A probabilistic model for the prediction of intra-abdominal infection after colorectal surgery

Abstract

Aim: Predicting intra-abdominal infections (IAI) after colorectal surgery by means of clinical signs is challenging. A naïve logistic regression modeling approach has some limitations, for which reason we study two potential alternatives: the use of Bayesian networks, and that of logistic regression model. Methods: Data from patients that had undergone colorectal procedures between 2010 and 2017 were used. The dataset was split into two subsets: (i) that for training the models and (ii) that for testing them. The predictive ability of the models proposed was tested (i) by comparing the ROC curves from days 1 and 3 with all the subjects in the test set and (ii) by studying the evolution of the abovementioned predictive ability from day 1 to day 5. Results: In day 3, the predictive ability of the logistic regression model achieved an AUC of 0.812, 95% CI=(0.746, 0.877), whereas that of the Bayesian network was 0.768, 95% CI=(0.695, 0.840), with a p-value for their comparison of 0.097. The ability of the Bayesian network model to predict IAI does present significant difference in predictive ability from days 3 to 5: AUC(Day 3)=0.761, 95% CI=(0.680, 0.841) and AUC(Day 5)=0.837, 95% CI=(0.769, 0.904), with a p-value for their comparison of 0.006. Conclusions: Whereas at postoperative day 3, a logistic regression model with imputed data should be used to predict IAI; at day 5, when the predictive ability is almost identical, the Bayesian network model should be used.