Advancements and Limitations: A Systematic Review of Remote-Based Deep Learning Predictive Algorithms for Depression

Abstract This systematic literature review explores the emerging field of remote-based deep learning predictive algorithms for depression, focusing on addressing the limitations of traditional diagnostic methods and examining the current state of research in this novel area. A systematic search was conducted in Embase, Medline, Web of Science Core Collection, CINAHL, and PsycINFO in June 2023. To capture relevant studies, titles and abstracts of the papers were reviewed against predefined inclusion and exclusion criteria using four groups of keywords addressing prediction, depression, validity, and deep learning. Eligible studies were systematically reviewed based on the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist. The risk of bias was assessed using the Prediction Model Risk of Bias Assessment (PROBAST) Tool for methodological quality. The synthesis of data was conducted using the Synthesis Without Meta-Analysis (SWiM) framework. From 286 studies initially identified, 6 studies met all inclusion criteria, published between 2020 and 2023. Performance metrics revealed the potential of deep learning models, with accuracy rates reaching as high as 98.23%. Convolutional neural networks (CNNs) emerged as the predominant model, with applicability across diverse data sources such as speech recordings, body motion data, and facial images. However, issues related to risk of bias were prevalent, with most studies lacking essential reporting details and employing relatively small sample sizes. The review identified limitations in the practical application of these models, including limited demographic representation, absence of external validation, and a notable absence of models capable of anticipating the onset of depression. While the current models focus primarily on identifying existing depression of any duration, there is a need for advancements that enable the anticipation of future depressive episodes. To advance this field, we recommend standardized reporting practices, larger and more diverse datasets, external validation, and the development of predictive models that anticipate depression occurrences in advance. These enhancements will contribute to the credibility and real-world relevance of these models. While remote-based deep learning predictive algorithms hold promise in revolutionizing depression prediction, they require refinement and validation to fulfil their potential in clinical practice. This review underscores the need for further research and development in this area to address the identified limitations and contribute to improved mental health assessment and intervention.