“We can now proactively act to prevent adverse health events at scale, based on prediction who is most likely to suffer osteoporotic fractures, cardiovascular events, or severe influenza complications before they happen,” says Prof. Ran Balicer, Founding Director of the Clalit Research Institute and CIO of Clalit Health Services. In an interview with ICT&Health Global, he explains how AI is helping physicians identify high-risk patients earlier and bringing predictive healthcare into daily clinical practice.
We’ve gotten used to a healthcare system where we go to the doctor when the first symptoms occur. Usually, we get prescribed medication or go to the hospital, and hopefully, we get better quickly. What’s wrong with this procedure?
This is basically the mainstay of healthcare today, which is fundamentally reactive. If you want to think about what’s wrong with it, I’ll ask you this: imagine you have a car, and you keep driving it until smoke comes out of the engine. Only then do you go to the repair shop. But why wait until smoke comes out? Obviously, you feel that’s wrong. That’s not the way to maintain your car, or an entire fleet of cars.
I think it’s the same thing when it comes to the most important thing we have, which is our health and the health of the people we love. It doesn’t make sense in this era to provide care only when patients already feel something is wrong, when there is already dysfunction, pain, or pathology that has become entrenched. Providing care at this point is, in most cases, associated with only increased complexity of care needed, increased costs, and only partial relief.
We should be providing predictive, proactive, and preventive care. And the technology for that has been with us for at least two decades. I know that because Clalit has been providing this type of care for more than 17 years already, using data from electronic health records. We had electronic health records dating back to the 1990s, so by the end of 2009, we already had at least a decade of data available. We used it to identify which patients were about to deteriorate, which patients were likely to develop chronic kidney disease a year or two later, and reach out to their general practitioners to advocate for proactive care.
This kind of data allows us to provide a completely different type of care. When care is proactive, it first and foremost provides our patients with the best possible care. It is also less costly, less painful, and ultimately far more effective in improving outcomes. And we now have many examples showing how this moves from theory into actual practice at the national scale.
The car analogy captures the problem well. We take our cars for annual checkups, and some people even clean them every Saturday, but we forget to take care of our health.
Think about it: in your car, you have all these little warning lights, right? One tells you the oil is low, another tells you the tire pressure is low. Nowadays, with AI, and even two decades before AI, once we have updated health data available in digital records, we can have this type of ‘warning lights’ for the human body, in most cases, not just for cars.
Let me give you one example of how we implemented this on a scale. One of the problems we face at Clalit, like many countries around the world, is hepatitis C. It’s a viral liver disease that, if left untreated, dramatically increases the risk of liver cancer and liver cirrhosis, both of which are severe and very costly diseases.
The good news is that we have a cure. There is a drug that, if I have been infected with hepatitis C and receive treatment, cures me in 98.8% of cases. That means I will not go on to develop those serious illnesses. The challenge is that hepatitis C often goes undetected until it is too late. People don’t know they have it.
Every year, Clalit screens around 50,000 people to identify silent carriers of hepatitis C. Out of those 50,000 first-time screenings, we found only 38 positive cases. Sure - we cured these 38 patients and changed their life course. But only in 38 cases. So the question became: can we do better than that?
Back in 2017, our research group at the Clalit Research Institute, led by Prof. Noa Dagan, developed a predictive model that could identify which patients were most likely to have hepatitis C. We screened the first 500 people on that list. Guess what? We found 38 positive cases within those 500 people alone.
So instead of screening populations and finding one positive case in every thousand people, we were suddenly finding one positive case in every ten people screened. That’s a hundred-fold improvement. It gives us a much better chance of meeting the World Health Organization’s goal of eradicating hepatitis C.
And this is not some theoretical construct. Clalit now uses these tools every single day. Our home-developed AI systems are embedded directly in primary care clinics, helping physicians by providing patient-specific recommendations on which patients require treatment and what actions should be taken. We joined know-how from different divisions and professions - community medicine, innovation, and information technology, to turn this vision into a daily reality.
Just this month alone, more than 250,000 care gaps were closed based on AI-generated recommendations that physicians reviewed and accepted. Physicians said, “Yes, I understand this recommendation because it is explainable AI. I trust it, and I want to act on it.” And those recommendations immediately moved into practice. In each and every such case, the patient is notified via SMS that their GP consulted with the Clalit AI and that the AI is suggesting a care modification. The patients are empowered to take an active part in this process.
Clalit is often seen as a global leader in data-driven preventive healthcare. What was the turning point when Clalit realized healthcare had to move from reactive care to predictive care?
Like every healthcare system, we faced several major challenges. One challenge was staffing. No healthcare system in the world has enough staff. Right now, there are around 10 million physician positions missing globally, and the shortage keeps growing.
We understood that as populations age and chronic diseases become more common across all age groups, we should have to find a different way to provide the same level of care with a shrinking workforce of physicians, nurses, and allied health professionals. That was one major pressure point.
The second pressure was the mounting cost of healthcare. We know that when medicine evolves from molecules to antibodies, and from antibodies to cell therapies, each transition can increase the cost per patient tenfold. There is, therefore, a huge financial imperative to make healthcare systems far more effective with the funding available. And of course, at the end of the day, our goal is always to improve patient care.
