Predictive models for health-related quality of life built on two telemonitoring datasets

Congestive heart failure (CHF) is an incurable disease where a key objective of the treatment is to maintain the patient's quality of life (QoL) as much as possible. A model that predicts health-related QoL (HRQoL) based on physiological and ambient parameters can be used to monitor these parameters for the patient's benefit. Since it is difficult to predict how CHF progresses, in this study we tried to predict HRQoL for a particular patient as an individual, using two different datasets, collected while telemonitoring CHF patients. We used different types of imputation, classification models, number of classes and evaluation techniques for both datasets, but the main focus is on unifying the datasets, which allowed us to build cross-dataset models. The results showed that using general predictive models intended for previously unseen patients do not work well. Personalization significantly improves the prediction, both personalized models and personalized imputation, which is important due to many missing data in the datasets. However, this implies that applications using such predictive models would also need to collect some self-reported labels of HRQoL to be able to help patients effectively.