Prediction of complications associated with general surgery using a Bayesian network.

BACKGROUND

Numerous attempts have been made to identify risk factors for surgery complications, but few studies have identified accurate methods of predicting complex outcomes involving multiple complications.

METHODS

We performed a prospective cohort study of general surgical inpatients who attended 4 regionally representative hospitals in China from January to June 2015 and January to June 2016. The risk factors were identified using logistic regression. A Bayesian network model, consisting of directed arcs and nodes, was used to analyze the relationships between risk factors and complications. Probability ratios for complications for a given node state relative to the baseline probability were calculated to quantify the potential effects of risk factors on complications or of complications on other complications.

RESULTS

We recruited 19,223 participants and identified 21 nodes, representing 9 risk factors and 12 complications, and 55 direct relationships between these. Respiratory failure was at the center of the network, directly affected by 5 risk factors, and directly affected 7 complications. Cardiopulmonary resuscitation and sepsis or septic shock also directly affected death. The area under the receiver operating characteristic curve for the ability of the network to predict complications was >0.7. Notably, the probability of other severe complications or death significantly increased when a severe complication occurred. Most importantly, there was a 141-fold higher risk of death when cardiopulmonary resuscitation was required.

CONCLUSION

We have created a Bayesian network that displays how risk factors affect complications and their interrelationships and permits the accurate prediction of complications and the creation of appropriate preventive guidelines.