Predicting Clinical Deteriorations using Wearable Sensors

Introduction

Acutely-ill hospitalised patients are at risk of clinical deteriorations such as cardiac arrest, admission to intensive care, or unexpected death. Currently, patients are manually assessed every 4-6 hours to determine the likelihood of subsequent deterioration. However, this is limited to intermittent assessments, delaying time-sensitive interventions. Wearable sensors, combined with an alerting system, could provide continuous automated assessments of the likelihood of deteriorations. To be suitable for hospital use, wearable sensors must be unobtrusive and provide reliable measurements of key vital signs including breathing rate (BR), a key predictor of deteriorations. The aims of this work were: (i) to develop a technique for monitoring BR unobtrusively using wearable sensors, and (ii) to assess whether wearable sensors provide reliable predictions of deteriorations when using this technique.

Monitoring breathing rate (BR) unobtrusively

Current methods for monitoring BR using wearable sensors are obtrusive. An alternative approach is to estimate BR from electrocardiogram or pulse oximeter signals, which are already acquired by wearable sensors to monitor heart rate and blood oxygen levels. Both signals are subtly modulated by breathing, providing opportunity to use them to monitor BR. I assessed the performance of previously proposed signal processing techniques for estimating BR from these signals in both healthy and hospitalised subjects. Although some techniques were precise enough for use with healthy subjects in the laboratory, they were imprecise when used with hospital patients. Therefore, I developed a novel technique, combining the strengths of time- and frequency-domain techniques. Its performance was assessed on data from 264 subjects. In hospital patients, the technique provided highly precise BRs 86% of the time, which exceeds the performance of manual observation, the current clinical standard.

Assessing the reliability of wearable sensors for predicting deteriorations

I implemented methods for rejecting unreliable sensor data, and for fusing continuous multiparametric data, to predict deteriorations. These were used alongside the novel technique for monitoring BR to predict deteriorations using wearable sensors. The system was assessed in a clinical trial of 184 hospital patients, conducted in collaboration with clinicians. The reliability of the system was assessed by comparing its predictions against documented deteriorations. Its predictive value was similar to that of the routine manual assessments (AUROCs of 0.78 vs 0.79). Crucially it provided continuous assessment, potentially providing predictions of deteriorations hours earlier than routine practice.

Conclusion

This work has demonstrated the potential for wearable sensors to reliably and unobtrusively predict deteriorations, when coupled with a novel technique for monitoring BR. This could improve patient outcomes, and reduce costs. Further work should investigate which patients would benefit most from this technology, and whether it could reduce clinical workload. In the future the technology could potentially be used with consumer wearables to improve patient safety in the community, where clinical expertise is less readily available.