Identifying patterns of high intraoperative blood pressure variability in noncardiac surgery using explainable machine learning: a retrospective cohort study

Figures

• 

  Figure 1.. The process of constructing the interpretable model in pattern recognition of high intraoperative blood pressure variability in noncardiac surgery; XGBoost, extreme gradient boosting;RF, random Forest; LGBoost, light Gradient Boosting; LR, logistic regression.

• 

  Figure 2.. Receiver operating characteristic (ROC) curves of four machine learning models for predicting high intraoperative blood pressure variability (HIBPV) during noncardiac surgery. The area under the curve (AUC) was 0.85 (95% CI: 0.84–0.86) for XGBoost, 0.84 (0.82–0.85) for random Forest (RF), 0.82 (0.80–0.84) for light Gradient Boosting Machine (LightGBM), and 0.71 (0.69–0.74) for logistic regression (LR). These results demonstrate the superior discriminative performance of tree-based ensemble methods in identifying HIBPV patterns; XGBoost, extreme gradient boosting: RF, random Forest; LGBoost, light Gradient Boosting; LR, logistic regression.

• 

  Figure 3.. The top 10 significant features for identifying HIBPV in noncardiac surgery. (A): overall patients: the SHAP values and feature rankings are for the entire patient cohort, providing a general overview of the significant features and their impacts. (B): pediatric patients (under 18 years old): here, the SHAP values and rankings are specific to pediatric patients. The data reflects the unique characteristics and relationships of these features within this age group.(C): adult patients (between 18 and 65 years old): the SHAP values and feature rankings for adult patients are shown. The differences in values and rankings compared to other populations may be due to the physiological and surgical differences in adults. (D): elderly patients (65 years old and above): this Sub-figure presents the SHAP values and feature rankings for the elderly patient population. The data helps to understand the importance of features in predicting HIBPV in this age group, considering their specific health conditions. The x-axis represents SHAP values, which convey the magnitude and direction of each feature’s contribution to the model’s prediction. Negative SHAP values (left of center) imply a…

• 

  Figure 4.. SHAP interactive explanation plot for all patients from the testing set of the XGBoost model. (A): interaction between the dose of sevoflurane and patient age on intraoperative blood pressure variability. (B): interaction between arterial blood calcium levels and the history of hypertension on intraoperative blood pressure variability. The SHAP value is displayed on the Y-axis, with a negative number indicating a beneficial impact on the forecast by lowering BPV, while a positive value indicates an adverse influence on the prediction by increasing blood pressure fluctuation. The X-axis coordinates indicate the variation in the value of the specified influencing factor, while the predictive factor on the right side of the chart represents the interacting component that interacts with the specified influencing factor. The color red symbolizes an augmentation in the value of the interaction component, whereas the color blue signifies a reduction in the value of the interaction factor.