A real-time novelty detector for artefact identification in physiological measurements

We report the design of a kernel-based on-line novelty detector (ADDaM - Automatic Dynamic Data Mapper) and its use in the detecting of artefacts in physiological data streams gathered during general anaesthesia. ADDaM is an on-line method, that produces a robust and principled statistically partitioned history of any ordered data stream. It then constructs a probability distribution function (PDF) of the values in the stream by placing suitable Gaussian kernels at the centres of each of the partitions. The novelty of the next point entering the stream is assessed by testing against the current PDF. The more novel the point the more likely it is to be an artefact. The partitions and the PDFs are then updated after the novelty of each new point is assessed. The performance of this method is compared with artefact detection using both conventional on and off-line methods including Kalman filters, ARIMA and moving median or mean methods. The study shows that the performance of our novelty detector is as least as good as the best alternative on or off-line methods. Typical, error rates for artefact identification of 9.2% were achieved by ADDaM, compared with 16.5% for the best Kalman filtering, 9.3% for the best ARIMA model tested, 5.3% for the best moving mean method and 9.6% for the best moving median method.