Catching patient decline early enhances recovery and reduces mortality. But identifying warning signs remains difficult in hospitals and across care settings.
New research explores how technology, including artificial intelligence and wireless wearable sensors, can help to improve early identification of patient deterioration.
But another new study finds that it’s not just a matter of technological performance – suboptimal staffing workflows can delay vital data.
Data delayed by shift changes
There is a bottleneck between patient monitoring and care response that points to operational and staffing workflows rather than technology, according to a report this past month in npj Health Systems.
Researchers from Oregon Health & Science University found median delays spanned 12 to 25 minutes in the availability of patient vital signs between their recording and nurse validation across 11.4 million measurements – with a third of abnormal readings delayed by more than an hour.
As part of their study, the researchers evaluated the relationship between time of day and measurement-to-validation lags and found lags spiked right before nursing shift changes – even with automated patient monitoring devices integrated with electronic health records.
“After adjusting for location, shift period, vital sign type and abnormal values, mean lag increased by 5.2 minutes (95% confidence interval 5.13-5.30) for every hour closer to shift change,” the researchers said.
They concluded that nurse validation lags directly impact the timeliness of EHR-based early warning systems and “skew conclusions about patient status or the performance of monitoring infrastructure.”
Discrepancies between patient vital sign changes and when an EHR reflects them hold implications for both patient safety and health system performance, particularly where the EHR “serves as the foundational data layer for hospital operational analytics and AI-driven predictive models,” they said.
HIMSS research has shown the value of AI-enabled predictive analytics, even as those benefits are far from fully realized.
But when done effectively, the clinical and financial ROI can be substantial, as we showed earlier this year in our report on a codeveloped Hartford HealthCare and MIT platform that uses machine learning to optimize hospital operational efficiency – improving staffing efficiency and reducing length of stay.
Benchmarking an AI deterioration model
Meanwhile, another new study offers fresh insights into the performance of AI-powered deterioration models.
Based on two years of data from a multistate health system, the Duke Deterioration Index, a predictive machine-learning model developed by Duke Health to identify hospitalized patients at risk of clinical decline, was shown to catch 55% more deterioration events with fewer alerts than NEWS 2, a rules-based scoring tool used at the bedside.
“Across 118,063 patient encounters at five hospitals, the Duke Deterioration Index outperformed a widely used vendor deterioration index at every possible alert threshold,” said researchers from Vega Health, the Duke-sprung startup that works with health systems on the integration and monitoring of AI tools.
“At the same alert rate, the Duke Deterioration Index caught 55% more deterioration events than the vendor model. With a four-hour snooze on repeat alerts, the Duke Deterioration Index needed three-quarters fewer alerts to catch the same share of events: about six false alarms for every true positive, compared with 28 for the vendor model.”
However, the strongest results came from pairing the ML model with the chart-based scoring system, which Vega said offers new evidence that health systems don’t have to choose between the two.
“The combined approach captured more deterioration events than either model individually,” Vega researchers said in their new report, Retrospective Evaluation of Adult Inpatient Deterioration Models.
“Over the two-year evaluation, NEWS 2 alone predicted 64.7% of deterioration events. By adding Duke Deterioration Index alerts, the system may have captured an additional 19.8% of total deterioration events.”
Publicly benchmarking health AI models could bring a new level of transparency to providers in their choosing of AI models, according to Vega. The company’s platform pulls historical and real-time data from the EHR and standardizes it so health systems can see how their data feeds and AI models are performing.
“With so many options available, clinical leaders need a way to identify the optimal model and approach for their care teams and patient populations,” the researchers said.
