Predicting hospitalization risk in children with Type 1 diabetes

Children with Type 1 diabetes (T1D) walk a tightrope every day, relying on diet and insulin regimens to maintain metabolic balance and avoid complications such as diabetic ketoacidosis or hyperglycemia and ketosis without acidosis. When these complications develop, patients risk critical damage and usually must go to the emergency department to be treated.

Helping his patients avoid those crises is a years long mission of Soumya Adhikari, M.D., Pediatric Endocrinologist at Children’s Health℠ and Associate Professor at UT Southwestern.

To that end, he and his team, including Dr. Juan Diego Mejia Otero, a former fellow in Pediatric Endocrinology, and Dr. Perrin White, the Division Chief, have developed the first local predictive model for T1D patients at risk of future hospitalization that is validated with data from customized electronic medical records (EMR). The model uses three key risk factors – hospitalization within the past year, access to insurance and hemoglobin A1c levels – to calculate an individual risk score for each patient. The team recently published a summary of their process and findings in Pediatric Diabetes.

“Calculating the risk for each patient can show us where to direct limited resources, such as real-time glucose monitoring and mental health services, and help our patients avoid medical emergencies,” says Dr. Adhikari. The team is now in the process of implementing individualized care pathways based on a stratification of their patients’ risk scores.

Enhancing patient records for better data collection

Ten years ago, Dr. Adhikari and his team weren’t thinking about predicting risk. Their goal was to collect data on a variety of treatment and outcome metrics – ranging from A1c levels to depression scores – so they could measure how well their patients and practice were doing.

“I wanted to be able to look back in four years and say, ‘We're doing better than we were four years ago,’” Dr. Adhikari says. “To do that, I needed proof.”

The problem was, they couldn’t get the data they needed from the patients’ records. Information was formatted inconsistently, and many metrics weren’t being recorded at all.

“The first time we tried to data mine the medical record, we couldn’t even systematically identify what kind of diabetes a patient had, let alone what their lab results were,” says Dr. Adhikari.

They spent years redesigning the medical record and training providers throughout the hospital to use them as intended. By 2014, the team had created a record that tracked the data they needed about each patient.

“Once we had the data, we were inspired to use it not only to measure the past but predict what might happen in the future,” Dr. Adhikari says.

Calculating risk of future hospitalization

As recently chronicled in Pediatric Diabetes, the team started by studying patient data from 2014-2018 to see which individual metrics in their database were strong predictors of diabetic ketoacidosis and hyperglycemic ketosis without acidosis. They identified five risk factors: high A1c levels, female sex, non-White race, possession of non-commercial insurance and hospitalization during the preceding 12 months. The strongest predictor was prior hospitalization, which netted a four- to five-fold increase in the risk of future hospitalization.

Then the team searched for the combination of variables that posed the most accurate predictions. The winning model combined just three factors: A1c, insurance status and previous hospitalization.

They trained their model against two years of data – for example, using 2014 data to predict outcomes in 2015, then comparing the prediction to known outcomes for 2015 – and validated it against the next two years of data. In the validation stage, the model identified 243 patients at high risk of hospitalization, of whom 92 (38%) actually did end up being admitted in the next year. Overall, the model is 95% accurate at predicting who won’t be hospitalized and 30% accurate at identifying who will.

While previous research has pointed to the same risk factors used in the model, Dr. Adhikari’s team is the first to calculate how much a patient is at risk based on cohorted data.

“It's one thing to say, ‘We think this child is high risk,’ and another to say, ‘This child is five times more likely to suffer a crisis,’” Dr. Adhikari explains. “That’s information you can act on.”

Goal: guide treatment decisions

The next step for Dr. Adhikari’s team is to use their model to guide treatment decisions. The most obvious course is to connect high-risk patients with education and encouragement about keeping up with their diet and insulin routines. Social workers and child psychologists may also step in to offer mental health services.

Children with diabetes experience depression at twice the rate of children without diabetes. Research shows they factor their disease into more than 100 decisions they make each day, because their diet and blood sugar levels affect their capacity to do everything from play sports to take a test at school.  

“Relieving the mental health burden on our highest-risk patients could keep them out of crisis,” says Dr. Adhikari.

Another area of support could be real-time monitoring. Continuous glucose monitors were only recently approved by Texas Medicaid, opening a new level of care to patients with noncommercial insurance. The team at Children’s Health has seen a several-fold increase in the adoption of these technologies since spring 2020, when they first gained approval from Texas Medicaid.

While new treatment pathways take time to emerge, Dr. Adhikari says that having quality data is a good place to start.

“Just beginning to ask those questions – who needs our attention most and what kind of attention do they need? – is helping us develop the quantified, personalized practice we want.”

Learn more about the innovative endocrinology care and research at Children’s Health

Research Library

Find the latest innovation, clinical trial and research articles.