AI and Machine Learning in Medicine

1. Given your extensive experience in hematology, how do you currently integrate AI and machine learning into your daily clinical practice at Kalnes Hospital and Rikshospitalet? What specific applications have proven most beneficial?

In my work with benign hematology, particularly at the thrombosis and hemostasis clinic at Kalnes, we’re actively exploring how AI and machine learning can enhance our ability to predict the risk of venous thromboembolism (VTE). One of our ongoing studies is focused on improving the predictive accuracy of traditional tools like the Wells score by integrating AI. We’re using machine learning to analyze data from our thrombosis registry, known as TROLL, which contains detailed information on over 5,000 patients diagnosed with VTE since 2005. This data is still being collected and offers a robust foundation for our research.

At Rikshospitalet, we’re leveraging AI and machine learning to optimize post-bone marrow transplantation care. Specifically, we’re using these tools to predict which patients can be safely managed in a home care setting and which require inpatient care. Additionally, we’re conducting a new study using machine learning to assess the benefits of early mobilization – like starting walking exercises soon after the transplant – and how group settings and scheduled appointments can impact recovery.

2. Could you elaborate on the ongoing studies you are involved in, particularly those focusing on machine learning for predicting thrombosis and personalized treatment? What preliminary results or insights have you observed?

In our thrombosis research, we're particularly focused on improving how we predict and diagnose VTE (venous thromboembolism) in primary care settings. The goal is to avoid unnecessary hospitalizations for patients who don't have VTE. Right now, there are long waits for ultrasound and CT-scans in emergency departments. If we can develop better tools to predict and, when necessary, rule out VTE, many patients could be treated directly by their general practitioners without needing hospitalization or unnecessary imaging tests.

We're also utilizing the TROLL registry with machine learning to better identify which patients need prolonged anticoagulation, determine the appropriate dosage (whether full or reduced), and assess the risk of bleeding. This approach aims to ensure that the right patients receive the right treatment at the right time, for the right duration.

Preliminary insights suggest that if we improve our initial assessments – using tools like blood tests and the Wells score – we can save a significant amount of time and resources. However, there's still a lot of room for improvement, and it's crucial that we share this knowledge more widely.

3. In your opinion, what are the most significant advantages of using AI and machine learning in diagnosing complex hematological conditions, such as thrombosis or rare disorders like PNH and CAD?

I believe AI and machine learning offer two key advantages when it comes to diagnosing rare diseases like PNH and CAD:

First would be faster diagnosis. Rare diseases often go undiagnosed for too long because they’re not immediately considered by specialists and doctors. AI can help by quickly identifying patterns and connections between symptoms across different organs, which speeds up the diagnostic process.

Second is personalized treatment. AI also plays a crucial role in deepening our understanding of these diseases, allowing us to deliver more personalized medicine. With the number of new and expensive treatments available, it’s often challenging to determine the right one for each patient. By using AI to analyze all relevant parameters, we can tailor treatments more effectively based on the complete clinical picture.

4. You mentioned that AI and machine learning cannot replace the nuanced judgment of experienced clinicians. How do you balance the use of these technologies with the need for human expertise in diagnosing and treating rare or complex cases?

I see AI and machine learning as powerful tools, but they work best when combined with human expertise. For example, in gastroscopy, you wouldn’t expect a trainee to achieve the same results with AI as an experienced gastroenterologist who has seen thousands of cases. However, when that experienced specialist uses AI, it can enhance the speed and accuracy of their work.

In hematology, we’re increasingly using machines to help read blood smears. While these machines are useful, they rely on archived images and might miss subtle abnormalities, especially at the edges of the smear where pathological cells often hide. This is why, as hematologists, we still need to take a second look if something seems off. AI can assist, but it’s the trained eye of an experienced clinician that ultimately makes the final call.

5. How do you address the limitations of AI and machine learning in recognizing patterns or diagnosing conditions that are not well-represented in existing datasets? What role does clinical knowledge play in overcoming these limitations?

AI and machine learning are incredibly powerful, but they’re only as good as the data they’re trained on. They excel at recognizing patterns in large datasets, but when it comes to rare diseases or smaller datasets, AI can struggle to make accurate diagnoses because there simply isn’t enough data to train it properly.

