Prediction of Mortality Using On-Line, Self-Reported Health Data: Empirical Test of the Realage Score

Objective: We validate an online, personalized mortality risk measure called “RealAge” assigned to 30 million individuals over the past 10 years. Methods: 188,698 RealAge survey respondents were linked to California Department of Public Health death records using a one-way cryptographic hash of first name, last name, and date of birth. 1,046 were identified as deceased. We used Cox proportional hazards models and receiver operating characteristic (ROC) curves to estimate the relative scales and predictive accuracies of chronological age, the RealAge score, and the Framingham ATP-III score for hard coronary heart disease (HCHD) in this data. To address concerns about selection and to examine possible heterogeneity, we compared the results by time to death at registration, underlying cause of death, and relative health among users. Results: The RealAge score is accurately scaled (hazard ratios: age 1.076; RealAge-age 1.084) and more accurate than chronological age (age c-statistic: 0.748; RealAge c-statistic: 0.847) in predicting mortality from hard coronary heart disease following survey completion. The score is more accurate than the Framingham ATP-III score for hard coronary heart disease (c-statistic: 0.814), perhaps because self-reported cholesterol levels are relatively uninformative in the RealAge user sample. RealAge predicts deaths from malignant neoplasms, heart disease, and external causes. The score does not predict malignant neoplasm deaths when restricted to users with no smoking history, no prior cancer diagnosis, and no indicated health interest in cancer (p-value 0.820). Conclusion: The RealAge score is a valid measure of mortality risk in its user population.