Abstract: Wearable devices are a useful and widely used source of continuous and temporal dependant data. In contrast to the traditional clinical environment, these devices allow time series data collection in an individual’s daily living environ- ment. However, missing data can occur while using them. Many techniques have been applied to solve these data gaps; nonetheless, missing time series data poses extra challenges, such as maintaining the temporal dependency. In this article, we addressed the forecast of sleep trackers data (sleeping heart rate (HR) and time asleep) for 2 main reasons: (1) to design models capable of accurately forecasting missing data from those devices, and (2) to apply those models to empower sleep interventions that may increase its quality, by forecasting future sleep events. We collected wearables data over 290 days (per individual) from 12 participants using a smartwatch and made this dataset publicly available. We then explored several hyperparameters of 2 Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). We further elaborated and compared the performance of 3 approaches to training those RNNs. Although similar performance, slightly more accurate results were obtained after training a GRU network on an entire population’s dataset, which was able to forecast the average, minimum, and maximum sleeping HR with a root-mean-squared error (RMSE) of 4.4 (± 1.4), 4.9 (± 2.6), and 12.1 (± 4.0) beats per minute, respectively. However, the total time asleep was impossible to forecast with low error.

Abstract: Wearable devices are a useful and widely used source of continuous and temporal dependant data. In contrast to the traditional clinical environment, these devices allow time series data collection in an individual’s daily living environ- ment. However, missing data can occur while using them. Many techniques have been applied to solve these data gaps; nonetheless, missing time series data poses extra challenges, such as maintaining the temporal dependency. In this article, we addressed the forecast of sleep trackers data (sleeping heart rate (HR) and time asleep) for 2 main reasons: (1) to design models capable of accurately forecasting missing data from those devices, and (2) to apply those models to empower sleep interventions that may increase its quality, by forecasting future sleep events. We collected wearables data over 290 days (per individual) from 12 participants using a smartwatch and made this dataset publicly available. We then explored several hyperparameters of 2 Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). We further elaborated and compared the performance of 3 approaches to training those RNNs. Although similar performance, slightly more accurate results were obtained after training a GRU network on an entire population’s dataset, which was able to forecast the average, minimum, and maximum sleeping HR with a root-mean-squared error (RMSE) of 4.4 (± 1.4), 4.9 (± 2.6), and 12.1 (± 4.0) beats per minute, respectively. However, the total time asleep was impossible to forecast with low error.

Abstract: Wearable devices are a useful and widely used source of continuous and temporal dependant data. In contrast to the traditional clinical environment, these devices allow time series data collection in an individual’s daily living environ- ment. However, missing data can occur while using them. Many techniques have been applied to solve these data gaps; nonetheless, missing time series data poses extra challenges, such as maintaining the temporal dependency. In this article, we addressed the forecast of sleep trackers data (sleeping heart rate (HR) and time asleep) for 2 main reasons: (1) to design models capable of accurately forecasting missing data from those devices, and (2) to apply those models to empower sleep interventions that may increase its quality, by forecasting future sleep events. We collected wearables data over 290 days (per individual) from 12 participants using a smartwatch and made this dataset publicly available. We then explored several hyperparameters of 2 Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). We further elaborated and compared the performance of 3 approaches to training those RNNs. Although similar performance, slightly more accurate results were obtained after training a GRU network on an entire population’s dataset, which was able to forecast the average, minimum, and maximum sleeping HR with a root-mean-squared error (RMSE) of 4.4 (± 1.4), 4.9 (± 2.6), and 12.1 (± 4.0) beats per minute, respectively. However, the total time asleep was impossible to forecast with low error.