Computer model has learned to predict the recurrence of tuberculosis

6 December 2018

One of the key health challenges set by WHO is to put an end to the tuberculosis epidemic by 2030. This requires the methods of early diagnosis and effective treatment of this disease. Scientists at TSU with specialists from Siberian State Medical University, Tomsk Oblast Tuberculosis Dispensary and the Research Institute of Tuberculosis (Novosibirsk), have developed a new approach to identifying drug-resistant forms of pulmonary tuberculosis. Researchers have created a computer model that has passed machine learning and can, with an accuracy of more than 95 percent, identify patients at risk.

- The increase in the incidence of tuberculosis, which is now registered throughout the world, is complicated not only by an increase in the number of primary patients but also by the fact that more patients have drug-resistant forms, - says Yury Kistenev, director of the StrAU Institute of Biomedicine. - In such cases, therapy does not give the desired effect, a person has a relapse of the disease, which leads to the infection of other people. The treatment of such patients is costly and time-consuming. To reduce the number of such cases, it is necessary at the beginning of treatment to assess the risk of recurrence in the treatment of a patient with standard drugs.

Doctors have provided a database with clinical indications of 850 tuberculosis patients and information about relapses. In the course of developing a computer model, various methods for classifying information were tested. Using the methods of machine learning, out of 120 initial parameters, scientists identified about 30, based on them a model was built that is capable of predicting a relapse with a very high probability, that is, identifying a risk group with drug resistance.

Verification of the created predictive model was carried out on data that did not participate in the learning process. It showed high accuracy in detecting the risk of recurrence, reaching almost 100 % on individual variants of the model.

The application of this program in practice will help the phthisiologist to determine the risk of recurrence and to select for the patients a set of drugs that will be effective in their form of the disease.