In medicine, much of our work is about recognizing patterns and drawing on our experience. We often ask ourselves, "What did we do the last time we saw a patient with these symptoms?" Relying solely on AI means we miss out on that critical experience, and we can’t teach the machine what to look for in those unique or rare cases.

The most important thing to remember is that patients aren’t machines – they don’t always follow the expected patterns, and they might have multiple conditions that complicate the diagnosis. For instance, consider a patient with a pulmonary embolism. Typically, the symptoms include shortness of breath and chest pain when breathing deeply. But what if this patient also has chronic obstructive pulmonary disease (COPD) and is always short of breath? It’s easy to mistake a worsening of their symptoms for just another COPD exacerbation. This is where clinical experience is invaluable – it helps us see beyond the obvious and consider all possibilities.

6. Can you discuss any specific challenges you’ve encountered when implementing AI-driven tools in your research or clinical practice? How have you addressed these challenges to ensure effective integration?

One challenge that really stands out involved a patient who came to our department with leukocytosis, but no signs of infection. He had been losing weight and feeling unwell for weeks, and my gut told me something was seriously wrong – possibly a hematological disease. We ran a blood smear, and the AI tool we use, called Cellavision, analyzed it and reported that everything looked normal. But when we manually examined the smear under a microscope, we found blasts – cells that are indicative of acute leukemia – all located in the outer part of the smear where the machine doesn’t typically scan.

If we had relied solely on the AI’s assessment, we might have missed this critical diagnosis. This experience reinforced for me that, even with advanced technology, if something doesn’t feel right, you need to double-check and often take a closer look yourself.

As a result, we’ve made it a practice to always double-check blood smears manually when the blood count is abnormal, ensuring that we don’t miss anything important.

7. What are the key factors that contribute to the successful use of AI and machine learning in clinical settings, particularly in terms of accuracy, reliability, and patient outcomes?

AI and machine learning are powerful tools that can help us make faster, more accurate diagnoses, and identify patterns in complex patients who don’t fit the usual categories. But the key to their success is how they’re used – especially by experienced clinicians.

For example, when an experienced gastroenterologist uses AI during a gastroscopy, they might spot small changes that could easily be missed by the naked eye. However, it’s the doctor’s expertise that interprets the significance of those changes.

Similarly, an ECG machine can help interpret wave patterns, but it’s important to remember that it’s just a suggestion. The final interpretation must be verified by a doctor and should align with the patient’s symptoms, as it might simply be a normal variant.

In my field of thromboembolism, AI can assist in analyzing CT scans, but it can’t assess the patient’s symptoms. That’s where doctors and other healthcare professionals come in – they’re essential for ensuring that the AI’s findings make sense in the context of the patient’s overall condition.

8. How do you see AI and machine learning evolving in the field of hematology over the next decade? What advancements or changes do you anticipate, and how might they impact clinical practice?

I believe AI and machine learning will play an increasingly significant role in both diagnosing and determining treatment options. As we continue to discover more diseases, subtypes, and emerging mutations, AI will be invaluable in sorting through all the data to help pinpoint precise diagnoses and predict the best treatment options.

In the future, AI might take over tasks like interpreting blood smears, bone marrow smears, and even histopathological examinations. However, I still think professionals will be essential for double-checking and verifying AI’s findings. We’ll still need to perform bone marrow and blood tests, but who knows – maybe robots will assist us in those tasks too.

The biggest impact will likely be that more samples and tests can be processed faster and more consistently. The technology is already there, but the cost is still a barrier for widespread use.

9. Given your role in educating healthcare professionals through your ultrasound online course company, how do you envision incorporating AI and machine learning into training programs for GPs, nurses, and paramedics?

This is definitely the future, and we’re already starting to see it in action. AI, along with social media and software, is making education more accessible, allowing more people to benefit from training. Instead of traditional one-on-one courses, we’re using AI and online platforms to teach from a distance. We can even be live on an ultrasound session with a student in another country – this is incredibly useful, for example, in war zones where soldiers can perform scans under our supervision while we’re back in a hospital in Norway.

