Portable screen-printed carbon electrode (SPCE) biosensors provide a rapid and low-cost method to detect microcystin-lysine-arginine (MC-LR), a highly potent toxin produced by cyanobacteria during harmful algal blooms in freshwater. Because MC-LR can damage the liver even at low concentrations and is associated with an increased risk of liver and colon cancer, the World Health Organization has set a guideline value for MC-LR in drinking water at 1 microgram per liter.
SPCE sensors work by measuring changes in electrochemical signals that reflect the concentration of toxins. However, the accuracy of these sensors is highly influenced by the water being tested. Factors such as pH, turbidity, electrical conductivity, and other water quality parameters can interfere with sensor readings, often requiring recalibration after each water sample.
Researchers from Hanbat National University in South Korea and the University of Central Florida in the United States have developed a machine learning framework that accounts for differences in water quality, enabling accurate MC-LR measurements without repeated sample-specific calibrations. The research was led by Professor Jungsu Park of Hambat National University and Professor Woo Hyoung Lee of the University of Central Florida. This paper was made available online on March 26, 2026 and was published in Volume 298 of the journal. water research June 15, 2026.
“This study provides a robust data-driven framework to characterize biosensor-water matrix interactions and provides a practical approach to improve the speed and accuracy of in situ MC-LR detection in complex environmental waters.” Professor Park says.
To build and train the model, the team collected 201 measurements from 27 sites across Florida representing a wide range of water quality conditions, including freshwater, estuarine, and transition environments. For each water sample, we measured pH, turbidity, electrical conductivity, total dissolved solids, ultraviolet absorbance at 254 nanometers (UV254), and electrochemical impedance (Z’) of the biosensor that changed in response to MC-LR. These measurements served as input variables and the model was trained to predict the actual concentration of MC-LR.
Among the different machine learning models evaluated, Extreme Gradient Boosting (XGBoost) showed the best performance, achieving a Nash-Sutcliffe efficiency of 0.89 and a root mean square error of 13.21. This level of performance demonstrated that a single integrated model can accurately predict MC-LR concentrations across a variety of water samples without the need for separate calibration models for each condition.
To determine which input variables had the greatest impact on the model’s predictions, the researchers used an explainable artificial intelligence technique called Shapley Additive Explains (SHAP). They found that the electrical impedance of the biosensor was the strongest predictor of toxin levels, followed by electrical conductivity, pH, UV absorbance, and turbidity, indicating that incorporating water quality parameters improved the predictive accuracy of the biosensor.
“This framework eliminates the need for repeated sample-specific calibrations, reducing time, effort, and sensor consumption. Compared to traditional workflows, sensor usage is reduced.” and Reduce costs and environmental impact while improving analytical efficiency.,” Professor Park says.
As harmful algal blooms become more frequent as the climate changes, this data-driven approach could make monitoring for toxins faster, more accurate, and easier to implement in drinking and recreational water testing.
sauce:
Hanbat University Industry
Reference magazines:
