Researchers announced on Mar. 19 that a new artificial intelligence model has uncovered a significant number of unrecognized COVID-19 deaths in the United States during the pandemic. The study, published in Science Advances, used machine learning to estimate that more than 155,000 COVID-19 deaths were not officially recorded as such between March 2020 and December 2021.
The findings highlight concerns about the accuracy and equity of public health reporting systems. Accurate mortality data is considered essential for officials to allocate resources and develop effective policies during emergencies.
According to the study, the machine learning model was trained using death certificate data from inpatient hospital deaths, where COVID-19 testing and cause-of-death reporting were assumed to be highly accurate. Sixteen different models were tested, with Extreme Gradient Boosting (XGBoost) selected for its predictive accuracy. The model then analyzed over 3.8 million out-of-hospital death certificates for adults aged 25 and older.
The analysis found that official records undercounted COVID-19 deaths by about 19 percent compared to the model’s estimates. Discrepancies were most pronounced for deaths occurring at home, where predicted numbers were more than double those reported. The study also identified significant gaps in hospice care and emergency rooms.
The research revealed that underreporting was especially high among marginalized racial groups—including Hispanic, American Indian/Alaska Native, Black, and Asian populations—as well as individuals with less education and residents of Southern states such as Alabama, Oklahoma, and South Carolina. Counties with lower household incomes and poorer health metrics also had higher rates of unrecognized deaths.
The authors concluded that "the US death investigation system undercounted COVID-19 deaths in a 'systematically inequitable' way." They noted that while their approach relies on certain assumptions—such as applying patterns from hospital data to home deaths—it offers an alternative method for identifying hidden mortality beyond traditional excess-death models. The researchers suggested future studies could use similar techniques to investigate other hidden causes of death.
