OBJECTIVE: Non-restorative sleep is prevalent among individuals with depression and is strongly associated with the severity of the condition. Therefore, identifying non-restorative sleep can aid in the early screening of depression. Investigating non-restorative sleep in depression necessitates long-term monitoring under naturalistic conditions. METHODS: In this study, we recruited 149 participants and collected electrocardiogram and tri-axial acceleration from them, resulting in a total of 761 nights of data. The period from midnight to 6:30 AM was segmented into 78 five-minute intervals, from which 40 physiological features were extracted for each interval. To deal with variations in sleep patterns across participants and dates, we reordered the sleep data based on levels of parasympathetic nervous system (PNS) activation to explore the underlying neural mechanisms of non-restorative sleep in individuals with depressive symptoms. RESULTS: We developed a model that integrated convolutional neural networks with an attention mechanism to identify non-restorative sleep in individuals with depressive symptoms. The model demonstrated impressive performance on an independent test set, achieving an accuracy of 81.25% and an F1 score of 77.85%. Additionally, Bayes' theorem was used to compute the posterior probability indicating non-restorative sleep in this population, assessing abnormal PNS activation. CONCLUSION: Finally, we designed a system capable of automatically evaluating nighttime sleep status and quantifying changes in non-restorative sleep associated with depression. SIGNIFICANCE: This system offers a novel tool and method for the early identification of individuals at risk of depression.