Feasibility of a Machine Learning Classifier for Predicting Post-Induction Hypotension in Non-Cardiac Surgery

Abstract

Purpose: To develop a machine learning (ML) classifier for predicting post-induction hypotension (PIH) in non-cardiac surgeries.

Materials and methods: Preoperative data and early vital signs were obtained from 3669 cases in the VitalDB database, an open-source registry. PIH was defined as sustained mean arterial pressure (MAP) <65 mm Hg within 20 minutes since induction or from induction to incision. Six different ML algorithms were used to create binary classifiers to predict PIH. The primary outcome was the area under the receiver operating characteristic curve (AUROC) of ML classifiers.

Results: A total of 2321 (63.3%) cases exhibited PIH. Among ML classifiers, the random forest regressor and extremely gradient boosting regressor showed the highest AUROC, both recording a value of 0.772. Excluding these models, the light gradient boosting machine regressor showed the second highest AUROC [0.769; 95% confidence interval (CI), 0.767-0.771], followed by the gradient boosting regressor (0.768; 95% CI, 0.763-0.772), AdaBoost regressor (0.752; 95% CI, 0.743-0.761), and automatic relevance determination regression (0.685; 95% CI, 0.669-0.701). The top three important features were mean diastolic blood pressure (DBP), minimum MAP, and minimum DBP from anesthetic induction to tracheal intubation, and these features were lower in cases with PIH (all p<0.001).

Conclusion: ML classifiers exhibited moderate performance in predicting PIH, and have the potential for real-time prediction.

Keywords: Anesthesia; artificial intelligence; general; general surgery; hypotension; machine learning.

Fig. 1. Study flowchart. AdaBoost, adaptive boosting; ARD, automatic relevance determination; CV, cross-validation; GBM, gradient boosting machine; HR, heart rate; MAP, mean arterial pressure; SMOTE, synthetic minority oversampling technique; XGB, extremely gradient boosting.

Fig. 2. Distribution of (A) duration of PIH and (B) time to PIH with cumulative distributions curves. PIH, post-induction hypotension.

Fig. 3. Hemodynamic measurements from the VitalDB database for a 73-year-old male patient (case number 268) undergoing exploratory laparotomy. (A) MAP and (B) HR measurements from the start of anesthesia to incision. (C) MAP and (D) HR measurements from the initiation of anesthesia to the end of anesthesia. HR, heart rate; IH, intraoperative hypotension; MAP, mean arterial pressure; Op, operation; PIH, post-induction hypotension.

Fig. 4. Receiver operating characteristic curves (A and C) and precision-recall curves (B and D) of machine learning classifiers for predicting post-induction hypotension based on training set (A and B) and test set (C and D). AdaBoost, adaptive boosting; ARD, automatic relevance determination; AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; GBM, gradient boosting machine; XGB, extremely gradient boosting.

Fig. 5. Shapley additive explanations summary plots for (A) random forest regressor and (B) XGB regressor. XGB, extremely gradient boosting.