Precision screening for cardiac amyloidosis using explainable AI: a threshold-optimised random forest pipeline with SHAP-based interpretability
J Tuennermann, M Sprang, M Molitor, P Wenzel, S Kramer, M Andrade, P Lurz, M BrandtAbstract
Background
Cardiac amyloidosis (CA) is an increasingly recognised but frequently underdiagnosed cause of heart failure, largely due to its heterogeneous clinical presentation and the complexity of current multimodal diagnostic algorithms. Because delayed recognition is associated with rapid disease progression and poor prognosis, early and accessible screening is vital. While machine learning (ML) offers a potential solution for early diagnostics, existing models often function as "black boxes" or rely on resource-intensive inputs, limiting real-world clinical adoption. This study aimed to develop and internally validate a transparent, high-sensitivity ML pipeline to identify patients at risk of CA using only routinely available, non-invasive clinical parameters.
Methods
We conducted a retrospective derivation and validation study using electronic health records from 952 adult patients with unexplained heart failure who were referred for myocardial biopsy at a tertiary centre (100 positive CA cases). A random forest pipeline was developed using 15 strictly selected features encompassing demographics, standard biomarkers, routine laboratory panels, and basic echocardiographic parameters. Class imbalance was addressed using an optimised algorithmic class weighting strategy, which proved superior to synthetic oversampling in a 10-fold cross-validated ablation study. Model explainability was audited using Shapley Additive exPlanations (SHAP).
Results
In the final held-out test evaluation, the optimised pipeline achieved a robust area under the receiver operating characteristic curve (AUROC) of 0.967 and an area under the precision-recall curve (AUPRC) of 0.767. To prioritise clinical safety and minimise missed diagnoses, an optimised decision threshold (0.34) was applied (Figure A), yielding a sensitivity of 90.0%, specificity of 94.1%, and a negative predictive value (NPV) of 98.8% (Figure B). Global SHAP analysis confirmed the model’s clinical validity (Figure C); predictions were driven by established pathophysiological hallmarks, including advanced age, structural pseudohypertrophy (interventricular septal thickness (IVSd) and left ventricular posterior wall thickness (LVPWd)), a proxy for wall stress (BNP), and relative preservation of left ventricular ejection fraction (LVEF). Furthermore, patient-level risk audits successfully decomposed these clinical features to explain individual predictions (Figure D).
Conclusion
Our threshold-optimised machine-learning approach provides a highly accurate, interpretable screening tool for cardiac amyloidosis. By relying exclusively on 15 routine, non-invasive parameters and achieving an NPV of 98.8%, this explainable AI tool allows clinicians to safely rule out CA, reducing the need for unnecessary invasive workups and accelerating the time-to-diagnosis for true positive cases.1. Threshold optimization and performance2. SHAP model explainability