa) impact on the risk prediction. The arrangement of the subplots follows the order of importance of the features for the model predictions. (a) Preoperative leak risk models. Red vertical lines depict: Age = 50 years, BMI = 30 kg / m2, Albumin blood = 4.4 g/dL, Protein blood total: 7.8 g/dL (b) Postoperative SSI risk models. Red vertical lines depict: Surgery length = 160 min, Albumin blood = 4.4 g/dL, Protein blood total = 7.2 g/dL, BMI = 25 kg / m2, Age = 60 years.
