A national assessment of drinking water across Montenegro has identified electrical conductivity and altitude as the most important and stable predictors of radon occurrence, according to an investigation utilizing statistical analyses and machine learning models. Researchers analyzed data collected under Montenegro’s national monitoring program to evaluate the distribution of radon and examine the environmental and physicochemical variables associated with its variability in groundwater supplies.
The investigation examined relationships among multiple water quality parameters, including radon concentration, gross alpha activity, gross beta activity, and various environmental and physicochemical measurements. To identify potential patterns, the researchers applied both bivariate and multivariate statistical analyses across the nationwide monitoring dataset.
Using the Kruskal–Wallis test followed by the Dunn–Bonferroni procedure, the authors assessed differences across three geographic regions and various lithological settings. The analysis revealed that differences in radon, gross alpha, and gross beta activity among the three regions were not statistically significant. In addition, while electrical conductivity and gross beta activity exhibited a certain association with limestone lithology, no statistically significant differences in radon concentrations were observed among the different lithological groups.
Correlation analysis indicated that altitude- and temperature-related factors may influence radon variability in drinking water. The authors noted that these associations potentially reflect underlying differences in radon transport mechanisms and groundwater residence conditions across the monitored locations.
To classify radon concentrations into three distinct categories, the researchers implemented a random forest machine learning model complemented by a leave-one-out feature-ablation sensitivity analysis. Evaluated through repeated stratified cross-validation and leave-one-out cross-validation, the random forest classifier achieved a mean accuracy of 53.8% (± 13.6%) and a weighted F1-score of 0.48 (± 0.13), with electrical conductivity and altitude emerging as the most stable predictive features.
Model interpretability was examined using SHAP analysis, which revealed a specific association between water temperature and elevated radon concentrations within the defined High-radon category. The authors concluded that these findings provide key insights into groundwater radon occurrence and demonstrate the utility of machine learning as an exploratory tool for identifying environmental parameters relevant to groundwater quality assessment and radiological risk evaluation.
Subject of Research: Factors associated with radon occurrence in drinking water in Montenegro
Article Title: Factors associated with radon occurrence in drinking water in Montenegro: a statistical and machine learning approach
Article References: Radonjic, S., Milicevic, B., Cimbaljevic, Z., Stajic, J. M., Nikic, B., & Marković, V. M. (2026). Factors associated with radon occurrence in drinking water in Montenegro: a statistical and machine learning approach. Environmental Geochemistry and Health, 48(17), Article 699. https://doi.org/10.1007/s10653-026-03600-2
Image Credits: AI Generated
DOI: 10.1007/s10653-026-03600-2
Keywords: radon occurrence, drinking water, Montenegro, national monitoring program, random forest model, electrical conductivity, altitude, water temperature, gross alpha activity, gross beta activity, limestone lithology, SHAP analysis, feature-ablation sensitivity analysis, Kruskal–Wallis test, Dunn–Bonferroni procedure, groundwater quality assessment, radiological risk evaluation