Constructing a Transformer-based Model to Infer Daytime Productivity from Biometric Information During Sleep

Abstract: Good sleep reduces tiredness and improves daytime work. Sleep quality is primarily determined by sleep satisfaction, sleep stage, and sleep duration. However, the impact of sleep quality on daytime performance, such as sleepiness and productivity, remains unclear. Clarifying the relationship between sleep, daytime productivity, and sleepiness can help quantify the vigor restored by sleep and identify appropriate work hours, ultimately reducing workplace errors and accidents.Existing methods for estimating sleep quality have limitations. EEG-based methods struggle with estimating daytime performance metrics, while methods using wearable sensor devices lack accuracy in estimating sleep parameters. Based on the correlation between biometric indices and sleep quality, this study aims to estimate the next day's productivity directly from non-EEG biometric data, which can be easily measured and analyzed. This approach avoids the loss of information inherent in estimating sleep quality from biometric data and extracts information more closely related to next-day productivity. To achieve our goal, we built a deep learning model based on Multi-Head Attention.Electrocardiogram (ECG), respiration (RIP), and electrodermal activity (EDA) were selected as sleep-related biometrics based on a survey report on sleep quality. Analyzing large amounts of raw data as-is would result in poor computational efficiency, so we extracted feature values from each type of biometric information. Given the uncertainty about which features would contribute to the estimation, we extracted 23 types of features and refined them based on their contribution ratio.To evaluate reaction time, working memory, and sleepiness restored by sleep, we selected the PVT-B, 2-back task, and Stanford Sleepiness Scale (SSS) to assess daytime performance and subjective sleepiness. To capture the time-series characteristics of sleep biometric data, we used a Transformer-based model with only an encoder to infer daytime productivity. We also output Shapley values to indicate the contribution of each feature to the prediction result, thus improving the model's interpretability.Samples were prepared by mapping the biometric time-series data obtained from the experiment each night to the productivity index data and sleepiness on the following day. Data were collected over 51 days (15 subjects), excluding days with equipment malfunctions. Twenty-three biometric features were extracted from the experimental data in 4-minute time windows. The model was trained to estimate reaction time, the accuracy rate of the 2-back task, and sleepiness. Sleepiness inference was based on a three-category classification: better, same, and worse.The trained model performed well in predicting reaction time, 2-back task accuracy, and sleepiness, with an R2 coefficient greater than 0.5 and an F1 score of approximately 0.8. Shapley values revealed which features contributed significantly to each objective variable. The Transformer model effectively captures daytime productivity from biometric information during sleep and refines the number of features based on Shapley values.This study demonstrates that using non-EEG biometric data and a Transformer-based model can accurately estimate next-day productivity and sleepiness, offering a novel approach to understanding the impact of sleep quality on daytime performance.

Keywords: Biometric, Sleep Quality, Daytime Productivity, Deep Learning

DOI: 10.54941/ahfe1005541

Cite this paper:

​