July 30, 2026 | News | Publication

Individual-Level Mortality Predictions: Statistical Models vs. Machine Learning

MPIDR study examined whether machine learning (ML) could improve individual-level lifespan predictions

Researchers investigated whether traditional statistical and machine learning models could achieve a similarly high level of accuracy in predicting individual-level lifespan. Overall, the results show that machine learning does not significantly improve mortality predictions compared to conventional methods.

Researchers have investigated whether machine learning models can improve the accuracy of lifespan predictions at the individual level. © istockphoto.com / mohd izzuan

Can machine learning help make individual-level lifespan predictions even more precise? Scientists addressed this question as part of the first edition of the "Population and Social Data Science Summer Incubator Program" at the Max Planck Institute for Demographic Research (MPIDR). This research question emerged from discussions between the mentors and students. "We asked ourselves a variety of questions: How does the accuracy of lifespan predictions vary across different socioeconomic groups? Which variables are most important for accurately predicting mortality? Can individuals predict their own lifespan more accurately than statistical and machine learning survival models" explained Ugofilippo Basellini, an MPIDR researcher and program mentor.

This approach is one of the first attempts to rigorously compare individual-level lifespan predictions using a large number of statistical models and covariates. In particular, they compared the performance of classical statistical models with newer machine learning methods. The researchers used data from the U.S. Health and Retirement Study (HRS 2022) and tested twelve statistical and machine learning models, as well as more than 150 predictors.

Integrated Brier score and mean area under the curve (AUC) with 95% confidence intervals, by gender, race and ethnicity, and education. The dashed vertical lines indicate the optimal Brier score (.00) and the AUC (1.00). The mean AUC for Kaplan-Meier, .5 by construction, has not been reported.

The findings revealed that statistical and machine learning models exhibit comparable accuracy and achieve relatively high discriminatory power. "However, they fail to account for most lifespan heterogeneity at the individual level," said Maria Laura Miranda, a participating student in the program who afterwards joined the MPIDR for her PhD studies. “There are consistent disparities in mortality predictability and risk discrimination, with lower accuracy for men, non-Hispanic Blacks, and individuals with low educational attainment," said Miranda.

The researchers observed similar performance between traditional and machine learning models. There are also persistent disparities in predictions for certain population groups, regardless of the model used. "Predicting mortality at the individual level is difficult, even for the most advanced machine learning models currently available. Machine learning models do not significantly improve mortality prediction accuracy compared to classical methods,” the researchers concluded.

Original Publication

Badolato, L., A. Decter-Frain, N.J. Irons; M.L. Miranda, E. Walk, E. Zhalieva, M. Alexander, U. Basellini, E. Zagheni (2026): "The Limits of Predicting Individual-Level Longevity: Insights From the U.S. Health and Retirement Study". Demography. DOI: 10.1215/00703370-12464628

Keywords

Mortality prediction | Survival analysis | Machine learning | Inequality