Background: Fatigue is a subjective, complex and multi-faceted phenomenon, commonly experienced as tiredness. However, pathological fatigue is a major debilitating symptom associated with overwhelming feelings of physical and mental exhaustion. It is a well-recognized manifestation in chronic inflammatory rheumatic diseases, such as Sjögren’s Syndrome and Systemic Lupus Erythematosus and an important predictor of patient’s health-related quality of life (HRQoL). Patient reported outcome questions are the key instruments to assess fatigue. To date, there is no consensus about reliable quantitative assessments of fatigue.
Method: Observational data for a period of one month were collected from 296 participants in the United States. Data comprised continuous multimodal digital data from Fitbit, including heart rate, physical activity and sleep features, and app-based daily and weekly questions covering various HRQoL factors including pain, mood, general physical activity and fatigue. Descriptive statistics and hierarchical clustering of digital data were used to describe behavioural phenotypes. Gradient boosting classifiers were trained to classify participant-reported weekly fatigue and daily tiredness from multi-sensor and other participant-reported data, and extract a set of key predictive features.
Results: Cluster analysis of Fitbit parameters highlighted multiple digital phenotypes, including sleep-affected, fatigued and healthy phenotypes. Features from participant-reported data and Fitbit data both contributed as key predictive features of weekly physical and mental fatigue and daily tiredness. Participant answers to pain and depressed mood-related daily questions contributed the most as top features for predicting physical and mental fatigue, respectively. To classify daily tiredness, participant answers to questions on pain, mood and ability to perform daily activities contributed the most. Features related to daily resting heart rate and step counts and bouts were overall the most important Fitbit features for the classification models.
Conclusion: These results demonstrate that multimodal digital data can be used to quantitatively and more frequently augment pathological and non-pathological participant-reported fatigue.
Read the full publication here.