Use of Machine Learning to Develop and Evaluate Models Using Preoperative and Intraoperative Data to Identify Risks of Postoperative Complications

Figures

• 

  Figure 1.. Flowchart of Complication Analysis and Cohort Split AKI indicates acute kidney injury, DVT, deep vein thrombosis; PE, pulmonary embolism.

• 

  Figure 2.. Results of Machine Learning Models A, Areas under the receiver operating characteristic curve (AUROCs) of best-performing learning models. B, AUROCs when using only preoperative data, intraoperative data, and combined data. C, AUROCs with added features in ascending order of missing rate. D, AUROCs with varied number of features. AKI indicates acute kidney injury; DVT, deep vein thrombosis; PE, pulmonary embolism. The error bars indicate 95% CIs.

• 

  Figure 3.. Complication-Specific Model Interpretation A, Evolvement of risks (from top to bottom) contributed by each variable (magnitude of contribution decreasing from left to right) compared with a group of patients who did not have pneumonia. B, Characterization of significant intraoperative time series (in this case, it is the noninvasive mean blood pressure [MBP]) by its statistical features. Each statistical feature is normalized to zero mean and unit variance; therefore, the magnitude reflects its deviation from the historical mean of patients. BMI indicates body mass index (calculated as weight in kilograms divided by height in meters squared); ETCO2, end-tidal carbon dioxide; SHAP, Shapley Additive Explanations.