Prediction of Decline in Activities of Daily Living Through Deep Artificial Neural Networks and Domain Adaptation

Abstract

In order to improve information available at the clinical level and to better focus resources for preventive interventions, it is paramount to estimate the general exposure to risk of adverse health events, commonly referred as frailty. This study compares the performance of shallow and deep multilayer perceptrons (sMLP and dMLP), and of long short-term memories (LSTM), on the prediction of a subject decline in activities of daily living, with and without a previous autoencoder based domain adaptation from an external dataset. Samples originates from two large epidemiological datasets: the English Longitudinal Study of Ageing (ELSA) and The Irish Longitudinal Study on Ageing, with 107879 and 15710 eligible samples, respectively. Deep networks performed better than shallow ones, while dMLP and LSTM performance were similar. Domain adaptation improved predictive ability in all comparisons. On the bigger ELSA dataset, sMLP attains a Brier score of 0.32 without domain adaptation, and 0.15 with domain adaptation, while dMLP attains 0.20 and 0.11, respectively. Thus, experimental results support the use of deep architectures in the prediction of functional decline, and of domain adaptation when data from another similar domain is available. These results may help improving the state of the art in predictive models for clinical practice and population screening.