This article substantiates the need to use data from an integrated electronic medical record of a patient to assess the risk of cancer. An exploratory analysis of the data of the integrated electronic medical record of patients in the Bryansk region who received a diagnosis of "malignant neoplasm" is being carried out. The influence of the patient's age on the risk of oncological diseases is evaluated by the example of the nosologies C50, C61. Provides an overview of the capabilities of the Auto ML Libraries and their limitations. The article describes the result of constructing models for assessing the risk of oncological diseases based on the ML.NET and Auto-WEKA libraries. It is concluded that it is impossible to constructing models for assessing the risk of oncological diseases based on the data of an integrated electronic medical record using Auto ML libraries without preliminary preparation and preprocessing of data. And since it is required to constructing separate models for each nosology and regular retraining of these models, it is advisable to develop an add-on over the Auto ML libraries that will extract and convert the data of the integrated electronic medical record into a form suitable for analysis. In addition, to improve the quality of the model, it is advisable to use patient history data, data obtained after vectorization of laboratory tests, aggregated data on visits to specialized specialists and related diagnoses, data from online patient questionnaires filled out during the course of medical examination, as well as data on environmental pollution.
Currently, in various fields of activity, data mining models (DM) are more and more actively used.
There is growing interest in the use of data mining models in medicine and in healthcare management
problems [
2021 Copyright for this paper by its authors.
To date, the regional information systems "Electronic medicine" of the Bryansk region have already accumulated a volume of biomedical data sufficient to start the process of constructing DM models, which, in turn, with an appropriate level of accuracy, can be used in the development of management decisions in the field of management health care at the regional level and the level of individual medical organizations. However, so far the DM models in the Russian Federation are not used in the process of health care management at the regional level, which is explained by the following reasons: The complexity of constructing DM models. The high cost of data scientists and the small number of such specialists in the regions. Failure to provide the level of accuracy required for quality control of all models. The need to construct an extremely large number of DM models for comprehensive automation of healthcare management, which is due to the number of different nosologies and the impossibility of constructing a single model for all nosologies. The International Classification of Diseases and Related Health Problems of the Tenth Revision (ICD-10) contains a description of about 14,400 different nosologies. More than a thousand of them are related to malignant and benign neoplasms.
According to the latest estimates of the International Agency for Research on Cancer (IARC) at the
World Health Organization (WHO) GLOBOCAN 2020 [
According to forecasts [
Some scientists associate an increase in the incidence of cancer, first of all, with an improvement in
diagnostics, as, for example, it happened in the United States in the 1990s. in the case of the introduction
of the prostate-specific antigen (PSA) test in prostate cancer [
One of the main indicators that determine the prognosis of an oncological disease is the degree of
prevalence of the tumor process at the time of detection. So, in Russia in 2019, 57.5% of malignant
neoplasms were diagnosed in stages I and II of the disease (32.3% in stages I and 25.2% in stages II)
and 42.5% in stages III and IV (17.6 % - in stage III and 24.9% - in stage IV) [
In connection with the above, the effectiveness of the functioning of the health care system in terms of detecting malignant neoplasms at an early stage (I-II) is extremely important in order to form risk groups in the process of population screening and preserve the health and quality of life of the population. An important factor hindering the timely detection of oncological diseases is the lack of clear criteria that allow doctors to identify a risk group during the period of clinical examination and conduct a set of additional examinations for patients belonging to this group. It seems promising to use a complex of data mining models for assessing the degree of risk in each patient during the period of clinical examination and subsequent informing the doctor about the degree of risk of patients undergoing medical examination, which will, on the one hand, increase the detectability, and, on the other hand, will not overload the system of functional and laboratory diagnostics.
