Interpretable multi-label classification model for predicting post-anesthesia care unit complications: a prospective cohort study

Abstract

Background: There are potential associations between post-anesthesia care unit (PACU) complications that significantly impact enhanced recovery after surgery. Timely identification of these signs is essential for implementing comprehensive, systematic management strategies and delivering personalized anesthetic care. However, relevant studies are currently limited. This study aimed to develop and validate an interpretable multi-label classification model to predict PACU complications concurrently.

Methods: This prospective cohort study enrolled adult patients who underwent general anesthesia and elective surgery and were transferred to the PACU after surgery. The patients were dynamically monitored and evaluated for the occurrence of six common PACU complications: respiratory adverse events, hypothermia, hemodynamic instability, nausea/vomiting, agitation/delirium, and pain. A multi-label classification model was developed on the basis of 16 key features, and a Markov network was embedded to quantify and analyze the association network among these complications. The SHapley Additive exPlanations (SHAP) method was applied to conduct interpretability analysis of the model.

Results: Of the 16,838 total patients, 6,830 (40.6%) experienced at least one complication. In the training cohort, 2,125 (57.0%) patients experienced two or more complications at the same time. The AUCs for the six complications in the three cohorts ranged from 0.735 to 0.914, 0.720 to 0.920, and 0.693 to 0.928, respectively. Respiratory adverse events performed best. Age, gender, BMI, duration of anesthesia, and postoperative analgesia emerged as the five most important features. The relative importance of respiratory adverse events to hemodynamic instability was the highest.

Conclusion: The integration of a multi-label classification model with interpretable methods has significant advantages in simultaneously predicting PACU complications, identifying the risk factors for specific complications, optimizing postoperative resource allocation, and improving patient outcomes.

Keywords: Model interpretability; Multi-label classification model; Post-anesthesia care unit; Risk assessment; SHapley additive explanations.

Fig. 1

Co-occurrence of PACU complications. The upset plot illustrates the degree of overlap among various complications. The left bar plot displays the total number of each individual complication. The black circles indicate every possible co-occurrence between complications, with their respective counts shown in the top bar plot

Fig. 2

(A) Structure of the Markov network. (B) Heat-map of pairwise correlation between PACU complications. The darker the color, the stronger the correlation; and the lighter the color, the weaker the correlation

Fig. 3

SHAP feature importance plot. The x-axis represents the mean absolute SHAP values for each feature, quantifying each feature’s overall contribution to the model output. Bars of different colors represent various complications; the longer the bar, the greater the contribution to the specific complication. Features are sorted in descending order based on their overall contribution to all complications

Fig. 4

SHAP dependency plot for PACU complications. Each point represents a sample from the data-set, showing the SHAP values under two different complications. A higher CR value indicates a stronger relative importance between the two complications