We realized that staying sustainable required fundamentally transforming how care is delivered. We had to move toward a modern healthcare model that is proactive rather than reactive, precise rather than intuitive, and data-driven and personalized, rather than relying on a one-size-fits-all, cookie-cutter approach in which a single guideline is applied equally to everyone.
Once we understood that, the real work was implementation. And the most important part of implementation is building tools that physicians and nurses will actually accept, adopt, and trust.
We approached this very carefully and gradually over many years. We built these tools together with frontline workers and for frontline workers. We continuously adapted them to fit existing clinical workflows, making sure the technology genuinely helped people do their jobs better.
Because if it simply feels like management introducing some shiny new tool and hoping people will use it, it will never work.
What did you do differently so that these initiatives didn’t end up as just another pilot project that dies after it’s completed because the financial model and long-term strategy are missing?
We actually call it “pilotitis.” It’s like an acute inflammation of an organization caused by too many pilots that lead nowhere.
The way we approach this problem is that when you start what people call a pilot, you cannot really think about it as a pilot. From the very beginning, you need to ask yourself: What are the project's scaling prospects?
We see the pilot not as an isolated effort but as step one in a sequence in which steps two, three, and four are already obvious. The checkpoint for moving from step one to step two must be predefined, and the funding source and organizational commitment for scaling must already exist.
So, you must ask yourself: if this pilot succeeds, what comes next? What is the scaling program? Who is going to pay for it? Who is going to sponsor it organizationally? And if you don’t have answers to those questions, then don’t start the pilot at all, because otherwise it becomes the graveyard of enormous effort and frustration.
At Clalit, we established what we call the Innovation Division, which I chair. Its role is to facilitate this work, help fund it, and provide operational support. We also created a network of innovation and research centers across all our hospitals, which we financially support to ensure these projects are implemented properly. And then, every successful project is scaled.
One of the key questions you mentioned is: who will pay for this? Prevention sounds politically attractive, but it often clashes with existing reimbursement models. How did Clalit make predictive healthcare economically viable?
We are very fortunate to operate in a country where the healthcare financing system differs from that of many other countries. In Israel, every citizen is entitled to a broad basket of healthcare services provided by one of four healthcare organizations, which patients can freely choose from. So there is strong competition and a high degree of patient choice, which is always beneficial.
At the same time, the funding comes from the government through a capitation model. As Clalit, the largest of these four organizations and the care provider for more than half of the population from cradle to grave, I don’t need to worry whether my patients are rich or poor. The funding does not come directly from them. It comes from the state.
Our responsibility is to provide the best possible care and compete on quality, service, and outcomes. Because we have cared for patients for many decades, we have a strong incentive not just to deliver care today or maximize volumes, but to keep people healthy.
This incentive structure, which has been thoughtfully built over the years in the Israeli healthcare system, aligns our interests with those of patients. Patients want to remain healthy. We also need them to remain healthy for the system to remain economically sustainable.
Our model is not based on filling hospital beds and generating revenue from hospitalizations. In many ways, it’s the opposite. We succeed when patients stay healthy over time.
Can you share one more example of a predictive model implemented in Clalit that actually led to better patient outcomes?
One of the first questions you need to answer when implementing proactive care is: who should receive proactive care first? Because, for now, we simply do not have enough resources to proactively manage everyone at the same time.
That’s why we built an entire set of predictive algorithms across different medical domains. We can identify which patients are most likely to develop severe osteoporosis and suffer fractures. We can predict who is at the highest risk of a cardiovascular event in the near future. We can even estimate who is most likely to develop severe influenza complications during the winter and therefore would benefit most from vaccination.
All of these predictive models come together into a unified system that ranks patients for each primary care physician. This allows physicians to immediately see which patients require attention first.
But knowing who to treat is only part of the challenge. You also need to know what action to take.
To do that, our AI system reviews the complete healthcare record of every patient every night. It cross-references that information with the full body of medical knowledge embedded into our models and then generates concrete clinical recommendations that are presented to the physicians.
Let me give you a practical example. Suppose I had a routine blood test yesterday, and the results suggest that my kidney function has deteriorated. My eGFR value drops below 30, which means my kidneys have decreased function. My physician may only review those lab results three days later, and even then, may not immediately remember that I am diabetic and currently taking a diabetes medication that is no longer appropriate for someone with impaired kidney function.
On the very same night the blood test is performed, our system flags the patient, identifies the likely onset of chronic kidney disease, recommends adjusting the diabetes medication, and even suggests which alternative medications could be considered instead.
All of this information is automatically prepared and delivered to the physician in a filtered, structured way. The physician reviews it, approves it, and then a message is automatically sent to the patient saying: “Your physician has used AI to review your health record. There is a new treatment recommendation. Please come in so we can explain the suggested changes.”
And again, this is not theoretical. We handle around 250,000 cases like this every month, and the number continues to rise.
