When AI detects hidden disease patterns
Systemic sclerosis is rare, takes many different forms and, in some cases, is life-threatening. Zurich’s own Professor Michael Krauthammer, physician and researcher at the University of Zurich, explains how artificial intelligence can help to detect hidden risk patterns.
Prof. Krauthammer, in layman’s terms, what kind of disease is systemic sclerosis?
Michael Krauthammer: Systemic sclerosis is a rare disease that can attack the entire body as well as various individual organs – typically the skin, but also the lungs and heart. Thickening of the connective tissue is one of its hallmarks. This may manifest as increasing hardening or stiffening in the affected tissues. It’s a chronic disease and, in cases where the organs are severely affected, it can be life-threatening.
In clinical practice, there are categories that have developed over the course of time, including milder and more diffuse forms of the disease. What are the limitations of this classification system?
A key issue is identifying patients with rapidly progressive forms of the disease at an early stage. Some people remain stable for many years after diagnosis, while for others it runs a swift and severe course. Medical observation plays a major role on this front: how severely is the skin affected? Are there any indications of organ involvement? This is a system based on observation and experience that has developed over time, but it has its limitations: there’s no precision tool for making prognoses at an early stage.
Why is differentiating between the different subgroups at an early stage so important for treatment?
The question of the right time to treat is not an academic one; it’s of immediate clinical relevance. There are effective treatments – immunomodulatory treatments, including intensive procedures in certain cases – but these can have significant side effects. For this reason, it’s not desirable to administer these to all patients, but only in as targeted a manner as possible where an unfavourable prognosis seems likely. In this context, precision means intervening early enough, but without administering unnecessary treatment.
To help achieve more precise diagnostics, you used artificial intelligence (AI) to analyse the disease progression of a large number of patients.
I have a long-standing interest in how artificial intelligence might be able to complement medical classification systems. A research collaboration with the Department of Rheumatology at University Hospital Zurich enabled me, together with my research team, to analyse the disease progression of around 14,000 systemic sclerosis patients from a European register. The register covers disease progression periods of several years, in some cases up to a decade. Data sharing of this kind is crucial for rare diseases in particular: without international cooperation, there simply isn’t enough data to identify clear patterns.
The AI worked with a "semi-supervised learning framework”. What does that mean?
The AI searched for patterns independently, but it didn’t work without guardrails. We wanted the findings to be rooted in concepts that assist rheumatologists with their work. Hence the AI was provided with clinical definitions, such as threshold values that determine when an organ is deemed to be affected. Without these guardrails, a fully unsupervised search would have produced findings that are mathematically coherent but of little clinical use.
What did the AI ultimately find?
The findings of the analysis were informative: the AI identified multiple subgroups – including one of particular clinical significance. In these patients, the lungs were severely affected, while the skin remained almost unremarkable. If the treating physician focuses primarily on the skin, patients like these appear at first to have a less severe form of the disease. But it’s precisely this group that needs to be identified at an early stage – because, in spite of the unremarkable skin findings, the progression of the disease can have serious consequences. For me, this was the paper’s most significant insight.
What has the impact of your discovery been?
This finding has given rise to two avenues of research. One avenue is developing a clinical algorithm based on the AI analysis that can be used in practice to give early warnings of severe courses of disease progression. The other avenue is asking the biological question: what is it that differentiates this patient group? What disease processes are at work here? The AI doesn’t just deliver a ready-made answer, but rather lays the groundwork for further research.
Does your research into systemic sclerosis also provide a model for devising methods to identify other disease patterns?
Yes. The paper is not just a study into systemic sclerosis, but also a methodological contribution. What we’re working with is known as patient journey data: disease data along a time axis, with laboratory values, imaging, clinical findings and disease progression over the course of years. This methodology can be applied to other fields, such as other rheumatological diseases, or to cardiology and endocrinology. The significant thing is that the AI is not just taking one point in time into consideration, but rather the overall trend observed in a patient over time.
Your study has been published to an international readership, thanks to an article in Nature Reviews Rheumatology for example. What does that mean to you?
For me, this paper is a kind of milestone. We have shown that this approach works when applied to a complex disease. I’m delighted that this study has gone on to be received as a relevant development in an international rheumatological context. It goes to show that AI-assisted stratification in rheumatology is becoming an important topic.
_
Links