Automatic artifact identification in anaesthesia patient record keeping: a comparison of techniques

The anaesthetic chart is an important medico-legal document, which needs to accurately record a wide range of different types of data for reference purposes. A number of computer systems have been developed to record the data directly from the monitoring equipment to produce the chart automatically. Unfortunately, systems to date record artifactual data as normal, limiting the usefulness of such systems. aper reports a comparison of possible techniques for automatically identifying artifacts. The study used moving mean, moving median and Kalman filters as well as ARIMA time series models. Results on unseen data showed that the Kalman filter (area under the ROC curve 0.86, false positive prediction rate 0.31, positive predictive value 0.05) was the best single method. Better results were obtained by combining a Kalman filter with a seven point moving mid-centred median filter (area under the ROC curve 0.87, false positive prediction rate 0.14, positive predictive value 0.09) or an ARIMA 0-1-2 model with a seven point moving mid-centred median filter (area under the ROC curve 0.87, false positive prediction rate 0.14, positive predictive value 0.10). Only one method that could be used on real-time data outperformed the single Kalman filter which was a Kalman filter combined with a seven point moving median filter predicting the next point in the data stream (area under the ROC curve 0.86, false positive prediction rate 0.23, positive predictive value 0.06).