Machine learning is used to solve many real-world problems in various fields and is a good substitute for traditional statistical methods. The object of research to create a model for forecast cognitive impairment of elderly people living in Russia based on machine learning. The prospective data from 3570 elderly people (from «Hospital for War Veterans», Vladivostok, Primorsky Region, Russia) from 2019 to 2023 were used for the study. Criteria for inclusion in the study are age from 60 years and older. Exclusion criteria were past mental illnesses and the presence of a serious illness (cancer or dementia). The results of the subjective analysis of physical and mental health tests (the nutrition scale, Lawton, Bartel index, and the Mini-kog test) were used as input data, and the MMSE scale was used as output. The following machine learning methods to assess cognitive impairment were used: logistic regression, support vector machine (SVM), the K-nearest neighbor, and random forest (RF). The optimal thresholds and arranged indicators were studied in the verification data. The average age of the persons examined was 80.2 years, 50.85% were women. Of the 45 parameters, five were ultimately selected for model development: age, nutritional status, Barthel index, daily physical activity, and baseline cognitive function. To improve the forecast, the data were subjected to logarithmic transformations to bring them to a single scale using the Standard Scaller function. Logistic regression did not perform well on multiclass classification problems and showed the lowest AUC = 0.50. The SVM method also showed low forecast accuracy—25% with an AUC of 0.68. Its value indicated poor model performance, given that an AUC value of 0.5 indicates no discriminatory power. When analyzing significant features using RF, none of the potential predictors showed low significance < 0.1, so there was no need to exclude any features, the AUC-ROC index was 0.73. Thus, the machine learning methods used to classify signs of dementia based on the data set showed an accuracy level of 25–97%. The RF method for predicting the cognitive status of older people shows the best accuracy and AUC, which indicates its preferable use in future work in about creating software. Conclusions. From the classification methods, we applied signs of cognitive impairment and essential results were achieved when working with random forest, its accuracy was 97%. Thus, the association of multiline with action, such as a convolutional deep neural network and a random forest algorithm in using survey data of patients, will significantly increase the opportunity to detect symptoms of dementia early.