Machine learning applied to digital phenotyping: A systematic literature review and taxonomy

Abstract

Health conditions, encompassing both physical and mental aspects, hold an influence that extends beyond the individual. These conditions affect personal well-being, relationships, and financial stability. Innovative strategies in healthcare, such as digital phenotyping, are strategic to mitigate these impacts. By merging diverse data sources, digital phenotyping seeks a comprehensive understanding of health, well-being, and behavioral conditions. Machine learning can enhance the analysis of these data, improving the comprehension of health and well-being. Therefore, this paper presents a systematic literature review on machine learning and digital phenotyping, examining the research field by filtering 2,860 articles from eleven databases published up to November 2023. The analysis focused on 124 articles to answer six research questions addressing machine learning techniques, data, devices, ontologies, and research challenges. This work presents a taxonomy for mapping explored areas in digital phenotyping and another for organizing machine learning techniques used in digital phenotyping research. The review found increased publications in 2023, indicating a growing interest in the field. The main challenges arise from the studies’ small participant samples and imbalanced datasets, limiting the generalizability of the results to broader populations and the choice of ML methods. Furthermore, the reliance on self-reported data can introduce potential inaccuracies due to recall and reporting biases. Beyond self-reports, authors explored different data types, including physiological, clinical, contextual, smartphone-based, and multimedia. Despite using video recordings in controlled experiments, studies have yet to investigate this method within intelligent environments. Researchers also analyzed neurophysiological phenotypes, suggesting the potential for interventions based on these characteristics.