Let Sleeping Patients Lie, avoiding unnecessary overnight vitals monitoring using a clinically based deep-learning model

“While overnight VS [vital sign] measurements disrupt sleep, they are often indicated and necessary for high-risk and potentially unstable patients. Identifying these patients in a reliable and timely manner is an area of active investigation, with efforts focused on models that vary from single parameter tools to weighted early warning scores and advanced predictive models using machine learning techniques. By contrast, relatively little work has been done to identify the low-risk cohort unlikely to benefit from such care that may, in fact, be harmed by these frequent assessments. Identifying this subset of patients has the potential to enhance recovery, improve patient sleep and satisfaction, and allow the redistribution of scarce resources (i.e., nurses, physicians) to higher-risk patients.

[..] By combining a deep recurrent neural network (RNN) for advanced predictive modeling with the clinical data generated by a multi-hospital health system, the derived tool enables the identification of low-risk inpatients and may improve outcomes by reducing overnight awakenings and enhancing sleep and recovery.

[..] The final data set obtained from a multi-hospital health system between 2012 and 2019 consisted of 2.13 million patient-visits (24.29 million VS measurements) in the retrospective cohort and 186,375 patient-visits (1.91 million VS measurements) in the prospective test set. We trained a deep RNN [recurrent neural network] with long short-term memory (LSTM) cells to predict individual patient stability for any hospital night, using a sequence of prior VS records during the hospital stay of each patient. The algorithm ingests a parsimonious list of longitudinal features, including respiratory rate (RR), heart rate (HR), systolic blood pressure (BP), body temperature (Tmpr), patient age, and a calculated risk score (Modified Early Warning Score [MEWS]), and produces a nightly assessment of overnight stability.

[..] Using the proposed deep-learning predictive model, we achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.966 (95% confidence interval [CI] 0.956–0.967) on the retrospective testing set, and 0.971 (95% CI 0.965–0.974) on the prospective set. Following model training and ROC curve construction, we established three different confidence thresholds, out of which the least conservative [..] can avoid overnight VS for 50% of patient-nights, while only misclassifying as stable 2 out of 10,000 patient-nights. [..] we established the clinically applicable region for this particular problem at a maximum of two false-positive predictions per 10,000 patient-nights (1.29% false-positive rate [patient-nights misclassified as stable divided by the total unstable patient-nights]), with the primary model threshold [..] lying on this region’s edge.

[..] To determine the benefit of using the proposed RNN architecture, we also evaluated a simple logistic regression model, receiving the same input variables as the proposed RNN, using the latest instance of VS measured right before the predicted patient-night rather than a sequence of VS. The logistic regression model achieved an AUC of 0.960 (95% CI 0.959–0.961) on the retrospective testing set and 0.964 (95% CI 0.962–0.965) on the prospective set.

[..] The benefits of reducing overnight VS monitoring extend beyond patient sleep. Nurses spend between 20 and -35% of their time documenting VS, and roughly 3 min per patient collecting them, accounting for ~10% of their shift with an eight patient census (average of 1.5 VS per patient per night). As healthcare systems seek to maximize efficiency and reduce waste, lean staffing models often hamper compliance with monitoring protocols as clinician capacity is exceeded, leading to delayed or incomplete care, particularly during periods of high acuity or census.”

Full article, Toth V, Meytlis M, Barnaby DP et al. npj Digital Medicine, 2020.11.13