Among the circumstances of the risk of malignant neoplasms, there are many exogenous and
endogenous factors, which are practically impossible for a doctor to take into account. According to the
literature [
The use of technologies for the analysis of biomedical data opens up new opportunities in the
development of medical information systems and improving the quality of healthcare management [
The data volume is about 100 Gigabytes. The database contains 304 tables with a total of 2062 fields, which, in the absence of documentation, makes reverse engineering and data extraction for analysis in an integral form a very difficult and non-trivial task. 53,182,278 records of doctors' appointments and hospitalizations of patients were extracted from the database, from which 1,773,148 records were selected for exploratory analysis containing diagnoses with codes of the international classification of diseases related to oncological diseases (ICD C00-C97).
To perform exploratory data analysis, a set of dashboards was developed on the Microsoft Power BI
platform (Fig. 1-11) using various visualization metaphors [
The X axis represents the quantized (10 years) age of the patient (Fig. 4) at which the diagnosis was made, the color indicates the stage at which the oncology was detected. It was found that for ICD C56 Malignant neoplasm of the ovary, almost half of the cases are detected only at the third stage, which is quite late, since at stages 3 and 4 the prognosis during treatment is much worse than at stages 1 or 2.
In the process of exploratory data analysis, an unusual dependence was found between the number of diagnosed oncologies on the age of the patient and the year of diagnosis. Fig. 5 shows the dependence of the number of diagnosed oncology cases on the patient's age for all years for ICD "C50 Malignant neoplasm of breast tissue" (25,195 patients) and C61 Malignant neoplasm of the prostate gland (indicate the number of patients 6,884). As mentioned above, these are some of the most common diagnoses. When we are analyzing Fig. 5, it can be seen that the age distribution is quite close to normal for both nosologies with a small additional peak on the right side.
The use of automated machine learning (AutoML) methods for the analysis (both exploratory and conventional) of biomedical data is gaining increasing popularity [17, 18], which allow in an automated mode to construct data mining models suitable for both exploratory analysis and decision making.
In article [17], the authors showed that AutoML methods can be used to analyze biomedical data and improve the productivity of specialists in solving a number of machine learning problems. It is also noted that the main limitation of AutoML at the moment is ineffective work on large datasets (dataset).
Fig. 13 shows a typical workflow for using AutoML to construct drill-down mining models. The scheme is based on the Cross-Industry Standard Process for Data Mining (CDISP-DM) methodology [19]. We reviewed the libraries of the most popular libraries that implement AutoML methods. We studied the capabilities of cloud services Google AutoML, Azure AutoML, Amazon Web Services AutoML and libraries Auto Keras, Auto-PyTorch, Auto-sklearn, Scikit-learn, Auto-WEKA, H2O AutoML, TPOT, MLBox and ML.NET. All of the above AutoML tools allow you to construct an DM model on a ready dataset in both automatic and automated modes. Many can perform selection of the optimal method from a predetermined set, selection of the optimal parameters of the selected method, limited data transformation (normalization, quantization, etc.) and optimization of the input data set (remove insignificant fields). However, none of the above libraries contains methods and tools that allow in an automated mode to select data that are optimal for constructing an DM model and perform their full preprocessing and transformation. The preparation of samples for training requires highly qualified specialists and a lot of time when processing large databases and data warehouses, which include sources of biomedical data. Selection of data mining
methods [The quality of the model
Is not sufficient]
Model validation Model deployment [The quality of the model is sufficient]
Hyperparameret optimisation
Shaping a dataset for exploratory data analysis Data perparation and
cleaning Shaping of the dataset for analysis
The use of the ML.NET [20] and Auto-WEKA [21] libraries for constructing a model for assessing the degree of risk of oncological diseases based on integrated electronic medical records of patients in the Bryansk region did not allow constructing a model of any value. The preprocessing methods available in these libraries are intended for normalized data, which did not allow using all the information available in the database to the fullest. Most of the AutoML libraries are focused on supporting neural network models, a significant drawback of which is the inability to explain and substantiate the results obtained. To assess the risk of oncological diseases, it is advisable to use scoring models based on logistic regression, which allow an expert to assess the contribution of each of the factors to the results obtained and more reasonably prescribe additional laboratory tests to the patient. Since AutoML methods applied directly to the available dataset did not lead to significant results and are unsuitable for replicating DM models, we have developed a technique for replicating DM models based on AutoML methods. Fig. 14 shows the process of replicating the DM models for a number of nosologies based on the reference model. Before starting the process, Data Scientist constructs a reference model for any nosology according to the scheme shown Fig. 11. After that, he needs to set the parameters for replicating models: a list of ICDs for which it is necessary to construct models, acceptable data mining methods for constructing models, a set data, a list of data preprocessing methods, model optimality criteria and accuracy requirements. The Meta ML library generates parameters for constructing a basic model for the i-th nosology based on the parameters of the reference model.
