Skipping overnight vitals for patients using AI

AI eliminates patient vitals monitoring

Date: 19th November 2020

Sleep disruptions due to unnecessary overnight vital sign monitoring (VS) are associated with a range of conditions such as delirium, cognitive impairment, weakened immunity, hypertension, increased stress, and even mortality.  However, determining which patients can forgo VS is subjective, and whilst those at high risk and potentially unstable should be monitored, those at low risk would benefit from uninterrupted sleep.  Now, for the first time researchers have developed an artificial intelligence-based (AI), deep-learning predictive clinical tool to identify which patients do not need to be woken up overnight — allowing them to rest, recover and discharge faster.

On average a patient is woken every four to five hours for VS, and it is a common complaint from patients which can lead to delayed recovery and discharge of patients.  Furthermore, VS takes up valuable nursing time both collecting the data and documenting it.

Now, a team led by Theodoros Zanos and Jamie Hirsch from the Feinstein Institutes for Medical Research, US, have developed a predictive deep learning model that predicts overnight stability for any given patient-night, allowing high confidence, safe avoidance of overnight monitoring.

The team started by collecting and analysing data from multiple Northwell Health hospitals between 2012 and 2019.  This consisted of ~2.3 million admissions and 26 million vital sign assessments.  The team split the data using ~2.13 million patient-visits which equated to 24.29 million VS measurements to develop and train a deep recurrent neural network (RNN), and the used the remaining data in a prospective test set to determine the accuracy of the algorithm.  The algorithm produced a calculated risk score or Modified Early Warning Score –MEWS- and produced a nightly assessment of overnight stability.

Following model training, the deep-learning predictive model was set to established three different confidence thresholds.  The least conservative was able to decrease overnight VS for 50% of patient-nights, and only misclassified 2 out of 10,000 patient-nights as stable.  When the misclassified cases were examined the majority corresponded to marginally unstable patient-nights – the VS of these patients stayed largely stable throughout a 3-day period, but then abruptly increased during the falsely classified nights.

Conclusions and future applications

The team have successfully developed a deep-learning predictive model that takes a sequence of prior VS for a given patient and predicts the likelihood of overnight stability.  The AI model then guides safe decision making allowing overnight VS monitoring to be decreased by half thereby minimising sleep disruption.

The model was highly predictable and offered extremely low risk of miscalculation.  However, it is configured such that clinical teams could adjust the confidence threshold to implement a more stringent patient assessment or to provide a more risk-averse solution where applicable.  Furthermore, it is likely that a simple visual inspection of sleeping patients, which is part of the standard care for nurses, would detect any of those patients at risk.

Whilst the benefits to the patients may be obvious, the AI model will also enhance and optimise staff time.  The team estimate that halving overnight VS would equate to a 20-25% workload reduction in a single overnight shift.  This new clinical tool would enable nurses to safely cut VS in a non-subjective manner, easing the burden on staff and releasing time that could be better spent on critically ill patients.

The team are now working on rolling out this clinical tool, called the “Let Sleeping Patients Lie,” in several hospitals across Northwell Health.

We are currently seeing wearable, health monitoring sensors gain huge traction, due to their non-invasive ability to record a variety of vital signs.  We recently reported the development of wearable sensors that can be printed directly on skin, one can imagine that these, or similar devices, could be coupled with “Let Sleeping Patients Lie” providing a new strategy for clinical monitoring.  The AI-model could potentially be used to predict a new category of patients that would benefit from overnight VS using the wearable sensors without the need for waking, allowing these more at risk patients the benefit of uninterrupted sleep.  It would however, discriminating them from those patients that require a nurse interaction, and those that don’t need any intervention at all, as is seen with the group predicted here.  This combo approach could better direct resources, with AI driving optimal clinical decisions.


For more information please see the press release from the Feinstein Institutes for Medical Research

Tóth, V., M. Meytlis, D. P. Barnaby, K. R. Bock, M. I. Oppenheim, Y. Al-Abed, T. McGinn, K. W. Davidson, L. B. Becker, J. S. Hirsch and T. P. Zanos (2020). “Let Sleeping Patients Lie, avoiding unnecessary overnight vitals monitoring using a clinically based deep-learning model.” npj Digital Medicine 3(1): 149.