Beyond NT-proBNP and troponin: How machine learning redefines light-chain cardiac amyloidosis risk assessment

To develop and validate a machine learning-based prognostic model that provides enhanced risk stratification for AL cardiac amyloidosis patients beyond existing staging system. We conducted a retrospective cohort study of newly diagnosed AL cardiac amyloidosis patients (2006–2024), randomly allocating participants into training and test sets (8:2 ratio). Cardiac involvement required elevated cardiac biomarkers (NT-proBNP > 332 ng/L) or increased wall thickness (mean wall thickness > 12 mm). Using all-cause mortality as the primary endpoint, we compared five machine learning algorithms (support vector machine, CoxBoost, random survival forest, multi-layer perceptron, and k-neighbors classifier) and a traditional Cox model against the 2015 European-modified Mayo staging system. Among 132 enrolled patients (median age 60 years; 56.8% male), 83 deaths (62.8%) occurred during median 14.5-month follow-up. Feature selection identified six key predictors: cardiac response (36% vs. 15%, P < 0.001), complete hematological response (35% vs. 20%, P < 0.001), E/e’ ratio (15.0 [12.0, 20.0] vs. 19.0 [15.0, 25.0], P = 0.009), left ventricular global longitudinal strain (-14.1 [-11.2 – -16.1] vs. -10.7 [-8.8 – -13.6], P = 0.003), serum uric acid (341.0 [292.0–439.0] vs. 429.0 [345.5–518.0], P = 0.002), and weight loss (30.6% vs. 65.1%, P < 0.001). The CoxBoost model demonstrated superior discrimination (AUC 92%) and calibration (brier score 0.11). Conversely, the predictive value of the revised Mayo 2004 staging system was unsatisfactory, with an AUC of 74% and a brier score of 0.19. Machine learning incorporating multi-dimensional parameters (including myocardial strain and dynamic clinical variables) provides significantly more accurate prognostication than NT-proBNP/troponin-dependent staging, enabling a new era of personalized risk stratification in AL cardiac amyloidosis.