Identification of patient deterioration in vital-sign data using one-class support vector machines

Adverse hospital patient outcomes due to deterioration are often preceded by periods of physiological deterioration that is evident in the vital signs, such as heart rate, respiratory rate, etc. Clinical practice currently relies on periodic, manual observation of vital signs, which typically occurs every 2-to-4 hours in most hospital wards, and so patient deterioration may go unidentified. While continuous patient monitoring systems exist for those patients who are confined to a hospital bed, the false alarm rate of conventional systems is typically so high that the alarms generated by them are ignored. This paper explores the use of machine learning methods for automatically identifying patient deterioration, using data acquired from continuous patient monitors. We compare generative and discriminative techniques (a probabilistic method using a mixture model, and a support vector machine, respectively). It is well-known that parameter tuning affects the performance of such methods, and we propose a method to optimise parameter values using “partial AUC”. We demonstrate the performance of the proposed method using both synthetic data and patient vital-sign data collected from a recent observational clinical study.