Faces of digital health

View Original

Taiwan: The State of Digitalization, AI and What Went Wrong With COVID? (Yu Chuan Jack Li)

Taiwan spends only 6.4% of its GDP on healthcare, but has high satisfaction rates with healthcare, and is also very digitalized. Taiwan was successful and an exemplary country at managing the COVID-19 pandemic in 2020, and later experienced a huge wave of infections in 2021. What happened?

Professor Yu.Chuan Jack Li is Editor-in-Chief for BMJ Health & Care Informatics journal, the elected president of the International Medical Informatics Association (IMIA) and has devoted himself to evolving the next generation of Al in patient safety and prevention ("Earlier Medicine"). He has been deeply involved in biomedical informatics development in Taiwan and international cooperation on various continents, including Asia, America, Europe, and Africa. 

See this content in the original post

State of healthcare IT

As he explains, the universal healthcare system in Taiwan triggered a movement towards digitalization already in the nineties. 600 hospitals and 20,000 clinics around the country are required to submit insurance claims in the electronic pool. That triggered hospitals to install computer systems, not just for billing but also for patient care. 

Taiwan achieved close to 90% adoption of its Electronic Medical Records System in its medical facilities and managed to transition from Stage 6 to Stage 7 as a nation in 2019 under the HIMSS Analytics Electronic Medical Record Adoption Model (EMRAM). However, Prof. Jack-Li warns that healthcare IT is underfunded and consequently vulnerable. “The banking industry invests about 10% of their revenue into their IT systems. In Taiwan, we invested only 1- 1.5% of the revenue into hospital IT systems, in the US it is about 3%, in the Europe they're aiming for 5%, but in most countries, most hospitals don't have that investment.”

Secondary use of data in Taiwan

Hospitals use different IT systems but are required to use common standards when sending the data out of the hospital. Taiwan also maintains a central repository of anonymized data for research. “It's under quite strict restrictions. When you go to a data center, you can’t bring anything with you.  You can not bring your cell phone in. So if you go in there and you memorize what you're trying to do, you can access all the data in the data center. Then all the results have to go through a committee that approves that the research is not violating privacy or ethical concerns, in order to take the research results out the center. When you go in there, the access you have is amazing. You have access to people’s health data for the last 25 years.  That means 12 billion, 15 billion outpatient visits, hundreds of millions of inpatient stays and all of that, and it's a very big database, so it takes a lot of computing. This database alone supports the creation of about 1,200 papers every year,” explains Prof. Jack-Li. 

The COVID-19 Pandemic in Taiwan

The rates of infections in Taiwan in 2020 and 2021.

Taiwan was among the countries with the best COVID management approach in 2020. As explained by Prof. Li, the country stopped flights from Wuhan to Taiwan approximately two weeks earlier than the rest of the world. This, and the country’s past experiences with managing epidemics led to a feeling of overconfidence in 2021. Because the country avoided a wave of infections in 2020, there was not a lot of interest for vaccinations in 2021. Once the virus started spreading in the country in May 2021, the demand for vaccines rose, but the country didn’t have a sufficient supply of vaccines for fast vaccination and eventually did suffer from a COVID wave.

“Doctors hate alerts 99% of the time”

Among Prof. Li’s research focuses is patient safety. Because he still works in the clinical practice and because he held several leadership positions in healthcare, he has a deep understanding of creating meaningful decision support systems. A good example is medication management.

“I'm also a clinician, so I also see patients twice a week, three times a week. Alerts and reminders are usually a pop-up dialog that stops you from doing things and you have to read and say, okay, or whatever responds to you. As doctors, we hate these alerts 99% of the time. When an alert is right, we love it. But when they got it wrong, it's a total waste of time. So every time I get this kind of alert, I am angry inside for three seconds, and then after I get back, I forget what I was doing with the patient and I have to rethink everything. Of course, after 20 years, those three seconds become 0.3 seconds, but still it's uncomfortable to be interrupted when you're doing something with full concentration and suddenly there's a stop sign that's really not important at all. In the hospital that I'm working in we published a paper that the physicians received 2 million reminders in eight months. And it's not even a huge hospital, it has 300, 400 physicians. If every reminder stops the physician from doing their job for three seconds and you multiply that by 2 million, that's 6 million seconds. It's a lot of physician time and it's quite expensive. So our warning systems are actually making hospitals lose money because they're wasting precious physician time,” explains prof. dr. Jack-Li.

Why is it Difficult to Create an Accurate Drug Interactions Decision Support Tool?

In 2007, when prof. Jack-Li was holding a position of Clinical Information Officer, he wanted to design a full-blown drug interaction alerts in his hospital.

“There were 12,000 different drug interactions in the decision support system. If we let that be, physicians would get 10 reminders with each patient. I talked to the pharmacy department and said let's look at the 12,000 interactions. How many are clinically significant? After a while, they came back to me and said it’s 700. Then I asked them how many of the 700 can happen in our hospital because we don't have all the drugs in the world, the answer was about 400. Then I asked them about the 100 clinically significant alerts that happened most frequently.  They told me that one of the most frequent interactions is digoxin and diuretics. If you put these two types of drugs together, it's very likely that you will suffer low potassium levels. So I asked pharmacists if we implement that alert, how many alerts are we going to get? They said hundreds or thousands per month. And I asked them who will get these alerts, and it was cardiologists. So I called the cardiologists and I said, guys, don't, you know that diuretics together with digitalis cause hypokalemia? They said, of course, we know. Then I asked them, then why aren't you still doing it? They said, of course, we add potassium to our prescriptions if we prescribe digitalis and diuretics. Then I said, oh we didn't know that. We did not check that. We only looked at the prescription and said, you have digitalis, you have diuretics, you should get an alert. But cardiologists were already mitigating the problem by adding potassium to their prescription. So for them, it's a totally safe order. But for the pharmacy department, it's a very dangerous structure interaction.”

The Hope of AI in the Future Decision Support Systems

The challenge, says Prof. Jack-Li is that different experts work in silos. AI could potentially address these challenges and make clinical decision support more clinically relevant. “AI, especially machine learning-based AI, is very capable of handling a high dimension of variables. They could actually handle millions of variables if you wanted. I think there are several things that need to be done to leverage that potential. One: the clinical department has to look at different sources of data, but usually that's getting out of their expertise. They're either an expert in lab results, they're experts in medication, but they're not experts in disease and don't want to get out of their comfort zone. But they will have to get out of their comfort zone and really look at all that data. The second thing is: the department has to know how to adapt new AI methodology that could handle the dimensionality of variables.”

Tune in to the full discussion. 

See this content in the original post