This study investigates the hydrochemical characteristics and controlling factors of water quality along the Axios River (Northern Greece) using a combination of statistical methods and artificial neural networks (ANN). It also aims to predict Electrical conductivity (EC) which is a dependent physico-chemical parameter. Electrical Conductivity is a key indicator of river water quality, reflecting both natural hydrogeochemical processes and anthropogenic pressures. Water samples were collected from three monitoring stations (Idomeni, Prochoma, and Malgara) during the period 2018–2023 and were analyzed to assess spatial and temporal variability and to identify dominant hydrochemical processes. Hydrochemical analysis revealed that river water is predominantly characterized by Ca–Mg–HCO₃ facies, indicating strong geogenic control associated with carbonate weathering, while spatial variations along the river course were generally limited. Pearson correlation analysis and factor analysis were applied to explore linear relationships among major ions and EC, and Multiple Linear Regression (MLR) models were subsequently developed for EC prediction. Although MLR demonstrated satisfactory performance, its predictive capability varied among monitoring stations, reflecting differences in hydrochemical complexity. To investigate potential nonlinear relationships, ANN models based on a Multilayer Perceptron architecture were implemented and evaluated using independent training and testing datasets. The relative importance of input variables was assessed through ANN-based Independent Variable Importance analysis, which was further employed as a feature selection tool. The refined ANN models showed improved predictive performance and enhanced interpretability, particularly at stations influenced by mixed hydrochemical processes. The comparative evaluation of linear and nonlinear modelling approaches indicates that EC in the Axios River is largely governed by linear geochemical relationships, while ANN-based methods can capture subtle nonlinear effects and provide additional insight when combined with classical statistical analysis. The results highlight the complementary role of machine learning techniques in river water quality assessment and support their use as exploratory and interpretative tools rather than as standalone predictive solutions.