Death certificate accuracy becomes crucial when public health decisions depend on mortality data, yet systematic undercounting can obscure the true burden of disease and mask health disparities that demand targeted interventions. Machine learning analysis of 2020-2021 US death certificates has identified approximately 155,536 COVID-19 deaths that were officially attributed to other causes, representing a 19% undercount of the pandemic's actual toll. The algorithm trained on patterns within death certificate data to distinguish COVID deaths misclassified as pneumonia, heart failure, or other conditions. This computational approach revealed that misclassification wasn't random but followed distinct demographic and geographic patterns. The unrecognized deaths occurred disproportionately among individuals with less than high school education and those identifying as Hispanic, American Indian, Alaska Native, Asian, or Black. Counties with lower household incomes and worse baseline health metrics also showed higher rates of COVID death misclassification. The geographic clustering was particularly pronounced in Southern states. This systematic bias in death certification represents more than statistical error—it reflects structural inequities in healthcare access, diagnostic capabilities, and death investigation resources. The findings suggest that communities already facing health disparities bore a heavier COVID burden than official statistics indicated. For mortality surveillance systems, this demonstrates how machine learning can reveal hidden patterns in cause-of-death assignments. The methodology could prove valuable for future pandemic preparedness, ensuring that public health responses account for undercounted deaths in vulnerable populations. However, the analysis relies on death certificate data quality and may still miss deaths in communities with the most limited healthcare access.