AI from Bern Improves Cardiac Risk Assessment
Two international studies led by Inselspital show that artificial intelligence can assess the disease progression of certain heart patients more precisely than established risk models.
A patient receives the same diagnosis after a procedure as another. However, their health status can develop completely differently. This is precisely where a research team from Inselspital Bern and the University of Bern is focusing. In two recent studies, the researchers investigated whether artificial intelligence can reliably detect such differences.
The answer is positive for at least two specific heart conditions: The new models were able to assess risks more accurately than established methods. The results were published in The Lancet Digital Health and JAMA Cardiology, respectively. The studies do not yet prove that AI already makes better medical decisions today. Rather, they show the potential that computer-assisted prognoses could have in the future.
More Data Provides a Clearer Picture
The first study focused on people with severe aortic valve stenosis who were treated with a so-called TAVI procedure. This minimally invasive procedure involves inserting a new heart valve via a catheter.
Different types of information were combined for the AI model, including clinical data, laboratory values, electrocardiograms, ultrasound and CT scans, and procedure data. In total, data from 3991 patients from Switzerland and Japan were included in the study. The model was developed and validated with 2985 cases from Switzerland. Another 1006 patients from Japan served as an external test group.
The result: The AI models performed better than conventional surgical risk scores in predicting long-term mortality. The average accuracy over five years improved by approximately 10 to 15 percent, depending on the model and comparison. The improved performance was also evident in the external test group.
This external test is particularly important for medicine. A model that only works with the data it was developed with would be of little use. The fact that the Bernese models also functioned in an independent patient group in Japan suggests better transferability. However, this is not yet a guarantee for later use in other hospitals.
Focus on a Rare Heart Condition
The second study focused on Transthyretin Amyloid Cardiomyopathy, or ATTR-CM. This condition involves the deposition of protein fibers in the heart muscle. The disease can lead to heart failure and shorten life expectancy.
For the new model, data from 850 patients from six Swiss centers and the Medical University of Vienna were evaluated. The goal was to assess the risk of death or hospitalization due to heart failure.
Here, too, the AI model performed better than the established Mayo and NAC risk classifications. Depending on the patient group and comparison, the improvement in predictive performance ranged from approximately two to ten percent; in certain analyses, the difference in the three-year prognosis reached up to 19 percent.
Particularly interesting is that the researchers not only wanted to know how well the AI predicts. They also investigated which factors were decisive for each prognosis. This is important because a medical AI should not simply output a number without doctors being able to understand how it was derived.
What Does This Mean for Patients?
The potential benefit lies primarily in more individualized follow-up care. Today, patients are often classified using risk scores that work with a limited number of factors. AI can consider significantly more information simultaneously and identify complex relationships.
In the future, this could help to monitor particularly vulnerable individuals more closely or adjust therapies earlier. Conversely, for people with a lower risk, unnecessary check-ups might be avoided.
However, this does not mean that AI will decide who receives treatment and who does not in the future. The researchers explicitly emphasize that the technology is intended to support, not replace, medical judgment.
AI Not Yet Part of Daily Hospital Practice
Perhaps the most important sentence of both studies is therefore at the end: The models must first prove themselves in clinical practice.
Both systems have already been externally tested and are available as online applications for scientific purposes. However, before they can become a regular tool for patient care, further studies with other patient groups and under real-world conditions are needed.
The question of reliability also remains central. Medical data is complex, and patients vary considerably. A model can work very well for a specific group and less well for another. Therefore, careful external and prospective validation is crucial.
Bern Researches Tomorrow's Medicine
Nevertheless, the development is remarkable for Switzerland. The two works demonstrate how medical research, data science, and artificial intelligence can be integrated. Inselspital and the University of Bern are among the Swiss centers that seek to understand AI not as a replacement for doctors, but as an additional tool.
However, the decisive step is still ahead: A good prognosis in a study must become a reliably functioning application in daily practice.
The new models can already more accurately predict the risks certain heart patients face. Whether this will actually lead to better treatment must first be demonstrated in clinical practice.
