IMPROVING LIFE EXPECTANCY PREDICTIONS: A COMPARATIVE ANALYSIS OF MLP AND LSTM VERSUS LCR AND LCX MODELS

Abstract Accurate life expectancy prediction is a critical component of demographic planning and public health strategy. This study builds upon the author's previously developed machine learning models—Life Calculator Random Forest (LCR) and Life Calculator XGBoost (LCX)—by introducing two new deep learning-based models: the Life Calculator Multilayer Perceptron (LCM) and the Life Calculator LSTM (LCL). All four models were evaluated using a dataset derived from Ukrainian respondents, supported by World Health Organization and national statistical data. Performance was assessed using regression metrics such as RMSE, MAE, R², MAPE, MSE, and the Concordance Index. Results indicate that both deep learning models (LCM and LCL) outperform the existing models (LCR and LCX) in reducing prediction error across most metrics. While all models produced negative R² values—reflecting challenges relative to mean-based baselines—fewer negative scores from the deep learning models suggest improved relative performance. These findings underscore the potential of neural networks to model the complex, nonlinear dynamics of life expectancy and point to future opportunities for enhancement through deeper architectures and enriched features.