De-escalating Vital Sign Checks

The overall goals for this study are: 1) to develop a predictive model to identify patients who are stable enough to forego vital sign checks overnight, 2) incorporate this predictive model into the hospital electronic health record so physicians can view its output and use it to guide their decision-making around ordering reduced vital sign checks for select patients.

Patients in the hospital often report poor sleep. A lack of sleep not only affects a patient's recovery from illness and their overall feeling of wellness, but it is a leading factor in the development of delirium in the hospital. One method for improving sleep in the hospital is to reduce the number of patient care related interruptions that a patient experiences. Vital sign checks at night are one example. In hospitalized patients who are clinically stable, vital sign checks that interrupt sleep are often unnecessary. However, identifying which patients can forego these checks is not a simple task. Currently, the hospital's quality improvement team asks physicians to think about this issue every day and order reduced, or "sleep promotion", vital sign checks on patients they believe could safely tolerate it. The investigators goal is to use a predictive analytics tool to reduce the cognitive burden of this task for busy physicians. The investigators plan to develop a logistic regression model, trained on data from the electronic health record (EHR), to predict, for a given patient on a given night, whether they could safely tolerate the reduction of overnight vital sign checks. The model will use variables, such as the patient's age, the number of days they have been in the hospital, the vital signs from that day, the lab values from that day, and other clinical variables to make its prediction. The outcome is a binary variable, whether the patient will or will not have abnormal vital signs that night. The training data is retrospective therefore it contains the nighttime vitals that were observed, which the investigators will code as a binary variable and use as the outcome variable for the model to train against. The investigators will incorporate this algorithm into an EHR alert so physicians can observe its output during their work, and use this information, complemented by their own clinical judgment, to decide about ordering reduced vital sign checks for a given patient. The investigators will study the effect of this EHR alert on several outcomes: in-hospital delirium (measured by nurse assessment), sleep opportunity (a measurement, based on observational EHR data, of patient care related sleep interruptions), and patient satisfaction (measured by nationally-administered post-hospitalization HCAHPS surveys). Balancing measures, to ensure that reduced vital sign checks do not cause patient harm, will be rapid response calls and code blue calls. Physician teams will be randomized to either see the EHR alert (intervention arm) or not see the EHR alert.

The investigators' intervention, which is a notification to the physician that is seen in the EHR, is randomized at the patient level. The patients randomized to the control group do not have a notification shown to their physician while the intervention patients do.

A pop-up window in the EHR will notify a physician that their patient has been judged by a predictive algorithm to be safe for reduced overnight vital sign checks.

Nursing Delirium Screening Scale (Nu-DESC score) - assessed by the nurse, can range from zero to ten, a score > 2 has good accuracy for delirium

a *novel* measurement based on observational EHR data - for every night in the hospital, the investigators can extract from the EHR all event timestamps that could have interrupted the patient's sleep (measured between 11 pm and 6 am). These are blood pressure recordings, fingerstick glucose checks, blood draws for labs, and not-as-needed medication administrations. The maximum time period between such events is considered the patient's sleep opportunity for that night (measured in hours). A higher sleep-opportunity on a given night is better. The investigators can calculate an average sleep-opportunity for a hospital encounter and then an average sleep-opportunity for all encounters in a clinical trial arm.

results from Hospital Consumer Assessment of Healthcare Providers and Systems (HCAHPS) surveys administered to patients after discharge from the hospital (scale is a categorical response: never, sometimes, usually, or always)

when a patient has a code blue (respiratory or cardiac arrest) called on them in the hospital, the resuscitation team that responds then writes a note documenting the event; the investigators can count these notes as a proxy for counting code blue events themselves (lower number is better)

average number will be calculated at study completion (6 months from study start date - Sep 11, 2019)

when a patient has a rapid response (significant change in vital signs or alertness) called on them in the hospital, the team that responds writes a note documenting the event and the investigators can count these notes as a proxy for counting rapid response events themselves (lower number is better)

average number will be calculated at study completion (6 months from study start date - Sep 11, 2019)

Inclusion Criteria: All physician teams that operate under the UCSF Division of Hospital Medicine Exclusion Criteria: N/A

Participants are physician teams. The investigators may submit their alert-response data to an online resource.

Najafi N, Robinson A, Pletcher MJ, Patel S. Effectiveness of an Analytics-Based Intervention for Reducing Sleep Interruption in Hospitalized Patients: A Randomized Clinical Trial. JAMA Intern Med. 2022 Feb 1;182(2):172-177. doi: 10.1001/jamainternmed.2021.7387.