BACKGROUND
Peritonitis remains one of the urgent problems due to its high lethality in case of untimely or insufficient treatment, as well as the occurrence of postoperative complications.
OBJECTIVE
Determining the possibilities of predicting the occurrence of complications in the postoperative period in patients admitted with the diagnosis of peritonitis, using the methods of artificial intelligence on the example of the method of gradient bousting.
MATERIAL AND METHODS
A dataset of 1558 patients containing coded data of clinical diagnostic and instrumental investigations performed during hospitalization of patients diagnosed with peritonitis was used to build the model. The target parameter is the outcome of the disease — the presence or absence of a complication. From the mathematical point of view, the task under consideration belongs to the binary classification problem. The method of Gradient Boosting on Decision Trees (GBDT) was used to build a binary classification model. GBDT was used to analyze the importance of the parameters, i.e. how strongly the parameters affect the target parameter — whether a complication occurred or not. The CatBoost library, which has proved to be one of the most effective among similar algorithms of gradient boosting on decision trees, was chosen for model building using the GBDT method.
RESULTS
It was found that the importance of parameters determined by artificial intelligence algorithms and statistical methods differ, which can be explained by the ability of artificial intelligence algorithms to capture internal implicit relationships between data. In this case, it is sufficient to collect only the 3 most important parameters to obtain a prediction. The constructed model showed a good predictive ability in terms of non-complications.
CONCLUSION
On the basis of model building, the potential possibility of predicting the occurrence or non-occurrence of complications in patients after surgery on the basis of data collected at the time of patient admission to the emergency room is shown. In further work with data and training of predictive models, it seems reasonable to use artificial intelligence methods to select important parameters.
