At Intermountain, Predictive Analytics Boost COVID-19 Outcomes

In healthcare, predictive analytics tools have always been one of the most promising applications of big data. The ability to anticipate negative consequences before they occur is a major asset for any organization looking to improve patient care.

With all the variability that’s come with the pandemic, predictive analytics models have only grown in prominence and prevalence. Healthcare entities have developed predictive algorithms to envision the spread of the virus, estimate resource demands, and perhaps most importantly, calculate COVID-19 outcomes and patient risk.

However, not all of these tools are created equal. Despite the advancement of the technology – and the urgency of the current healthcare crisis – predictive analytics models still must be trained and validated on accurate, quality data.

At Intermountain Healthcare, leaders understood that this solid data foundation was critical. The team kept it in mind when designing three predictive analytics tools to forecast COVID-19 risk and identify the best treatments for individual patients.

One of the models Intermountain developed is a biomarker tool used to risk stratify patients who have severe COVID-19. The algorithm is based on widely available laboratory data and can determine a patient’s risk of needing mechanical ventilation or dying after they are admitted to the hospital.  

“We’ve used this tool because biologically, it’s also accurate in identifying patients who have a high risk of developing hyper inflammatory syndrome from the virus,” Brandon Webb, MD, an infectious diseases physician and chair of the COVID therapeutics committee at Intermountain, told HealthITAnalytics.

“We know that those are the patients who would benefit the most from anti-inflammatory treatments like steroids or tocilizumab.”

The second tool stratifies COVID-19 patients who are not admitted to the hospital but are at high risk of developing more severe disease. With this tool, caregivers can proactively connect to these patients and provide them with the appropriate care – namely, monoclonal antibody treatment therapy.

“This tool is used for identifying patients who are at high risk for hospitalization and would therefore benefit from drugs that prevent hospitalization, like monoclonal antibodies,” Webb explained.

“We have actually derived and validated that tool in a large Utah cohort of more than 20,000 patients with the intent that it would be used by providers throughout the state. We intentionally adapted it to become a simple points-based scoring model so that it wouldn't require advanced analytics or EHR integration – providers can calculate it with a pencil and pad if needed.”

For the third tool, the team tailored an individual risk prediction model to help determine whether COVID-19 patients are at risk for being admitted to the ICU.

“In developing the model, we identified those patients that we considered to be at high risk for hospitalization from COVID-19. We then collaborated with our population health company to engage in outreach activities with those patients – including texts and phone calls – to help high-risk individuals understand how to best take care of themselves and avoid contracting the virus,” said Greg Nelson, assistant vice president of analytics services at Intermountain.

“We worked with our clinicians to identify those factors that were important and to ensure the tool was accurate. Our individual risk prediction model actually worked really well in this patient population because the same risk factors for hospitalization were also the same risk factors that often result in elevated levels of care.”

Throughout the pandemic, these models have played an instrumental role in ensuring all patients remain healthy.

“Being able to avert hospitalizations is a really important factor,” said Nelson.

“We want to keep people out of the health system and have them stay well. This is especially true for our most vulnerable populations who otherwise may not have access to care or the ability to afford care. So, these are critical tools in our arsenal.”

These tools have also helped Intermountain identify and address health disparities. The models have incited the health system to develop strategies related to outreach, testing, and treatment as the pandemic has worn on.

“When the treatment isn't unlimited, we get into a situation where we want to target our treatment to the patients who are most at risk and most likely to benefit. During the pandemic, we found that for reasons we don't fully understand, persons of color are at higher risk and therefore at higher likelihood of being hospitalized,” said Webb.

“When these monoclonal antibody products became available, we said, ‘Look, we need a way of identifying high risk patients and targeting limited resources towards those patients,’ and these types of risk prediction or risk stratification models do just that.”

In this case, adding patients’ social determinants data into predictive tools served to help, not hinder, prevention and treatment efforts.

“Historically, we don't usually include race or ethnicity in predictive models for obvious reasons. We don't want to perpetuate the fallacy that race or ethnicity are biological constructs, and we don't want to perpetuate any type of profiling behavior,” said Webb.

“But when a prediction model is intended to connect at-risk patients with potentially effective, preventive therapies, then it is appropriate to include this kind of information in the model. For that very reason, we did include race and ethnicity in our risk prediction model so that a patient of color who is statistically at equivalent risk of being hospitalized at a much younger age, for example, isn't penalized systematically from having access to treatments that do work.”

Organizations developing predictive models – for COVID-19 or otherwise – should also remember that these tools are only half of the equation.

“It's always a balance between good clinical decision support and identifying the opportunities. You want to empower clinicians who know how to deliver the best care with effective analytics techniques,” Nelson said.

“Analytics give us a lot of leverage. These tools enable us to understand multiple patient populations and help us use the power of that data. My advice is to get started, but to also understand that this technology is not a replacement for human intelligence in decision-making. It's about augmenting those decisions.”

Building responsible AI tools is critically important as well, Nelson noted.

“At Intermountain, we have developed a playbook for the use of artificial intelligence and machine learning to ensure that our models are responsible, ethical, and unbiased. We want to make sure that the sample data we're using to train our models is truly representative of the population,” he said.

“The other side of that is helping clinicians and others understand what's actionable. We could predict something with a neural network, but it may not be terribly explainable. We always are very conscious of making sure that the model has some interpretability, so that it resonates with the people who are supposed to act on it.”

Webb added that it’s crucial for healthcare leaders to acknowledge that risk prediction and AI tools still aren’t perfect or infallible.

“It is really important to be responsible and transparent in communicating what the limitations of this technology are. A lot of the limitations are often intrinsic to the data that feeds the model. If the data is simply not complete enough to describe the clinical scenario that you're hoping to interpret and predict, then no level of complexity will make up for that,” he said.

“That's really important for the public and executive leaders to understand. While our technology has improved, the effectiveness of these tools still depends on the quality and completeness of the data.”

While COVID-19 has brought significant challenges and longstanding issues to the forefront of the healthcare industry, the crisis has also accelerated the development of innovative tools and strategies.

“The pandemic has resulted in knowledge gain of a magnitude that we've never seen in medicine before. As part of that knowledge gain, we've seen innumerable different types of prediction models proposed and published. In some cases, as the cream rises to the top, those that are best validated and accurate have actually been implemented in some cases,” said Webb.

“The pandemic has accelerated not only our ability to build these tools, but also C-suite awareness and awareness among the public about the power of analytics and risk stratification models.”

As the pandemic continues – and hopefully, begins to wane – big data analytics tools will remain a vitally important part of care delivery.

“The pandemic introduced a new vocabulary and a language around uncertainty in healthcare. Organizations had questions about how many patients they were going to get, how many people would contract the virus, and how many people would be hospitalized. Even during vaccinations, there's a lot of uncertainty about human behaviors. Will people get the vaccinations? Will they go back for the second dose for those therapies that require it?” Nelson concluded.

“Data analytics and predictive models are really good at helping us share in a language with our leadership about making decisions in light of uncertainty. People have a growing appreciation for the power of analytics in general.”