Data-driven approach for modeling of control action impact on anemia dynamics based on energy-informational health state criteria

In this work, spatial and time parameters were used to estimate the state of a biosystem on the strength of all attributes from integral criteria based on the Kullback information metric that is considered as a metric of a preferential behavior of a bio-object. Proximity of observed and preferential (reference) states is analyzed in the attribute space, where the proximity is normalized in the Mahalanobis model by an intraset distance of the reference state. Proposed approach can be considered as decision support instrument for medical experts evaluating possible outcomes of several treatment procedures. By modeling dynamics of individual integral criteria, the authors demonstrated the possibility of predicting the impact of sorbifer durules and mecsidol on the health state of pregnant women diagnosed with anemia. Experimental data included 8 blood properties from 92 pregnant women at different gestation periods measured at 3 timesteps: before; while; after applying treatment defined by medical experts. Treatment was considered as control action on biosystem and implicitly embedded into basic machine learning models such as linear regression and multilayer perceptrone. “Brute” approach to cross-validation identified several unique cases that could not be learned by models, due to lack of representativeness in original dataset and not model’s relative simplicity. Revealed anomalies were proven to be correct by medical experts from Tomsk neonatal health center.