2019 № 3 Prospects for the using of machine learning methods for predicting cardiovascular disease
Morbidity and mortality from cardiovascular diseases (CVD) has remained the leading rate in recent decades worldwide. Primary prevention methods based on the management of cardiovascular risk factors are most effective in reducing the burden of CVD. In preventive medicine for risk management of CVD use the riskometers – scales that was obtained as a result of long prospective studies.But the practical application of the developing scales has showed the limitations in the forecast accuracy. Machine learning makes it possible to improve the accuracy of cardiovascular risk prediction due to nonlinear relationships of their deeper adjustment between risk factors and disease outcomes.
2236 patients’ data were used. We trained the model on the features used in the Framingham scale construction. We compared the resulting model and the Framingham scale for the accuracy of the cardiovascular event prediction. Thus, according to the ROC analysis for the Framingham scale, the indicators are as follows: precision Accuracy: 70,0%, the AUC: 0.59. At the same time for the model obtained using machine learning similar indicators were: Accuracy: 78,8%, AUC: 0.84. Thus, the use of machine learning algorithms including deeplearning algorithms can significantly improve the accuracy of cardiovascular risk prediction of trained models.
2020 № 4 Algorithm for forming a suspicion of a new coronavirus infection based on the analysis of symptoms for use in medical decision support systems
The course of the COVID‑19 pandemic imposes a significant burden on healthcare systems, including on primary care,
when it is necessary to correctly suspect and determine further management. The symptoms non-specificity and the manifestations versatility of the COVID‑19 impose difficulties in identifying suspicions. To improve the definition of COVID‑19 symptom checkers and medical decision support systems (MDSS) can potentially be useful. They can give recommendations for determining the disease management.
The scientific analysis shows the manifestations versatility and the occurrence frequency COVID‑19. We structured the manifestations by occurrence frequency, classified them as “large” and “small”. The rules for their interaction were determined to calculate the level of suspicion for COVID‑19. Recommendations on patient management tactics were developed for each level of suspicion. NLP models were trained to identify the symptoms of COVID‑19 in the unstructured texts of electronic health records. The accuracy of the models on the F-measure metric ranged from 84.6% to 96.0%. Thus, a COVID‑19 prediction method was developed, which can be used in symptom checkers and MDSS to help doctors determine COVID‑19 and support tactical actions.