<p>Groundwater arsenic contamination poses a significant health risk in coastal region of Bangladesh. However, existing studies have rarely applied advanced machine learning (ML) algorithms to predict arsenic concentrations using comprehensive water quality parameters (WQPs). Thus, this study evaluated multiple ML algorithms to predict arsenic concentrations based on measured WQPs including pH, water temperature (WT), electrical conductivity (EC), salinity (WS), dissolved oxygen (DO), total dissolved solids (TDS), water hardness (WH), total alkalinity (TA), chemical oxygen demand, chloride (CL), and biological oxygen demand. Descriptive statistics showed moderate variability and symmetrical distributions in WQPs, with arsenic level varying between 0.01 and 0.72 mg L⁻¹. Correlation analysis revealed positive associations between arsenic and all WQPs (WT, r = 0.29 to TDS, r = 0.63), except for a negative correlation with DO (r = − 0.41). Among the models, MLP exhibited better predictive performance, achieving R² increments of 24.75%, 13.06%, 16.92%, and 5.15%, and RMSE reductions of 34.88%, 26.32%, 30%, and 12.5%, compared against MLR, RFR, SVR, and XGB algorithms, in that order. According to the finding, model performance was ranked as MLP &gt; XGB &gt; RFR &gt; SVR &gt; MLR. Notably, input combination 6 (TDS, Cl, EC, WH, TA, WS) enabled the MLP model to explain over 78.70% of the total variance and contributed to 91.83% of the total prediction accuracy. This research may offer important insights for assessing arsenic contamination in groundwater across Bangladesh's southeastern coastal area.</p>