Online Anomaly Detection in Microbiological Data Sets

To prevent health risks caused by waterborne bacteria, significant changes of the bacterial community have to be detected as soon as possible. The aim of this study was to research suitable methods and implement a prototype of a system that can immediately detect such anomalous data points in microbiological data sets. The method chosen for the detection of anomalous cell counts was prediction-based out-lier detection: auto generated models were used to predict the expected number of cells in the next sample and the real number was compared to the prediction. Significant changes in bacterial communities were identified using Cytometric Fingerprinting, a method that provides functionalities to compare multivariate distributions and quantify their similarity. The prototype was implemented in R and tested. These tests showed that both methods were capable to detect anomalies but have to be optimized and further evaluated.