Machine learning-based model for predicting the outcome of abdominal paracentesis drainage

Abstract

Abstract Background Abdominal Paracentesis drainage (APD) is a useful treatment for acute pancreatitis (AP) patient with pancreatitis associated ascitic fluid, however, researches seldom mentioned whether every patient benefit from this treatment. Here, we described a machine learning model to predict the outcomes of APD on certain AP patients. Methods The EHR data of 464 AP patients admitted between 2014 to 2020 were used in our study in a de-identified way. A machine learning model using random forest algorithm was established and validated under the stratified 10 fold cross validation strategy. The patients were labelled as “apd_cure” and “apd_serious” group according to their outcome, and the accuracy, sensitivity, specificity, positive prediction value, negative prediction value and ROC curve as well as its area under curve were used to value the efficacy of the model. A logistic regression model was established in the same strategy to compared their predictability. Results The random forest model has an excellent overall properties in predicting the outcomes of APD treatment for the AUC was 0.703 ± 0.118 [95%CI 0.64–0.77]. The accuracy, specificity and NPV (Negative Predictive Value) of the model was 0.786 ± 0.038, 0.940 ± 0.037 and 0.817 ± 0.037, respectively, indicates the model was more able to correctly classify patients who improved after APD treatment. The sensitivity and PPV(Positive Predictive Value) of the model was 0.208 ± 0.144 and 0.486 ± 0.232, which means that the model has insufficient ability to identify patients who may be more likely to have a worsening condition after APD treatment. Finally, the random forest model was statistically better than logistic regression model in accuracy and specificity. Conclusion The random forest model described in this study is a validated model in predicting the outcome of APD treatment on acute pancreatitis patients. It has higher overall performance than the logistic regression model. We hope it may help doctors choose treatment options appropriately and may enhance treatment efficacy in this group of patients.