Well-being Forecasting using a Parametric Transfer-Learning method based on the Fisher Divergence and Hamiltonian Monte Carlo

INTRODUCTION: Traditional personalised modelling typically requires sufficient personal data for training. This is a challenge in healthcare contexts, e.g. when using smartphones to predict well-being.

OBJECTIVE: A method to produce incremental patient-specific models and forecasts even in the early stages of data collection when the data are sporadic and limited.

METHODS: We propose a parametric transfer-learning method based on the Fisher divergence, where information from other patients is injected as a prior term into a Hamiltonian Monte Carlo framework. We test our method on the NEVERMIND dataset of self-reported well-being scores.

RESULTS: Out of 54 scenarios representing varying training/forecasting lengths and competing methods, our method achieved overall best performance in 50 (92.6%) and demonstrated a significant median difference in45 (83.3%).

CONCLUSION: The method performs favourably overall, particularly when long-term forecasts are required given short-term data.