Portable screen-printed carbon electrode biosensors provide a rapid and low-cost method to detect microcystin-lysine-arginine, a potent toxin produced by cyanobacteria during harmful algal blooms in freshwater. The World Health Organization has set a guideline value of 1 microgram per liter for this toxin in drinking water, as even low concentrations can damage the liver and have been linked to an increased risk of liver and colon cancer.
These biosensors work by measuring changes in an electrochemical signal that corresponds to the concentration of the toxin. However, their accuracy is often affected by varying water quality parameters such as pH, turbidity, and electrical conductivity, which usually require recalibration for each new sample.
Researchers from Hanbat National University in South Korea and the University of Central Florida in the United States announced on July 11 that they have developed a machine learning framework capable of accounting for differences in water quality. This allows for accurate measurements without repeated calibrations specific to each sample. The study was led by Professor Jungsu Park from Hanbat National University and Professor Woo Hyoung Lee from the University of Central Florida. Their findings were published online on March 26 and appeared in Volume 298 of Water Research on June 15.
"This work provides a robust data-driven framework for characterizing biosensor-water matrix interactions and offers a practical approach to improving the speed and accuracy of on-site MC-LR detection in complex environmental waters," says Prof. Park.
To train their model, researchers collected 201 measurements from 27 field sites across Florida representing diverse water conditions including freshwater, estuarine, and transitional environments. For each sample, they measured several variables: pH, turbidity, electrical conductivity, total dissolved solids, ultraviolet absorbance at 254 nanometers (UV254), as well as the biosensor's electrochemical impedance (Z'). These inputs were used to predict actual MC-LR concentrations using various machine learning models.
Extreme Gradient Boosting (XGBoost) performed best among tested models with a Nash-Sutcliffe efficiency of 0.89 and root mean square error of 13.21—demonstrating that one unified model could accurately predict toxin levels across different samples without needing separate calibration models.
The team also applied Shapley Additive Explanations (SHAP) to determine which input variables most influenced predictions; they found that biosensor electrical impedance was most important, followed by electrical conductivity, pH, UV absorbance at 254 nm, and turbidity—showing improved prediction when incorporating these parameters into analysis.
"This framework eliminates the need for repeated sample-specific calibration, reducing time, labor, and sensor consumption. Compared to conventional workflows it can reduce sensor usage thereby lowering cost and environmental burden while improving analytical efficiency," says Prof. Park.
As harmful algal blooms become more frequent due to climate change, this data-driven approach may make monitoring toxins faster, more accurate—and easier—to deploy both in drinking water supplies as well as recreational settings.
