Classifying missense variants is often challenging due to limited pathogenicity evidence. Consequently, many remain categorized as variants of uncertain significance (VUS). VUS are at the core of healthcare disparities, as individuals from underrepresented race, ethnicity and ancestry (REA) populations are more likely to receive a VUS result. To improve VUS interpretation, we designed machine learning (ML), gene-specific algorithms utilizing public genomic, protein, and population data. To evaluate ML utility in arrhythmia genetic testing across REA groups. From 1/1-11/15/22, various ML algorithms were validated at Invitae, using existing models such as SpliceAI and developing our own by leveraging gnomAD, AlphaFold, and others. ML evidence was incorporated into Sherloc, a semi-quantitative ACMG/AMP-based variant interpretation framework. Only evidence that met a negative or positive predictive value >80% was incorporated. At least one evidence model was available for 42 arrhythmia genes. Variant classifications impacted by ML were evaluated, stratified by REA groups. Out of 2,715 US-based individuals that underwent arrhythmia panel testing, 1,303 (48%) had ML evidence applied to at least one variant and 796 (29%) were from an underrepresented group. Models contributed to classifying at least one benign/likely benign (B/LB) variant in 840 (31%) and at least one pathogenic/likely pathogenic (P/LP) variant in 43 (1.6%) individuals. A higher percentage of Asian (42%), Black (38%), and Hispanic (35%) individuals had an ML-dependent definitive classification (P/LP or B/LB) relative to White (30%) individuals (p <0.03). The average amount of benign or pathogenic ML evidence applied per reclassified variant was similar across populations, showing that the models contributed comparably across REA groups. By leveraging publicly available datasets to create gene-specific ML algorithms for variant interpretation, the resultant evidence impacted a significant number of individuals who had arrhythmia genetic testing. For those with at least one variant with ML evidence applied, 67% would have otherwise received a VUS result. Additionally, ML appears to provide more definitive variant classifications for underrepresented individuals. ML can assess factors in ways that are agnostic to population ancestry and, when appropriately implemented, can narrow the VUS gap rates across REA groups.