On-line novelty detection for artefact identification in automatic anaesthesia record keeping

Abstract

We report the design of a kernel-based on-line novelty detector (ADDaM — Automatic Dynamic Data Mapper) and its use in the detection of artefacts in an automatic anaesthesia record keeper (AARK). ADDaM produces a partitioned history of any ordered data stream and constructs a probability distribution function (PDF) from that history using Gaussian kernels. Two forms of PDF are possible: a s tatic PDF where the prior probability of each kernel is determined by the number of observations it represents and a temporal PDF where more recent observations have a higher prior probability. Testing against the current PDF assesses the novelty of the next point entering the stream. The performance of this method for artefact detection in heart rate data was compared to Kalman, ARIMA and moving mean filters using receiver operator characteristic (ROC) curves. Performance was measured using the area under the curves (AUC), and the false positive rate (FPR) and positive predictive value (PPV) calculated at the optimal cost-point on the curves. The results obtained were: ADDaM (Static PDF) AUC 0.92, FPR 0.12, PPV 0.12 and ADDaM (Temporal PDF) AUC 0.97, FPR 0.12, PPV 0.15. Both ADDaM-based methods out performed all other on-line methods tested.