Abstract
The importance of sleep is paramount for maintaining physical, emotional and mental wellbeing. Though the relationship between sleep and physical activity is known to be important, it is not yet fully understood. The explosion in popularity of actigraphy and wearable devices provides a unique opportunity to understand this relationship. Leveraging this information source requires new tools to be developed to facilitate data-driven research for sleep and activity patient-recommendations. In this paper we explore the use of deep learning to build sleep quality prediction models based on actigraphy data. We first use deep learning as a pure model building device by performing human activity recognition (HAR) on raw sensor data, and using deep learning to build sleep prediction models. We compare the deep learning models with those built using classical approaches, i.e. logistic regression, support vector machines, random forest and adaboost. Secondly, we employ the advantage of deep learning with its ability to handle high dimensional datasets. We explore several deep learning models on the raw wearable sensor output without performing HAR or any other feature extraction. Our results show that using a convolutional neural network on the raw wearables output improves the predictive value of sleep quality from physical activity, by an additional 8% compared to state-of-the-art non-deep learning approaches, which itself shows a 15% improvement over current practice. Moreover, utilizing deep learning on raw data eliminates the need for data pre-processing and simplifies the overall workflow to analyze actigraphy data for sleep and physical activity research.
Introduction
Quantifying the relationship between daytime physical activity and nocturnal sleep quality is a significant challenge in clinical research. While Polysomnography (PSG) remains the gold standard for sleep assessment, its intrusive nature and high cost limit its use in large-scale, naturalistic studies. Actigraphy, which utilizes wearable motion sensors, offers a portable alternative but traditionally relies on "lossy" data transformations. Conventional methods often involve manual feature engineering or automated Human Activity Recognition (HAR) to categorize movement into discrete intensity buckets (e.g., sedentary, light, vigorous). This process discards the high-dimensional nuances of raw accelerometer data. This study investigates whether deep learning can bypass these bottlenecks by identifying predictive patterns directly within raw sensor output.
Methodology
The researchers compared two primary modeling strategies using data from the QUEST (Qatar’s Ultimate Education for Sleep in Teenagers) study, which monitored 92 adolescents for one week using clinical-grade ActiGraph GT3X+ wearables.
- Feature-Engineered Approach: Raw data was pre-processed into four activity intensity levels using the RAHAR tool. These features were then fed into classical models (Logistic Regression, SVM, Random Forest, AdaBoost) and a standard Multi-Layer Perceptron (MLP).
- End-to-End Deep Learning Approach: Raw triaxial accelerometer data, aggregated into one-minute epochs, was fed directly into various deep architectures, including MLPs, Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs/LSTMs).
The CNN architecture (shown above) was specifically designed to capture local temporal patterns, such as bursts of high-intensity exercise, which may serve as critical indicators of sleep quality. In contrast, recurrent models were employed to explore long-term sequential dependencies across the entire day.
Results and Discussion
The experimental results demonstrated a clear advantage for deep learning on raw data. The CNN model emerged as the top performer, achieving an AUC of 0.9456, an accuracy of 0.9286, and an F1 score of 0.9444. This represents an 8% improvement over state-of-the-art non-deep learning methods and a 46% improvement over basic Logistic Regression on raw data.
A key insight from this performance is that the specific "shape" or nuance of physical acceleration contains vital information lost during traditional categorization. Interestingly, the CNN outperformed the LSTM, suggesting that local activity patterns (e.g., the intensity and duration of specific movements) are more predictive of sleep quality than the exact chronological sequence of every movement throughout a 24-hour period.
Conclusion
By utilizing deep learning—specifically CNNs—on raw wearable sensor data, researchers can achieve higher predictive accuracy while significantly simplifying the clinical workflow. This approach eliminates the need for complex HAR pre-processing, allowing for a direct pipeline from device to diagnosis. Future research should focus on scaling these models with larger datasets to mitigate overfitting and exploring more sophisticated architectures to improve specificity in identifying poor sleep patterns.
Summary updated