Deep Learning vs. Tree Ensembles for Life Expectancy Prediction: Evidence from Ukrainian Life Calculator

Life expectancy must be accurately predicted to allow effective planning of healthcare, resource allocation and public policy. This paper presents the comparison of four predictive models for the Ukrainian humanitarian project Life Calculator: the standard tree based ensemble machine learning models, Random Forest (LCR) and XGBoost (LCX), and deep learning models, Multilayer Perceptron (LCM) and Long Short-Term Memory (LCL). Data based on 260,000 Ukrainian participants were used to train all of the models using a single pipeline and evaluated based on common regression measures (RMSE, MAE, MSE, R2, MAPE) and the C-Index. Findings indicate that MLP and LSTM are always better than LCR and LCX in terms of error reduction in most measures, and MLP has the largest concordance score. In spite of this progress, negative values of R2 in all models indicate that there are still challenges regarding the noisy and incomplete health data. Deep learning models can capture more nonlinear, fine-grained associations among demographic and health variables; however, additional feature enrichment and more advanced models are required for clinically viable life expectancy predictions.