Introduction. The article describes an attempt to predict the mortality rate of patients under dispensary observation by narcologists using a trained neural network based on data obtained from electronic cards of dispensary observation of patients who died between 2019 and 2023. Purpose: predicting the mortality rate of patients suffering from narcological diseases using a trained neural network. Materials and methods. Data on deceased patients were used for training and further prognosis. To determine the optimal model for training, the following were used: LightGBM (Light Gradient Boosting Machine) - gradient boosting library; CatBoost - gradient boosting library; Random Forest - Decision Tree Ensemble; Ridge is a type of linear regression with L2 regularization; Bidirectional LSTM is a type of recurrent neural network. Further, the model selected based on the results of the training made a forecast of the mortality rate in the group of patients suffering from narcological diseases at the dispensary follow-up in 01.01.2024. When evaluating the training results of each of the models, the indicators were evaluated: RMSE (Root Mean Squared Error) - the square root of the mean square error; MAE (Mean Absolute Error) - average value of absolute differences between model forecasts and actual values; R2 (R-squared) is the fraction of variance of the dependent variable that is explained by the model. Results. When evaluating, the best model was Bidirectional LSTM (neural network), which showed the best results. As a result, a neural network based on anonymized electronic records of living patients made a mortality forecast for 2024 and 2025. According to the results of this forecast, mortality in the group of people suffering from narcological diseases and under dispensary observation will be: in 2024: 1.97 ‰; in 2025: 1.16 ‰. Conclusions. The neural network is able to predict the mortality rate of a fairly high accuracy (the average discrepancy between the absolute differences in the real and predicted date of death obtained on the test part of the data was 1.27 years) even with a small amount of available data in electronic form. An increase in the initial data both in terms of sample size and quality (more filled fields) can significantly improve the quality of the forecast.