Why Can’t Patient Outcomes Be Predicted More Like the Weather?

Weather predictions are notorious for changing — a morning forecast for sunny skies might be updated at noon to call for rain. But as inconvenient as it may be to leave your umbrella at home, these changes reflect how the forecast’s accuracy is dynamic. The weather models these forecasts are based on are constantly being updated with new data about changing climate conditions.

Cleveland Clinic researchers wondered if the same approach could be applied to patients in the ICU. In a new paper appearing in BMJ Open, they demonstrate that dynamic modeling can predict patient outcomes with 90% accuracy, significantly higher than traditional, static models. They also show that dynamic modeling becomes even more accurate over time.

“We were surprised by how good the results were,” says lead author Abhijit Duggal, MD, a pulmonologist at the Cleveland Clinic. “This is as strong as it gets.”

Moving beyond snapshots

Conventional models fall short because they base predictions on snapshots in time, explains Dr. Duggal. Most predictions are based either on the patient’s condition and risk factors on their first day of treatment or on samples of patient data taken over several days. A critical flaw in these models is that they assume the patient’s trajectory will continue in a straight line, he says.

“We know for a fact that’s not the case because patients are very dynamic,” explains Dr. Duggal. “Even on a day-to-day basis, you can see massive changes happening based on their disease severity and other factors that come into play.”

Another issue with traditional models is that they typically focus on the patient’s outcome: recovery or death. They provide little information about how the patient’s condition may change daily.

To apply dynamic modeling to the ICU, researchers looked towards weather forecasting and other fields that use models that reflect changing conditions.

“These models aren’t new, but they haven’t been applied in a clinical setting,” Dr. Duggal says.

Researchers also looked at how these models could be used to predict short-term changes in the patient’s condition, and whether the patient would improve, deteriorate, or remain stable over the next 24 hours.

A model that improves over time

For the study, the team analyzed data from more than 5,200 patients who had been admitted to the ICU with COVID-associated pneumonia. They developed dynamic models to predict the disease trajectory over the first seven days of the patients’ ICU stay and compared the predictions with conventional, static modeling.

Researchers found that both approaches started well and were able to predict the patient’s first 24 hours with 80% to 90% accuracy.

“However, with each passing day, the ability of the static models to predict outcomes diminishes,” Dr. Duggal says.

Meanwhile, the dynamic models continued to improve, predicting patient outcomes with more than 90% accuracy after six days.

Researchers then compared the results with data about clinical interventions and found a very strong correlation; patients who were predicted to be deteriorating saw an escalation of care, while patients who the model predicted would be stable or improve saw de-escalation.

“It really shows we are identifying the right patients and correctly identifying their disease trajectory,” Dr. Duggal says.

Next, the team plans to study dynamic modeling in greater detail and in wider patient populations with the goal of developing tools for clinical use.

“If we could tell clinicians their patient has an 80% chance of deterioration in the next 24 hours, or an 80% chance of improvement, that will have a significant clinical impact on care,” he says.