Multicategory CNN for Prediction Sleep Quality from Feature Physiological and Behavioral Daily

Abstract

Sleep quality is a vital indicator of overall human health and cognitive performance. Traditional methods for sleep assessment rely on subjective questionnaires or single-modality sensor data, which limits precision and adaptability in real-world monitoring. This study aims to develop a multiclass Convolutional Neural Network (CNN) model to predict Sleep_Quality levels— poor , fair , good , and excellent —by integrating physiological features (eg, heart rate, heart rate variability, skin temperature) and daily behavioral features (eg, step count, sedentary ratio, and smartphone usage patterns). The outputs were fused through fully connected layers followed by a Softmax classifier. Model performance was evaluated using subject-independent validation, with metrics including macro-F1, macro-AUROC, and Expected Calibration Error (ECE). The proposed multiclass CNN achieved an accuracy of 78.8%, a macro-F1 score of 0.76, and a macro-AUROC of 0.87, outperforming classical machine learning baselines such as SVM, Random Forest, and XGBoost . Multimodal fusion improved macro-F1 by 8% over the best unimodal model. Post-training temperature scaling reduced ECE from 0.094 to 0.064, indicating improved reliability of probabilistic outputs. Grad-CAM analysis revealed interpretable temporal patterns linking stable HRV and consistent bedtime routines to higher sleep quality categories. The results demonstrate that the proposed CNN multiclass model effectively captures complex physiological–behavioral interactions associated with sleep quality. This approach provides a foundation for personalized, data-driven sleep health monitoring systems and highlights the potential of deep learning in predictive sleep analytics. Future research will expand validation using cross-device and polysomnography datasets to enhance clinical generalizability.