Deep Multitask Learning for Sleep-Wake Classification with Signal Quality Index

Photoplethysmography (PPG) is well known for monitoring blood volume changes and has been adapted to detect sleep-wake states using deep learning approaches. However, in the field of sleep classification, many studies focus on single-task learning, where the output classifies wake or sleep without considering signal quality, which could enhance sleep-wake classification. Therefore, this study proposes a multi task learning approach to predict both the signal quality index (SQI) level and sleep-wake states per window using a CNN-BiLSTM model, comparing it with state-of-the-art sleep-wake models. The performance of the proposed model surpasses baseline models in cases of acceptable SQI accuracy: 75.65 ± 10.91 %, F1-score: 77.85 ± 15.96 %, sensitivity: 72.94 ± 20.09 %, specificity: 77.46 ± 13.88 %) but not in cases of excellent SQI. This underscores the resilience of the proposed model in sleep-wake classification, suggesting its potential for application with wearable sensors, where data quality is more sensitive to noise, for home monitoring of sleep-related conditions in the future.