Predictive modeling of damage and repair for disease and activity of daily living status in ELSA dataset using machine learning models

Abstract

A good predictive model is useful in health sciences for predicting onset of disease, as well as damage or repair of health deficits. One can predict one or more of these quantities depending on the nature of the collected data. In this thesis, we predict the future binary health states between successive waves of the English Longitudinal Study of Aging (ELSA) dataset. The predicted health states are 19 diseases and 25 activities of daily living states (ADLs) of individuals in the ELSA study. While we can directly predict those states with a high prediction quality, we cannot directly predict damage and repair probabilities or individual binary damage transition probabilities with the similar high prediction quality. However, we could predictively model damage and repair probabilities using the predicted health states. We applied model selection between deep neural networks (DNN), random forests, and logistic regression, then found that a simple one-hidden layer 128-node DNN was best able to predict future health states (AUC $\geq 0.91$) and average damage and repair probabilities ($R^2 \geq 0.92$). We applied feature selection for 134 full explanatory variables and found that 33 variables are sufficient to predict all disease and ADL states well. The prediction quality of individual damage transition probabilities are analyzed by the deciles of the probabilities and found to be well calibrated. We also studied the correlations between predicted health states which were stronger than the observed correlations. The hazard ratios (HRs) between high-risk deciles and the average were between 3 and 10 where high prevalence damage transitions typically had smaller HRs. We did not find a significant relation between model predictions versus individual ages.