Machine learning score using multiparametric assessment for death prediction in cardiac amyloidosis

Abstract

Background

Cardiac amyloidosis (CA) is a severe disease with poor prognosis and increasing incidence. Available scoring systems for prognostic stratification in light chain (AL) and transthyretin (ATTR) amyloidosis are based on limited biological parameters. Allowing process of a greater number and complexity of variables, machine learning (ML) could improve prognostic assessment.

Objectives

To investigate the feasibility and accuracy of supervised ML algorithms using clinical, biological and imaging features to predict all-cause mortality in CA patients.

Methods

Data were collected from the French Referral Center for Cardiac Amyloidosis database (Hôpital Henri-Mondor, Créteil), including 1513 patients with wild type ATTR (n = 777), hereditary ATTR (n = 304) and AL (n = 432) CA between 2010 and 2023 (Fig. 1). Based on comprehensive clinical, biological and imaging features, we assessed accuracy of several supervised ML algorithms (Random Forest, Random Forest Ranger, XGBoost and LASSO) to predict all-cause mortality and compared with traditional logistic regression.

Results

Among 1513 CA included, 636 (42%) died during a median follow-up of 1.5 years (IQR: 0.5–3.1). ML score using XGBoost exhibited a higher area under the curve compared with logistic regression for prediction of all-cause mortality (AUC 0.76 vs 0.67, P < 0.001; Fig. 2). In ATTR cohort, ML score using Random Forest ranger evidenced better performance compared with logistic regression (AUC ML score 0.77 vs 0.72 with logistic regression, P = 0.008). However, in AL cohort, ML scores were not associated with an incremental prognostic value.

Conclusion

A ML-model including clinical, biological and imaging parameters showed the best accuracy to predict all-cause mortality in CA patients compared with any traditional methods.