AI can also assist by flagging when something looks abnormal, but it’s crucial that we, as professionals, identify and interpret those abnormalities ourselves rather than relying solely on the machine.

10. What ethical considerations do you believe are essential when utilizing AI and machine learning in patient care? How can clinicians and researchers address issues such as data privacy, algorithmic bias, and the lack of emotional understanding in these technologies?

Absolutely, these ethical considerations are crucial. For instance, using AI to sift through various journals and find patterns can be tough. It's a lot simpler when AI is focused on one patient at a time, but then we have to ensure patient IDs are anonymized to protect privacy.

Another challenge is that AI struggles with rare or unique cases. It’s great with large datasets and broad patterns, but it can easily miss the outliers – the one-in-a-billion cases that don’t fit the norm. This is why AI should support, rather than replace, human expertise.

And when it comes to the emotional side of care, AI just can’t replace the human touch. Many patients, especially the elderly, need someone who listens and provides comfort – sometimes just a simple hug or a kind word. They don’t need a machine to deliver tough news without any emotional support.

11. Could you provide examples of how AI and machine learning have improved patient management or decision-making in your field? Are there particular cases where these technologies made a significant difference?

Definitely! AI and machine learning have really changed the game for us. For instance, Cellevision has been a game-changer for screening blood smears. Instead of looking at every smear ourselves, AI helps us by flagging the ones that seem off or where the blood count raises a red flag. This way, we can spend more time on the cases that really need it and get through more smears, which helps us rule out diseases faster.

Another great example is how we use machine learning to figure out the risk of deep venous thrombosis (DVT). It lets us assess a patient’s risk right after their initial tests, so they don’t have to wait around for hours for an ultrasound. It speeds things up and makes the whole process a lot smoother for everyone involved.

12. In terms of research and collaboration, how do you work with interdisciplinary teams to enhance the development and application of AI and machine learning in hematology? What role do these collaborations play in advancing your studies?

In hematology, and specifically at Kalnes, we really rely on teamwork across different specialties to make the most of AI and machine learning. We partner with departments like radiology for ultrasound and CT scans, clinical medicine and pathologists for blood tests and smears, and cardiologists for echocardiography. We also team up with the technical university, where experts in machine learning help us with the latest tech and methods.

These collaborations are crucial for our success. As doctors, we’re great at clinical work, but when it comes to AI and tech advancements, we depend on specialists who really know their stuff. Working together with these experts allows us to stay on the cutting edge and get the best results from our research and applications.

13. What advice would you give to other clinicians and researchers who are considering integrating AI and machine learning into their practice or studies? What best practices and pitfalls should they be aware of?

If you're thinking about integrating AI and machine learning into your practice or research, here's my advice: definitely go for it, but keep your critical thinking hat on. AI can be a fantastic tool, but it’s important to remember that it's not infallible and shouldn’t be relied on blindly.

Always combine AI insights with your own clinical judgment. For example, we’re seeing more patients come in with whole genome sequencing results that suggest they have an increased risk of thrombosis. But when we dig deeper, we often find that these results also show genes that might lower the risk or even increase bleeding risk. So, it’s essential to consider the whole picture rather than just focusing on isolated bits of data.

In other words, treat AI as a valuable assistant rather than the final authority. It’s like the difference between a single vitamin and a whole apple – you need the full context to really understand what’s going on.

14. How do you envision the future relationship between clinicians and AI technologies in healthcare? What is the ideal scenario for their coexistence and collaboration to improve patient care?

I see AI as a permanent and powerful ally in healthcare, and I’m excited about the potential it holds. The ideal scenario is one where AI and clinicians work hand in hand to enhance patient care. AI can definitely make things like diagnosing, treating, and personalizing medicine faster and more precise, which is fantastic.

However, it’s crucial that we don’t lose sight of the human side of healthcare. Patients need compassion, empathy, and support, things that AI just can’t provide. So, while we should fully embrace the benefits of AI, we must remember that it should serve as a tool to support us, not replace the essential human touch. We need to stay in control and ensure that AI complements our care, helping us to be more effective but not taking over the personal aspects of patient interactions.

--Issu 04--