After that, using the Auto ML library, the process of constructing a basic DM model is carried out. Next, Meta ML generates K sets of parameters for constructing a set of DM models and prepares data sets for constructing these models. Model constructing is done using Auto ML. All built models are saved to the repository. When the models are built for all combinations of parameters and data sets, the optimal model for the i-th nosology is selected from the repository and its parameters are compared with the parameters of the base model. The best model is published. The generation of models is repeated for N nosologies. As a result, based on the reference model, models will be built for all the indicated nosologies. 5. Conclusions 1. To effectively flight socially significant diseases, a personalized preventive approach to risk assessment, early diagnosis of malignant neoplasms in the screening process and targeted correction of negative changes at the level of the individual, cohort, and population are required. 2. It seems promising to use data mining models to increase the likelihood of detecting malignant neoplasms in the early stages during medical examination. To use biomedical data in the task of assessing the risk of malignant neoplasms oncology requires significant preprocessing and data preparation, which cannot be performed using the existing Auto ML libraries. 3. The use of various visual metaphors in the process of exploratory analysis of biomedical data allows you to better understand the specifics of the data and discover various patterns, dependencies and phenomena that must be taken into account when constructing data mining models. 4. As shown by exploratory analysis, risk assessment with an acceptable level of accuracy requires the construction of a separate predictive model for each nosology, and regular (at least once a year, or better more often) adjustment (retraining) of these models to correctly take into account the patient's age. 5. Although the overall distribution of patients by age is close to normal distribution, there is a shift in the peak incidence over time, which will have a significant impact on the estimated level of risk. In all likelihood, this process obeys certain patterns, and it is extremely important to find them. This will allow predicting the peak incidence for a specific year or period and identifying the degree of risk of malignant neoplasms for different age groups. 6. Since the total number of nosologies is large enough, a single model will have very mediocre quality indicators, and it is very costly to construct (and keep up to date) a large number of models, it is advisable to use the proposed method of replicating models and develop a special library for sampling and preprocessing data from integrated electronic medical records and subsequent launch of Auto ML methods and tools.
The process of building multiple models from a reference Setting the parameters for constructing models: a list
of ICD, admissible methods data mining, data set, a list of data processing methods, criteria for optimality of a model, requirements for
accuracy View the report
For i=[1; N]
Generate the parameters for building a basic model for the i-th ICD based on the reference model Saving the base model
to repository For j=[1; K]
Preparing a dataset for building the j-th model Genrate the parameters of building the model j-th
Saving the j-model to the repository Extract the optimal model
for the i-th ICD [Optimal model better] [Basic model better] Publishing the optimal model
Publishing the base model 7. When constructing models for assessing the risk of malignant neoplasms in addition to data from the integrated electronic medical record, it is advisable to use patient history data, data obtained after vectorization of laboratory tests, aggregated data on visits by specialized specialists and related diagnoses, data from online patient questionnaires filled in the process of undergoing medical examination, as well as data on environmental pollution (chemical, physical, biological contamination) and the work of the population in dangerous manufacture.
The reported study was funded by RFBR, project number 19-07-00844.