Machine learning prediction of coded transthyretin amyloid cardiomyopathy among adults with heart failure: a systematic review and meta-analysis

Background Structured electronic health record and claims data could support scalable case-finding for transthyretin amyloid cardiomyopathy (ATTR-CM), but reported models usually predict diagnosis codes or proxy labels rather than clinically adjudicated disease. We synthesized the performance, transportability, and clinical implications of these models in adults with heart failure. Methods We searched PubMed, Embase, Scopus, CINAHL/EBSCO, Cochrane CENTRAL, and Google Scholar through 30 June 2026 for supervised machine-learning classifiers trained on retrospective structured health-record data. Sensitivity and specificity were pooled with a bivariate random-effects model; reported area under the receiver operating characteristic curve (AUROC) was pooled separately on the logit scale using Hanley-McNeil variances, DerSimonian-Laird random effects, and Hartung-Knapp-Sidik-Jonkman confidence intervals. Risk of bias was assessed with PROBAST + AI. Predictive values were projected across assumed prevalence levels. Results Seven studies contributed eight cohort evaluations (3,658 participants; 1,841 cases). All selected estimates were from random-forest models and all study targets represented wild-type ATTR-CM diagnosis codes or coding-based proxies. Pooled sensitivity was 79.8% (95% CI 65.7–89.0%) and specificity 73.3% (65.5–79.9%); the summary receiver operating characteristic area was 0.809. Pooled reported AUROC was 0.861 (0.746–0.929), with high heterogeneity (I²=93.9%). Excluding the UK and German transfer evaluations yielded sensitivity 84.2% (75.3–90.3%), specificity 73.4% (65.5–80.0%), and AUROC 0.892 (0.799–0.945). At 1% assumed prevalence, pooled performance implied a positive predictive value of 2.9% and negative predictive value of 99.7%. All eight evaluations were at high overall risk of bias. None reported empirical calibration or decision-curve analysis. Conclusions Random-forest models show moderate ability to reproduce coded or proxy-defined ATTR-CM labels, but the evidence does not establish accurate detection of clinically confirmed disease. Given the high risk of bias among included studies, the pooled performance estimates warrant cautious interpretation. Their potential role is cohort enrichment or targeted case-finding rather than stand-alone population screening. Prospective evaluation with a uniform adjudicated reference standard, fixed prediagnostic prediction times, calibration, and workflow-level utility assessment is required. Registration PROSPERO CRD420261436051.