Culture-independent data can be utilized to identify heterotrophic plate count (HPC) exceedances in drinking water. Although HPC represents less than 1% of the bacterial community and exhibits time lags of several days, HPC data are widely used to assess the microbiological quality of drinking water and are incorporated into drinking water standards. The present study confirmed the nonlinear relationships between HPC, intact cell count (ICC), and adenosine triphosphate (ATP) in tap water samples (stagnant and flushed). By using a combination of ICC, ATP, and free chlorine data as inputs, we show that HPC exceedance can be classified using a 2-layer feed-forward artificial neural network (ANN). Despite the nonlinearity of HPC, the best binary classification model showed accuracies of 95%, sensitivity of 91%, and specificity of 96%. ICC and chlorine concentrations were the most important features for classifiers. The main limitations, such as sample size and class imbalance, were also discussed. The present model provides the ability to convert data from emerging measurement techniques into established and well-understood measures, overcoming culture dependence and offering near real-time data to help ensure the biostability and safety of drinking water.
Keywords: Flow cytometry; Free chlorine; Heterotrophic plate count; Machine learning; Tap water.
Copyright © 2023 Elsevier Ltd. All rights reserved.