Missing Data Imputation in Daily Wearable Data for Improved Classification Performance

In the realm of wearable technology, the continuous monitoring of health parameters through smartwatches provides a wealth of daily data for research and analysis. However, this data often encounters missing values, presenting a challenge for interpretation and utilization. Remarkably, there exists a notable gap in the literature concerning the imputation of missing daily data from smartwatches. To address this gap, our study systematically explores a diverse set of imputation methods with Fitbit wearable data, encompassing various scenarios and missing rates. Our primary objectives are: (i) measure the influence of missing values rate and distribution on the proposed imputation methods; (ii) assess the role of data imputation in enhancing the performance of machine learning algorithms. Our results underscore the pivotal role of missing data patterns in imputation method selection. Furthermore, we demonstrate that more advanced data imputation approaches positively contributes to the efficacy of classification algorithms, improving 4,4% and 0,4% in terms of F-measure for the proposed classification tasks. This study not only addresses the challenges associated with missing data in wearable daily monitoring but it also provides practical insights for the optimization of machine learning applications in health monitoring.