A nurse who works at Hallym University Chuncheon Sacred Heart Hospital was at her nurse station looking at the electronic medical record (EMR) screen as usual. Suddenly, she noticed that the status of a diabetic patient was different from usual. On one side of the screen, the patient's status was displayed in red, indicating a high probability of hypoglycemia. The nurse immediately rushed to the ward where the patient was hospitalized, explained to his guardian that he could have a hypoglycemic episode, and gave him a snack. The patient’s blood sugar returned to normal, and he could avoid hypoglycemic shock.

That’s what happened recently at the hospital. It is an example of a hospitalized patient's risk being predicted in advance without medical staff needing to visit the patient or for the patient or guardian to seek help from medical personnel.

This was made possible by an “AI prediction model” that Hallym University Medical Center developed in 2019 and introduced to five of its hospitals. The AI prediction model is a system that calculates the probability of 42 symptoms in real-time, ranging from falls and bedsores to delirium, diabetes complications, and dysphagia.

Various AI solutions have recently been introduced into the medical field, but Hallym University Medical Center's AI prediction model has attracted attention because it was developed based on its data and experience in the field. According to Hallym University Medical Center, the response from medical workers and patients after the introduction of the AI prediction model has exceeded expectations.

Son Eun-jin, director of the Department of Nursing at Hallym University Chuncheon Sacred Heart Hospital, who is actively using AI prediction models, made a similar assessment and elaborated on it during a recent interview with Korea Biomedical Review .

“It was the nurses who first recognized the need for an AI system (to predict patient conditions),” Son said.

When the medical center developed the Order Communication System (OCS) and EMR, the nurses at the affiliated hospitals participated in the process and expressed their opinions. However, as they used the programs, they asked for more features, which led to the development of AI prediction models.

“Based on various patient indicators entered into the EMR, we came up with the idea of ‘what if we could predict the patient's condition in advance and prepare for it,’” Son said. “When the needs of the field and the medical center's goals were aligned, we started developing the AI prediction model in 2019 and introduced it to the field in 2020.”

The AI prediction model uses machine learning technology to learn from 10 years of patient data recorded in the EMR (medical specialty, age, gender, day of visit, diagnosis code, and others).

After continuous development based on the needs of the field, the AI prediction model currently predicts 42 conditions, including falls and pressure ulcers, narrowing of arteriovenous fistula in dialysis patients, phlebitis, hypertension complications, diabetes complications, CRE-CPE, pressure ulcers in emergency room patients, and delirium.

Head Nurse Koo Hyun-joo demonstrated on screen how the AI prediction model works. She selects a patient in the EMR system and shows the possible conditions and their predicted rates. Nurses could also see the status of hospitals that have adopted AI models for medical prediction, learning variables, and average prediction rates.

So, are there any errors in these predictions? In healthcare, as in any other field, errors can be fatal and require extreme caution. Naturally, the nurses in the field were even more aware of this.

“It wasn't done perfectly the first time,” Director Son said. “The nurses proposed as many different learning variables as possible, but in some cases, there were limitations (in the prediction rate).”

The medical center developed them by referring to existing foreign cases and papers and communicating with the Information Management Bureau. As a result of these efforts, she said the prediction model currently has an average prediction rate of 87 percent for 42 symptoms.

‘Preemptive responses are possible as proactive measures manage patients efficiently’

Since implementing the AI prediction model, the hospital has been able to preemptively respond to abnormalities in patients' conditions and prevent them from escalating into larger outbreaks, Son said.

According to an AI prediction model satisfaction survey conducted by Hallym University Gangnam Sacred Heart Hospital in August on nurses working in 108 wards, 97 percent of respondents were satisfied with the AI prediction model. The reasons cited for their satisfaction included “being able to grasp the patient's condition in real-time,” “customized patient management 24 hours a day, 365 days a year,” and “lower incidence of severe illness.”

“We can see the patient's condition in real-time and take proactive action based on the predicted rate, which may seem like more work at first glance,” Director Son said. “However, if the patient's condition worsens, the medical team must jump in. Acting proactively reduces the workload that could have been far heavier otherwise.”

AI predictive models have also helped fill in some of the gaps during the recent medical crisis.

“Part of trainee doctors’ job is screening, and the nurses in the field are using AI prediction models (to replace it partially) to solve the patient's problem according to the prediction rate or to notify the professors so they can act quickly,” Son said. “AI prediction models are helping us a lot.”

“I think it can compensate for the gap of medical staff to some extent in terms of predicting the patient's condition in advance and managing it so that it doesn't get worse,” she said. “I think it will also help build an environment where doctors can focus on their training.”

Patients and caregivers were also more reassured when informed of predictive systems such as AI prediction models.

“For diabetic patients whose AI predicts that they are likely to experience hypoglycemia, we notify their caregivers in advance to remind them to take a snack or inject glucose in consultation with their doctors,” Son said. “We also have an aspiration pneumonia prediction model, and if the prediction is high, we manage them more carefully, including suctioning, more frequently.”

“When we tell patients and their guardians that we are proactively responding with AI prediction models at the time of admission, both patients and guardians are relieved,” she said. “Also, when I go to check on patients after seeing the results of AI prediction models, patients ask me, ’Why did you come here if you're okay,' and they are happy when I tell them that I came because of AI prediction models.’”

Son emphasized that Hallym University Medical Center's AI prediction model is an example of how nursing record data can contribute to patient safety, and she hopes that other hospitals will follow suit.

“Nurses sit in front of a computer for 70 to 80 percent of their time. I want them to realize that the nursing records they fill out every shift can be developed into quality data that can be used to impact patient safety positively.” Son said. “Hallim University Medical Center's AI prediction model is a case in point. We hope that such cases will spread.”