Predicting Oxygen Saturation Levels in Blood Using Autoregressive Models: A Threshold Metric for Evaluating Predictive Models

This paper presents preliminary results for using data driven models to describe the natural dynamics of the Pulse Oximetry Monitoring signals. Linear autoregressive discrete time models are used to predict future levels of oxygen saturation in patients' blood. While standard modeling methods are used in identifying dynamic system models for these physiological signals, a performance objective based on a threshold is proposed to evaluate the predictive models. We discuss why standard evaluation metrics that have been commonly used in analyzing engineering systems may not be relevant for physiological ones even though standard modeling techniques may still give acceptable results. Using the proposed evaluation metric, we show that the combination of predictive models with frequent pulse oximetry measurements can be used as a warning of critical oxygen desaturation events that might have adverse effects on the health of patients.