Machine learning and mobile sensing for naturalistic infant behavior (0-36 months): A comprehensive review and...

The recent rapid development of mobile and wearable sensing technologies and computational modeling has allowed for high-density and continuous measurement of infants' real-world behavior in natural settings. However, developmental science still struggles practically and methodologically to capture, label, integrate, and understand these data streams, especially in a way that is generalizable and inclusive. This review explores the state of the art in sensing and machine learning (ML) methods to study infants 0-36 months in natural environments. It delineates (i) sensing technologies, (ii) data collection and annotation approaches, (iii) ML approaches for behaviour recognition, prediction and modelling, and (iv) translational paths towards screening and early intervention, with a focus on feasibility, inclusivity, and ethics. A systematic review of literature published between 2015 and 2025 was performed in the following databases (Scopus, Web of Science, PubMed, IEEE Xplore, ACM), in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. Inclusion criteria involved studies of infants between 0 and 36 months, naturalistic or free-flowing contexts, and mobile/wearable sensing, computer vision, signal processing or ML. The review examined the sensor modalities (egocentric and exocentric camera and microphone, inertial measurement unit/actigraphy, physiological sensors), data properties (temporal resolution, density, duration), and settings (home, clinic, community). The synthesis places a primary emphasis on naturalistic motor behavior, which is the area most well-represented in the studies included. Selectively considered sensing and ML evidence for vocal and communicative behavior (including language exposure and caregiver-infant interaction), social-affective behavior, and sleep-wake regulation are presented where relevant. It also explored ML paradigms (supervised, semi- and weakly supervised, self-supervised, multimodal, sequential, and transformer-based), validation approaches, and reporting quality.