LANTERN Models Health-State Transitions for Long-Term Care Insurance.

Accurately predicting long-term care transition probabilities is critical for disability insurance pricing, reserving, and solvency. Traditional actuarial models often rely on restrictive assumptions like Markov or semi-Markov processes, which may not adequately capture the complexities of irregular longitudinal health data, nonlinear aging patterns, and diverse covariate histories. This paper presents LANTERN, a novel framework that develops a well-calibrated estimator for multi-state transition probabilities using irregular longitudinal health data. The model learns from individual health histories, accounts for time between observations, and conditions probabilities on demographic and socioeconomic attributes. It outputs a valid probability distribution across four health states: healthy, mild disability, severe disability, and death. Individual probabilities are then aggregated by age group and origin state to form transition matrices compatible with actuarial cohort projections. Evaluated against logistic regression, gradient-boosted trees, a recurrent neural network, and a last-state persistence benchmark using data from the Health and Retirement Study, LANTERN demonstrated improved severe disability discrimination, strong calibration, and the lowest transition matrix error, proving its utility for long-term care transition modeling.