OBJECTIVES: During the pandemic, the numbers of Medicare beneficiaries dying increased including those with and without documented COVID-19. We used administrative, insurance, and clinical data from beneficiaries within a Medicare accountable care organization (ACO) to predict all-cause mortality within the next year, then compared actual and predicted mortality (2020-2022).

METHODS: We used linked insurance claims and electronic health record data to examine all-cause mortality among beneficiaries aged 65+ years. Using data from 2017-2019 and an 80/20 split, we developed and tested an algorithm for predicting death using baseline clinical and demographic data, e.g., AUC=0.890 (testing dataset). We compared actual versus predicted mortality by decile of predicted death for each year. The mortality prediction algorithm included age, biological sex, original reason for eligibility, race/ethnicity, prior ACO enrollment, CMS-HCC indicators, claims-based frailty measures, geriatric serious illness measures, and utilization in the prior year.

RESULTS: There were 190,343 beneficiaries; mean age in first enrollment year was 74.3 years and 58.3% were female. The predicted mortality varied: range=0.001 to 0.91; lowest decile mean=0.003; and highest decile mean=0.23. Between 2017-19, observed and predicted mortality was comparable (e.g., 2017 difference=0.0004, 95%CI: -0.0024-0.0032); between 2020-22, observed mortality exceeded predicted in all years (e.g., 2020 difference=0.004, 95%CI: 0.0027 - 0.0059); the difference increased as the predicted risk decile increased. For example, observed mortality exceeded predicted for subjects in the decile of highest predicted mortality (e.g., 2020 difference=0.023, 95%CI: 0.018-0.034); pre-pandemic observed and predicted mortality were comparable.

CONCLUSIONS: The one-year probability of death varied considerably across beneficiaries. The increase in all-cause mortality during the pandemic was concentrated among those with the highest mortality risk before consideration of COVID-19; among those with low predicted mortality, there were few deaths even during the pandemic in this population of older beneficiaries. Risk stratification could help target future interventions during pandemic-type outbreaks or disasters.