Using Neural Networks to Predict Micro-level Elderly Mortality and Analysing its Risk Factors Using Causal Discovery
Predicting mortality risks at individual level is a challenging and yet essential task for both private and public decision-making. Accurate longevity forecasts and awareness of causal factors underlying mortality guide the individual life planning and shape the public policy by allowing for more targeted social interventions. Past research in individual-level lifetime estimation is either restricted to clinical settings or has not tested the potential of neural networks for predicting mortality. We implement five traditional and novel neural networks-based survival techniques using Health and Retirement Study (HRS), a nationally representative panel data of people aged 50 and older. The risk factors studied in this work are comprised of nearly 190 predictors spanning biological, behavioral, demographic and social sciences. The empirical results of survival models reported discriminative performance considered ”excellent” by statistical thresholds (the highest mean AUC of 85.58%). Furthermore, we show that modeling survival data with time-varying covariates allows the model to adjust the mortality forecasts according to the progression of variables through time. Finally, we discover that a combination of health, habits, and socioeconomic indicators contribute causally to mortality at older ages. The findings from this research highlight the potential of neural networks for predicting micro-level mortality, as well as guide future studies on what kind of causal factors should be more thoroughly investigated to reduce the perils of mortality.
E. Zhalieva, "Using Neural Networks to Predict Micro-level Elderly Mortality and Analysing its Risk Factors Using Causal Discovery", M.S. Thesis, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2022