Modern medicine and medical science are increasingly resorting to the use of complex mathematical methods for analyzing large amounts of large-dimensional data and developing algorithms for making diagnostic and therapeutic decisions based on them.
One of these main analytical methods is mathematical statistics. Doctors collect data on morbidity, treatment effectiveness, drug side effects, and other factors, and then analyze this data using statistical methods. This makes it possible to evaluate the effectiveness of treatment, identify risk factors, plan studies and make predictions.
Speaking at the conference "Tatiana's Day – Elements of Artificial Intelligence in Modern Hematology," Candidate of Technical Sciences, Head of the Information and Analytics Division of the NMRC for Hematology of the Ministry of Health of the Russian Federation, Sergey Kulikov presented the report "Search for genetic predictors of response to the treatment of diseases of the blood system using machine learning algorithms."
Addressing the participants of the conference, Sergey M. Kulikov noted that the purpose of his report is to protect researchers from false expectations of a miracle. The speaker emphasized that machine learning and artificial intelligence are two different things. Machine learning is a set of modern statistical methods for in-depth analysis of solving complex high-dimensional problems when classical multifactorial analysis ceases to work. The main purpose of machine learning is to rank features according to their predictive significance for classification, regression, event analysis and the construction of decision rules with the possibility of their interpretation.
In recent years, all algorithms that claim to be artificial intelligence and machine learning have been required to be interpretable and explicable. The resulting algorithm built by machine learning systems should be able to explain its decisive rule to a person. This is especially important in medicine.
Kulikov showed the work of machine learning algorithms using the examples of chronic myeloid leukemia, multiple myeloma and acute promyelocytic leukemia, when faced with the fact that the number of characteristics is comparable or greater than the sample size.
This analysis used: The machine learning method, known as "random forests,” was used to classify and analyze events. Methods of classical statistical analysis were used for comparison and interpretation. Examples of using free text processing algorithms for preprocessing and encoding purposes were also provided.
One example of the use of machine learning methods was the analysis of the relationship between copy number changes with a focus on the genes involved and the survival of patients with B-cell acute lymphoblastic leukemia. The dataset included 36 patients with Ph–B-ALL receiving therapy at NMRC for Hematology according to the ALL-2016 protocol from 2019 to 2023.
For the analysis, 46 binary signs were used, containing information about the presence of translocation, an increase or decrease in copy number involving a certain gene, as well as the presence of hyperplasia. Using the random forest method, the division staff ranked the initial symptoms according to the degree of importance, i.e. the degree of connection with survival.
The list included known factors influencing the outcome: the presence of translocation t(4:11) - a specific chromosomal aberration, which is most often found in acute lymphoblastic leukemia (ALL) and is considered an unfavorable factor, a change in the number of copies involving the CDKN2A gene (this gene encodes several proteins that slow cell division and acts as a tumor suppressor) - a neutral factor, and translocation t(7;14) - a favorable factor.
The researchers concluded that the most unfavorable genetic factor, according to the results of the analysis, is the presence of translocation t(4:11) and an increased copy number of the BIRC3 and ATM genes.
The ATM gene carries information about a protein that recognizes DNA damage and participates in the process of correcting it. If there is a mutation, the cell does not cope well enough with the restoration of its genetic material, which means that the risk of developing tumors increases.
An article written by a large group of authors from our center with the results of this study has been published in the International Journal of Molecular Sciences ( https://doi.org/10.3390/ijms242417602, Risinskaya et al.)
Also, the report presented an analysis of the effectiveness of therapy of patients with CML with asciminib. It is the first and only allosteric inhibitor of BCR::ABL, which was registered in Russia in January 2023.
Scientists and doctors were interested to learn what genetic factors influence a response to therapy. The study included 29 patients with failure in two lines of therapy with other drugs. The mutations ASXL1 and RUNX1 had the greatest prognostic significance.
Asciminib showed clinically significant efficacy in patients with the T315I mutation (a mutation in this gene causes leukemia cells to be resistant to all known tyrosine kinase inhibitors) almost 48.9% by the 96th week of therapy.
The report also showed the results of a study on the prognosis of the outcome of neurological disorders in hematological patients. An important element in solving this problem was the primary linguistic analysis for the preliminary processing of textual information, turning it into a set of numerical features suitable for further statistical analysis. The data of the conclusions of MRI, EEG and evoked potentials in the form of a free text were analyzed at the time of the first consultation. The stages of preprocessing the text of the conclusions for digitization and formalization of expert opinions were presented in detail. The researchers obtained a ranking of “signs of abnormalities” in the initial examination, which are most associated with a poor prognosis.
In conclusion, it was noted that in scientific research, a situation often arises when there is a lot of information, but there are few objects of research. This conflict forces us to look for new algorithms. However, despite the power of the new analysis tools, this does not eliminate the need to do data preprocessing in order to reduce the dimension of the original information.
«Miracles don’t exist. In addition to the methods of machine learning, biological, mathematical and statistical models should be used. It is necessary to make an attempt to reduce the dimension and only then use these methods. Random forests, in our opinion, are the most stable and effective machine learning tool. It is mainly intended for search analysis. Subsequently, the random forest results should be checked using classical biostatistics methods. No advanced methods reduce the requirements for data quality and representativeness of the training sample. It is important to remember that garbage begets garbage. One must be very careful in practical application. You cannot use machine learning if the machine does not explain the solution. Repositories (libraries, archives, collections) of data are the foundation and key to the success of using machine learning and artificial intelligence», said Sergey M. Kulikov.
