Portable screen-printed carbon electrode (SPCE) biosensors provide a fast and cost-effective solution for detecting microcystin-LR (MC-LR), one of the most toxic compounds released by cyanobacteria during harmful algal blooms in freshwater ecosystems. Even at very low concentrations, MC-LR poses serious health risks, including liver damage, and has been associated with a higher likelihood of liver and colorectal cancers. To protect public health, the World Health Organization recommends that drinking water contain no more than 1 microgram of MC-LR per liter.
These biosensors estimate toxin levels by monitoring changes in electrochemical signals generated when MC-LR is present. However, their performance is often influenced by the characteristics of the water sample itself. Variations in pH, turbidity, electrical conductivity, and other water quality parameters can alter sensor responses, making it necessary to recalibrate the device for each individual sample.
A research team from Hanbat National University in South Korea and the University of Central Florida in the United States has developed a machine learning-based approach that overcomes this limitation by compensating for differences in water quality. The study, led by Professor Jungsu Park of Hanbat National University and Professor Woo Hyoung Lee of the University of Central Florida, was first published online on 26 March 2026 before appearing in Volume 298 of Water Research on 15 June 2026.
According to Professor Park, "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."
To develop the predictive model, the researchers gathered 201 datasets from 27 sampling locations across Florida, covering freshwater, estuarine, and transitional water bodies with diverse environmental conditions. For every sample, they recorded key water quality indicators—including pH, turbidity, electrical conductivity, total dissolved solids, ultraviolet absorbance at 254 nanometers (UV254), and the biosensor's electrochemical impedance (Z'), which varies with MC-LR concentration. These measurements served as model inputs, while laboratory-determined MC-LR concentrations were used as the target outputs.
The team evaluated several machine learning algorithms, with the Extreme Gradient Boosting (XGBoost) model delivering the strongest performance. It achieved a Nash-Sutcliffe efficiency of 0.89 and a root mean square error of 13.21, demonstrating that a single generalized model could accurately estimate MC-LR concentrations across a wide range of water conditions without the need for sample-specific calibration.
To better understand how different variables contributed to the model's predictions, the researchers employed Shapley Additive Explanations (SHAP), an explainable artificial intelligence technique. Their analysis revealed that the biosensor's electrochemical impedance had the greatest influence on prediction accuracy, followed by electrical conductivity, pH, UV254 absorbance, and turbidity. These findings highlight the importance of incorporating water quality characteristics into the predictive framework to improve biosensor reliability.
Professor Park noted, "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 and thereby lowering cost and environmental burden while improving analytical efficiency."
With climate change contributing to an increase in the frequency and severity of harmful algal blooms, this machine learning-assisted biosensing approach has the potential to make monitoring of toxic cyanobacterial contaminants faster, more accurate, and more practical for routine testing of drinking water supplies and recreational water bodies.
These biosensors estimate toxin levels by monitoring changes in electrochemical signals generated when MC-LR is present. However, their performance is often influenced by the characteristics of the water sample itself. Variations in pH, turbidity, electrical conductivity, and other water quality parameters can alter sensor responses, making it necessary to recalibrate the device for each individual sample.
A research team from Hanbat National University in South Korea and the University of Central Florida in the United States has developed a machine learning-based approach that overcomes this limitation by compensating for differences in water quality. The study, led by Professor Jungsu Park of Hanbat National University and Professor Woo Hyoung Lee of the University of Central Florida, was first published online on 26 March 2026 before appearing in Volume 298 of Water Research on 15 June 2026.
According to Professor Park, "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."
To develop the predictive model, the researchers gathered 201 datasets from 27 sampling locations across Florida, covering freshwater, estuarine, and transitional water bodies with diverse environmental conditions. For every sample, they recorded key water quality indicators—including pH, turbidity, electrical conductivity, total dissolved solids, ultraviolet absorbance at 254 nanometers (UV254), and the biosensor's electrochemical impedance (Z'), which varies with MC-LR concentration. These measurements served as model inputs, while laboratory-determined MC-LR concentrations were used as the target outputs.
The team evaluated several machine learning algorithms, with the Extreme Gradient Boosting (XGBoost) model delivering the strongest performance. It achieved a Nash-Sutcliffe efficiency of 0.89 and a root mean square error of 13.21, demonstrating that a single generalized model could accurately estimate MC-LR concentrations across a wide range of water conditions without the need for sample-specific calibration.
To better understand how different variables contributed to the model's predictions, the researchers employed Shapley Additive Explanations (SHAP), an explainable artificial intelligence technique. Their analysis revealed that the biosensor's electrochemical impedance had the greatest influence on prediction accuracy, followed by electrical conductivity, pH, UV254 absorbance, and turbidity. These findings highlight the importance of incorporating water quality characteristics into the predictive framework to improve biosensor reliability.
Professor Park noted, "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 and thereby lowering cost and environmental burden while improving analytical efficiency."
With climate change contributing to an increase in the frequency and severity of harmful algal blooms, this machine learning-assisted biosensing approach has the potential to make monitoring of toxic cyanobacterial contaminants faster, more accurate, and more practical for routine testing of drinking water supplies and recreational water bodies.