Problem Statement Patient safety in the modern hospital is supported by multiple human and technological systems, and this type of high-risk/high-complexity domain requires multiple layers of protection against harm. This fault-tolerant approach is reflected in the Failure to Rescue (FTR) metric, which measures an institution's ability to respond to patients who have suffered an adverse event. But clinical resources are limited for identifying these cases on the general care unit, and threshold-based physiologic state monitors must be configured with wide limits to minimize false alarms. A deterioration detection system which improves sensitivity while maintaining high specificty would have the potential to provide an improved system of protection against harm and further reduce FTR rates. Methods Previous studies have shown a 6-8 hour period of instability precedes many forms of patient deterioration, especially cardiopulmonary arrests. To test the hypothesis that this period could be automatically identified using a physiologic monitoring system, an archiving system was constructed to preserve continuous inpatient pulse oximetry monitoring data and a dataset was collected comprising both patients who had required emergency intervention and control cases who had not. Satistical analysis was applied to determine if the two groups displayed different physiological characteristics. Several classification techniques were then tested for sensitivity and specificity in identifying deteriorations. Results A set of time domain and complexity features was found to differ significantly between the deterioration and control cases. Naive Bayes' and neural network classifiers achieved better-than-chance performance in identifying the deterioration cases up to two hours before the call for emergency assistance. Detection performance was especially strong for those cases identified as primarily respiratory in nature. Conclusions The physiologic archive developed in this work is the largest extant collection of high-frequency data on non-critical inpatients. The feature analysis reflects earlier findings that deteriorating patients display abnormal vital signs and increasing instability, and classifier performance demonstrates the potential for an added layer of FTR prevention.