The experience of neural network prediction of the need for surgical treatment in patients with the diseases of hepatopancreatoduodenal zone

Aim. Using multilayer perceptron artificial neural network, to develop a mathematical model for predicting the need for surgical intervention in patients admitted for hepatopancreatoduodenal zone diseases and to assess the capabilities for its clinical application.

Methods. The study was performed using the data of 488 patients with peptic ulcer, cholecystitis or pancreatitis, analyzed using multilayer perceptron artificial neural network, trained to distinguish vectors of data on risk factors of patients who did or did not require surgical intervention during current hospitalization.

Results. Patients in the training sample who had required surgical intervention during hospitalization were different from patients who had undergone conservative treatment by such characteristics as gender, age, duration of the disease, state on admission, and the structure of risk factors. The acquired data made it possible to train the artificial neural network. The ROC analysis of the mathematical model demonstrated the area under the curve (AUC) equal to 0.880 for the training group (n=385) and 0.739 for the clinical approbation group (n=103).

Conclusion. The AUC indicators of the created model can be characterized as very good in terms of predicting the need for surgical treatment in the training group and good for the clinical approbation group: sensitivity and specificity of the model exceed 80% in the training group and are highest in patients with peptic ulcer disease; in the clinical approbation group these parameters were lower as expected, however, remained at the level of 60-70%.