A machine learning-based artificial intelligence (AI) model can help clinicians more accurately predict long-term prognosis of patients with the rare heart disorder transthyretin amyloid cardiomyopathy
As reported in JAMA Cardiology, the researchers reported that the tool’s accuracy was three percent to 19% better over a three-year follow up period than conventional scores for predicting negative outcomes in these patients
Transthyretin amyloid cardiomyopathy is a progressive heart disease caused by a normally circulating protein, transthyretin (TTR), becoming unstable, misfolding and forming amyloid deposits in the heart. Those deposits make the heart muscle stiff, so it fills and pumps less effectively, leading to heart failure and rhythm or conduction problems
It is thought to impact around 120,000 people in the U.S., with 5,000-7,000 new cases identified each year, although figures could be higher as it is often missed or diagnosed late. The most common form is not inherited and involves the TTR protein becoming more unstable with age, but there is also an inherited form caused by different mutations in the TTR gene that varies in severity depending on the mutation
“In untreated cohorts, median survival ranges from approximately 20 to 66 months, depending on disease stage. The introduction of disease-modifying therapies has fundamentally altered the natural course of transthyretin amyloid cardiomyopathy, offering a demonstrable survival benefit and inaugurating a new therapeutic era,” explain lead author Christoph Gräni, MD, PhD, University of Bern, and colleagues
“Nevertheless, clinicians still lack a widely validated and contemporary risk model capable of accurately stratifying prognosis at diagnosis, an essential step in guiding treatment decisions and patient follow-up.”
To try and improve prognosis prediction for these patients, Gräni and colleagues analyzed medical records from 850 people with transthyretin amyloid cardiomyopathy treated at specialist centers in Switzerland and Austria. The typical participant was 79 years old, and most were men. The team collected information available around the time of diagnosis, including symptoms, medicines, blood-test results, kidney function and heart ultrasound measurements and used it to train a machine-learning program to look for patterns linked to death from any cause or hospital admission because of heart failure.
Rather than simply splitting one pooled dataset into training and test sets, they used internal-external cross-validation. This involved repeatedly training the model on two geographic cohorts and testing it on a third cohort, as a more demanding assessment of whether a model may transfer between centers
The new tool was better than the two standard staging systems, the Mayo Clinic and National Amyloidosis Centre scores, at separating people more likely to have one of these outcomes from those less likely to do so
The improvement was modest to meaningful depending on the hospital group tested. At three years, the researchers reported that the tool’s accuracy was three percent to 19% better than conventional scores, although these figures varied between centers
The researchers acknowledge the study had some limitations, for example, it used existing records from specialist centers in Europe, so the tool now needs testing in larger, more diverse groups. It finds patterns linked to risk but cannot prove what causes poor outcomes, and it was not able to capture repeat heart-failure admissions
However, they conclude: “These findings highlight the potential of machine learning-based risk prediction to improve individualized prognostication in transthyretin amyloid cardiomyopathy.”
News & FeaturesArtificial intelligenceCardiomyopathiesHeartPrognosisProteinsRare diseases
