Identification of at-risk populations for cardiac amyloidosis: a multimodal machine learning development and validation study

Abstract Background Early diagnosis is crucial to improve outcomes in patients with cardiac amyloidosis (CA) by initiation of life-saving treatments. Bone scintigraphy is highly sensitive and specific for CA diagnosis but the early identification of at-risk patients remains challenging. Purpose We aimed to develop and validate a multimodal prediction approach for identifying patients at risk of CA, guiding the referral process for confirmatory scintigraphy. Methods Consecutive all-comer referrals for bone scintigraphy at a large tertiary referral center between 2010 and 2023 were retrospectively enrolled. Patients referred before August 2020 were used for development and after August 2020 for independent validation. Machine learning modeling as well as univariate marker identification was performed using trait-wide association analysis and included 65 baseline (pre-scintigraphy) parameters extracted from routinely created electronic health records from echocardiography, cardiac magnetic resonance, laboratory, demographic parameters, and comorbidities. The primary endpoint was the prediction of CA-suggestive uptake on scintigraphy (Perugini grade≥2). All models were assessed for their prognostic value in terms of overall survival and heart failure-associated hospitalization (HFH). Results In total, 11,616 patients were included, of whom 279 (2.3%) exhibited high-grade cardiac uptake. The machine learning model demonstrated excellent performance in identifying patients with high-grade uptake in both the development (AUC 0.937 95% CI 0.925; 0.949) and independent validation cohort (AUC 0.912 0.882, 0.936). Results were consistent across subgroups with conditions commonly associated with left ventricular thickening (Figure 1). Of the 38/279 (14%) patients not identified by cardiologists in clinical routine, 24/38 (63%) were detected by the machine learning model. After a median follow-up time of 66 months (IQR 35-103) after scintigraphy, 590/11,342 (5.2%) patients had experienced HFH and 2,877/11,616 (24.8%) patients had died. The model predictions were significantly associated with mortality in the development (adjHR 3.05 95% CI 2.78; 3.35, p0.0001) and validation (adjHR 1.78 95% CI 1.38; 2.30, p0.0001) cohorts. Similarly, the predictions were significantly associated with time to first HFH for both the development (adjHR 5.79 95% CI 4.82; 6.96, p0.0001) and the validation cohorts (adjHR 5.16 95% CI 3.37; 7.90, p0.0001; Figure 2). Conclusions We propose a multimodal machine learning approach, Amylo-Detect, available online and ready for integration into electronic health records, that enables automated identification of patients at risk for CA. This model may facilitate timely referral for confirmatory CA testing, otherwise overlooked by clinicians.Prediction performance across subgroups Association of predictions with outcomes