Auto-Calibration for Neonatal Condition Monitoring
Masters thesis, University of Edinburgh.
This project describes our approach for automating the calibration stage needed by the Factorial Switching Linear Dynamical System (FSLDS) for neonatal condition monitoring.
Using observations collected by the monitoring equipment, the FSLDS proposed by Prof Christopher Williams and Dr John Quinn  is highly successful in inferring the probability distributions of both physiological and artifactual hidden factors affecting the measurements. The detection of these factors presents vital importance to clinicians. However, the FSLDS must be calibrated using manually selected data segments corresponding to normal dynamics. Our goal
is to automatically find such intervals.
The proposed solution uses a binary classifier able to discriminate normal sections from a set of fixed length measurement intervals. A “channel-based” classification is justified, together with a new labeling for the data, feature extraction solutions and a performance criterion especially
developed for the task. Experimental results have demonstrated that our classifiers are able to correctly extract normality sections from the observations on hand. More importantly, a number of comparative tests between the auto- and manually-calibrated FSLDSs provide evidence that our approach is indeed successful in uncovering the hidden factors influencing the observational data.