Clinical deterioration detection for continuous vital signs monitoring using wearable sensors

Abstract

Surgical patients are at risk of experiencing clinical deterioration events, especially when transferred to general wards during the postoperative period of their hospital stay. Cur rently, such events are detected by combining Early Warning Scores (EWS) with manual and periodical vital signs measurements, performed by nurses every 4 to 6 hours. Hence, deterioration may remain unnoticed for hours, delaying patient treatment, which might lead to increased morbidity and mortality. Also, EWS are inadequate to predict events so physiologically complex. So that early warning of deterioration could be provided, it was investigated the potential of warning systems that combine machine learning-based prediction models with continuous vital signs monitoring, provided by wearable sensors. This dissertation presents the development of such a warning system, fully indepen dent of manual measurements and based on a logistic regression prediction model with 85% sensitivity, 79% precision and 98% specificity. Additionally, a new personalized ap proach to handle missing data periods in vital signs and a novel variation of a RR-interval preprocessing technique were developed. The results obtained revealed a relevant im provement in the detection of deterioration events and a significant reduction in false alarms, when comparing the warning system with a commonly employed EWS (42% sensitivity, 14% precision and 90% specificity). It was also found that the developed sys tem can assess patient’s condition much more frequently and with timely deterioration detection, without even requiring nurses to interrupt their workflow. These findings sup port the idea that these warning systems are reliable, more practical, more appropriate and produce smarter alarms than current methods, making early deterioration detection possible, thus contributing for better patients outcomes. Nonetheless, the performance achieved may yet reveal insufficient for application in real clinical contexts. Therefore, further work is necessary to improve prediction performance to a greater extent and to confirm these systems reliability.