Neural Network Interpretation of the Intensity of Damage Processes to Biological Membranes of Human Cells, Depending on the Degree of Polymetallic Contamination of the Territory

Background: Lipid peroxidation is a primary driver of biological membrane damage and mediates the relationship between environmental exposure and adverse health outcomes. Malondialdehyde (MDA) is a widely recognized biomarker for quantifying oxidative stress intensity. Despite numerous studies on oxidative stress and metal exposure, nonlinear relationships between physiological characteristics, serum metal profiles and MDA levels in pubertal children remain insufficiently studied. Methods: The study included 105 conditionally healthy children aged 12–14 years from urban and rural regions of Tatarstan, Russia. Serum MDA concentrations were determined spectrophotometrically using the thiobarbituric acid assay, while Zn, Cu, Fe, Sr and Pb concentrations were measured by atomic absorption spectrometry. A multilayer perceptron neural network was applied to model nonlinear relationships between MDA levels, environmental exposure indicators and morphophysiological characteristics. Because the original relational dataset contained partially replicated participant-derived relational structures, primary validation was performed using independently reconstructed datasets without repeated observations. Additional repeated cross-validation and SHAP-based feature importance analysis were performed. Results: Urban-residing children demonstrated significantly higher serum MDA levels than rural counterparts, independent of sex, with girls consistently showing higher values. Reduction of predictor dimensionality improved model generalization behaviour. Validation using independently reconstructed datasets without repeated observations demonstrated reproducible exploratory predictive behaviour of the reduced neural network model, with independently reconstructed validation datasets yielding mean R2 values of 0.901 ± 0.052 and 0.914 ± 0.046, respectively. SHAP analysis demonstrated that zinc, copper and iron consistently represented the dominant contributors to the nonlinear model, although substantial variability in the relative ranking of zinc and copper was observed between validation datasets. Conclusions: The proposed neural network model demonstrated the ability to capture reproducible nonlinear relationships between oxidative stress markers and environmental exposure parameters in a limited biomedical dataset. The model should primarily be interpreted as an exploratory explanatory tool rather than an individual clinical prediction instrument. Because of the limited dataset size, partially reconstructed relational structure and exploratory study design, the findings require cautious interpretation and further external validation.