OBJECTIVE

To analyze the results of developing a neural network dialogue system for the formation of relevant responses to written requests of the population to a medical institution.

MATERIAL AND METHODS

The study was conducted at the Volga District Medical Center of the FMBA of Russia. The authors used a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model. As part of the study, a phased approach to fine-tuning the BERT model was used: adaptation in everyday vocabulary (Yandex/geo-reviews), optimization using loss functions (TripletLoss, ArccosMarginLoss), and classification using the SVM method. The dataset (2.623 records) was divided into training and test samples in a proportion of 8/2.

RESULTS

Citizens’ requests were classified using a feedback database of 2.623 records and divided into 18 categories. Categories with fewer than 20 entries were removed due to a lack of representativeness. It was found that messages have a substantial semantic overlap between categories. A single email could simultaneously contain several requests belonging to different categories. When the entry contained several similar but multidirectional statements, it resulted in a model failure. To avoid this, it was decided to use sets of templates that the model could offer the operator to facilitate the preparation of a response to a patient’s request.

CONCLUSION

The study results present compelling evidence that it is currently not feasible to completely exclude human participation in the full-fledged dialogue between artificial intelligence and a patient. The technological imperfection of modern artificial intelligence platforms does not yet allow solving the problems of subtle differentiation of a semantically complex message and its emotional coloring, thus preventing their wide autonomous use. However, they can already significantly facilitate the work of medical personnel in remote interaction with the served population.