Clinical deterioration in the inpatient setting happens rapidly and often results in adverse patient outcomes such as cardiac arrest and mortality. Patients who do not experience mortality may have a longer length of stay and incur increased cost when experiencing deterioration while admitted to the hospital. Predictive models embedded in the electronic health record (EHR) use discrete data to identify deterioration risk, often sooner than could be detected by human assessment alone, and can prompt clinician action, both actively through real-time alerts and passively by display of risk scores throughout the EHR. Predictive models support early detection of clinical deterioration, encourage timely intervention, and decrease negative patient outcomes.
A multidisciplinary taskforce, led by informatics nurses and including front-line clinicians, physicians, nursing leadership, advanced practice providers, nurse educators, and IT experts, was formed to evaluate system readiness and design a framework for model implementation using a combination of improvement science, adaptive leadership, and data analytics. A lookback validation utility was applied for identification of the threshold at which deterioration was most likely and patient populations for inclusion were selected based on this customized evaluation. An expected alert response workflow was finalized, merging predictive technology with human action. A predictive model dashboard was created to monitor performance outcomes.
A baseline median was calculated using the first four months of data following implementation, reflecting the period during the project when the active alert was turned on but comprehensive staff education had not been completed. A new baseline median was calculated using data from the 8 months following comprehensive, refocused education for comparison.
During the 8 months following intervention 2, an improvement in mortality for patients triggering an alert was observed (24.5%, n=242 to 18%, n= 480), confirmed by six consecutive data points below the median. Cardiac arrest mortality also decreased from 76.9% of all patients with a DI alert (n=87) to 72.9% (n=168). A similar shift in mortality was observed when the rapid response team proactively rounded on patients for an elevated DI score (16.4%, n=162 to 14.3%, n=337).
When comparing expected deaths (number of deaths expected per month at a baseline mortality rate of 24.5% using the first 4 months of data) with actual deaths per month by observed monthly mortality rate following intervention 2, a total of 123 lives were saved.
Taken together, findings suggest that predictive models, when coupled with an organized system response, can improve patient outcomes such as mortality and cardiac arrest.