News|Articles|September 7, 2026

Machine Learning Model Beats Standard Staging in ATTR-CM

Listen

0:00 / 0:00

Key Takeaways

A random survival forest trained on 850 ATTR-CM patients achieved c-indices 0.72–0.77 and 3-year AUC 0.73–0.80 across centers using internal-external cross-validation.
Compared with Mayo Clinic and NAC staging, discrimination improved by 2%–10% (c-index) and 3%–19% (3-year AUC), addressing risk compression into lowest stages.
Calibration was largely acceptable, but 3-year overestimation occurred in one cohort with shorter follow-up, underscoring time-horizon sensitivity and cohort-specific event accrual.
SurvSHAP explainability ranked NT-proBNP, age, eGFR, NYHA class, and amyloid-specific therapy as dominant predictors, supporting clinical plausibility and interpretability.
Deployment via a free web application and open code enables external validation, though broader geographic and prospective testing is needed before widespread adoption.

A random survival forest model trained on 850 patients outperformed the Mayo Clinic and National Amyloidosis Centre staging systems for ATTR-CM.

A machine learning (ML) model built to predict outcomes in transthyretin amyloid cardiomyopathy (ATTR-CM) demonstrated better prognostic discrimination than the 2 staging systems most widely used in clinical practice today, according to a study published in JAMA Cardiology.¹

“In this multicenter cohort study, we developed and validated an open-source time-to-event ML model for predicting the composite end point of all-cause mortality and hospitalization for HF [heart failure] among patients with ATTR-CM, using multimodal data,” wrote the researchers.

ATTR-CM is an underdiagnosed but increasingly recognized cause of heart failure, particularly in older adults.² Its clinical course has shifted substantially since the arrival of disease-modifying therapies such as tafamidis, but clinicians have lacked a contemporary tool to stratify prognosis at diagnosis, since the Mayo Clinic and National Amyloidosis Centre (NAC) staging systems were both developed before these therapies became standard of care.

Study Design and Population

Researchers used data from the Swiss Cardiac Amyloidosis Registry and the Cardiac Amyloidosis Registry of the Medical University of Vienna to train a random survival forest (RSF) model on 850 patients with confirmed ATTR-CM (median age, 79 years; 88.2% male).¹ The cohort was split into 3 groups—Bern-Swiss (n = 352), Other-Swiss (n = 260), and Vienna (n = 238)—and evaluated using internal-external cross-validation, in which each center was held out in turn to validate the model built on the remaining centers. The primary outcome was a composite of all-cause mortality and hospitalization for heart failure.

Improved Discrimination Across Cohorts

The RSF model achieved Harrell concordance indices of 0.74, 0.77, and 0.72 in the Bern-Swiss, Other-Swiss, and Vienna cohorts, respectively, with 3-year area under the curve (AUC) values of 0.73, 0.80, and 0.75. Both metrics exceeded those of the Mayo Clinic and NAC systems, with concordance index improvements of 2% to 10% and 3-year AUC gains of 3% to 19% across cohorts. Calibration was generally acceptable across cohorts and time points, though it deteriorated at 3 years in the Other-Swiss cohort, where the model tended to overestimate risk—a finding the authors attributed partly to that cohort's shorter median follow-up. Model performance held up in the subgroup of patients receiving amyloid-specific therapy.

The investigators also noted that both the Mayo Clinic and NAC systems tend to classify most patients into the lowest-risk category, limiting their ability to separate patients by risk. The ML model produced clearer stratification, with 3-year event-free survival in its low-risk group ranging from 85% to 91% across cohorts, compared with 68% to 90% for the lowest-risk category under the existing staging systems.

Explainability and Deployment

Explainability analysis using the SurvSHAP algorithm identified NT-proBNP level, age, estimated glomerular filtration rate, New York Heart Association class, and use of amyloid-specific therapy as the strongest drivers of predicted risk were all consistent with established clinical understanding of ATTR-CM progression.

The authors trained a final model on the full pooled cohort and made it available through a free web application, alongside publicly posted code, to support external validation and clinical use. They cautioned that the cohorts were all from Central European centers and that prospective, geographically diverse validation is still needed before the tool can be broadly adopted.

Managed Care Implications

The study's authors noted that the model's predictions could help identify patients who need closer follow-up, including more frequent clinical assessments, home monitoring of body weight, and timely adjustment of diuretic therapy. They also noted that an updated risk-stratification tool's clinical relevance lies not only in improved predictive performance but also in its ability to account for the effects of disease-modifying therapies—effects the Mayo Clinic and NAC systems, developed before these therapies were available, cannot capture.

“The model demonstrated good and stable discriminative performance and provided incremental prognostic value beyond the NAC and Mayo Clinic staging systems,” wrote the researchers. “These findings highlight the potential of ML-based risk prediction to improve individualized prognostication in ATTR-CM.”

References

Baj G, Ciocca N, Mohammadi Kazaj P, et al. Machine learning-driven risk prediction model in transthyretin amyloid cardiomyopathy. JAMA Cardiology. Published online August 28, 2026. doi:10.1001/jamacardio.2026.3496
Zeldin L, Brailovsky Y, Maurer MS. Transthyretin amyloid cardiomyopathy: a rapidly evolving landscape. Annual Review of Medicine. 2026;77:59-74. doi:10.1146/annurev-med-050124-030735