The same principle is already implemented in hospitals. For example, if you go to a Clalit hospital emergency department today and receive an X-ray of your hand, an AI system will automatically highlight potential fractures for the clinician using a small yellow rectangle on the image.
This is now fully integrated into routine care. It is almost impossible to perform imaging without receiving some form of automated decision support. At the end of the day, however, the physician still makes the final decision. AI is there to support expertise, not replace it.
We would never ask airline pilots to fly blindly simply because they are experienced. They are highly trained professionals, but we still provide them with every instrument and every piece of information available.
Physicians, nurses, and pharmacists are no different. Our responsibility is to help them stay at the very top of their profession, and that is exactly what we are trying to do.
So, to implement AI models and AI algorithms in clinical settings, you need to have everybody on board, from clinicians and nurses to patients, healthcare systems, and regulators. The question is: were physicians enthusiastic about AI from day one?
I’m a physician myself, so I can say that we are hard to convince, and justifiably so. We’ve been disappointed too many times by promises that never materialized. So, whenever a new gadget, a new tool, or a new recommendation comes along, we need to see, in a meticulous and evidence-based way, that we can actually trust it, and that it won’t create more problems, consume more time, or waste more energy.
Those are the two things physicians need to feel before they are willing to adopt something new.
That’s why, when we built these tools, we designed them around clinicians’ needs and created a strong evidence base from the beginning. We engaged key opinion leaders across all domains to validate the algorithms and confirm they work. We used clinicians themselves as designers when shaping the workflows, and we ensured these systems would not add time to clinical work but would instead save time.
We now actively allocate time for proactive care within physicians’ schedules, and we pay for it. We tell physicians: take these dedicated hours each week, not for opportunistic visits, but to proactively bring patients in.
Patients love it, and honestly, physicians do too. At this point, it has become part of the system itself.
To make this work at scale, however, you also need strong governance. And governance in AI is still a huge gap today. It’s extremely difficult.
This could not have happened without a unique act of joining forces – our community clinical leadership, our IT and development division, and our clinical innovation team. If you want to go far, you must go together, as a group.
Because we have been working in this field for nearly 20 years, we have gained significant experience managing AI systems long before AI became fashionable. Back in 2010, we didn’t even call it AI. We called it data mining, logistic regression, or predictive analytics. Then we moved into machine learning, deep learning, and now foundation models. Today, one model can generate 1,200 predictions in a single click.
But ultimately, the terminology doesn’t matter. Technology evolves, but the central question remains the same: can you trust the output?
Clalit has even created a governance framework for introducing AI in clinical settings.
You mean OPTICA, which stands for “Organizational Perspective Checklist for Artificial Intelligence Solutions Adoption”. It was published about a year ago in the New England Journal of Medicine AI edition, and it is open for everyone to use. Optica is essentially a set of criteria, checklists, and decision workflows that healthcare organizations can use when developing AI systems internally or purchasing them externally. It helps organizations systematically ensure that what they are implementing is truly responsible AI.
For example, you need to verify that your data matches the type of data the model was trained on. You need to make sure clinicians can integrate the tool into their workflows. You need to ensure cybersecurity, data privacy, proper machine learning operations, and reliable model performance on testing datasets.
But perhaps most importantly, you need a continuous plan for follow-up, monitoring, and evaluation, because AI systems are dynamic. You can deploy a model today and see excellent performance, but one year later, its behavior may change dramatically, even though nobody intentionally modified it.
We have case studies showing how changes in datasets or organizational workflows have completely transformed the accuracy of predictive models over time.
If you do not build a comprehensive governance framework for AI, you will eventually run into trouble. Today, every single AI solution used at Clalit, whether developed internally or purchased externally, goes through the OPTICA framework. Other health systems around the world have adopted it as well, and governments are now beginning to implement similar approaches.
Where should a healthcare organization that wants to apply AI start?
My first recommendation is simple: data is king, take better care of your data.
You can build the most sophisticated model imaginable, but if your data is incomplete, corrupted, inconsistent, or full of missing elements, you will fail. So the first thing any organization should ask is: what key outcomes do we actually want to improve? Once you know that, focus on making sure the critical data elements are collected systematically, reliably, and in a machine-readable format.
You don’t need every possible data point. In many cases, 80% of the value comes from 10% of the data elements. Identify those essential elements and ensure they are trustworthy. Once you have that foundation, you can achieve a great deal.
And you don’t need to begin with the most advanced models or massive foundation models. Start with simple equation-based tools that already make a meaningful difference. Build trust gradually, then introduce more sophisticated methods over time.
Nowadays, there is another approach in which organizations simply buy integrated AI solutions as complete packages. That can work too, but if you go on that route, you must first ensure the system performs as advertised in your own organization, data environment, and unique circumstances before fully committing to it.
In which areas do you believe predictive healthcare will have the biggest clinical impact?
I believe every healthcare organization should begin with imaging, because imaging decision support is currently the most advanced field in terms of evidence-based toolsets and clinician workflows. After that, I think organizations should focus on integrating predictive modeling into preventive medicine to identify the right patients for the right interventions at the right time - this is your best chance to truly transform care in terms of patient-relevant outcomes.
