Combining Resampling and Machine Learning to Improve Sleep-Wake Detection of Fitbit Wristbands

Consumer sleep tracking wristbands such as Fitbit have been increasingly used in scientific studies to measure sleep outcomes. Nevertheless, many validation studies indicate that Fitbit wristbands often misclassify wake as sleep (i.e. low specificity). This study aims to develop classification models that leverage Fitbit data to generate accurate sleep/wake classifications. The problem of interest was formulated into an imbalanced binomial classification problem, as a normal night of human sleep consists dominantly more sleep time than wake time. We investigated the effectiveness of four resampling strategy (i.e. random up sampling, random down sampling, ROSE, and SMOTE) in combination with two machine learning algorithms (i.e. decision tree and random forest). Our results showed that the specificity of Fitbit was improved by 75% in the best case. The classification models can be used to post-process Fitbit sleep data for better accuracy.