Bacterial biofilms cause biofouling of surfaces across industries. Removal of bacterial biofilms from surfaces is required to reduce energy consumption by industrial processes and mitigate human health risks. As biofilms are frequently located on hard-to-reach surfaces, such as within pipelines or on medical devices, it is difficult to assess efficacy of biofilm removal from surfaces after implementing biofilm removal strategies. Here, we develop a machine learning approach to predict biofilm removal from interfaces based on the biocolloidal properties of bacteria released from biofilms as cells released from biofilms are easier to characterize than biofilms in real-world applications. A small, low-throughput experimental dataset (N = 20) of analyzed confocal laser scanning microscopy images of biofilms and biofilm-released cell clusters is used in conjunction with a larger experimental dataset (N > 1,000) of analyzed Keyence digital microscopy images of biofilms and biofilm-released cell images to develop models predicting biofilm removal rates. We train three supervised machine learning models (non-linear dimension reduction, principal component analysis, autoencoder) on either the high-throughput dataset alone or a combination of the high-throughput dataset and half of the low-throughput dataset. We then compare the performance accuracies of the trained algorithms on the remaining low-throughput data using cross-validation. This work advances the development of machine learning methods for small experimental datasets and develops predictive models of biofilm removal using cells released from biofilms as indicators of surface removal rate.
