Deep Learning and Noninvasive Sensors for Detecting Physiological Dysregulation: A Scoping Review

TL;DR

This paper reviews how non-invasive sensors and deep learning can detect early signs of health issues like pain or stress in patients over 13 years old.

Contribution

It provides a scoping review of recent studies combining non-invasive sensors and deep learning for physiological monitoring.

Findings

Deep learning algorithms using sensor data showed high accuracy in predicting clinical events like pain or hypotension.

Most studies came from China, Australia, and South Korea, with applications in ICU and emergency settings.

Retrospective and experimental study designs were most common in the reviewed literature.

Abstract

Early detection of pain, stress, or hemodynamic instability is key to preventing serious clinical events. In recent years, non-invasive sensors and deep learning algorithms have gained relevance as tools for accurate and continuous monitoring. To map and synthesize the scientific evidence on the use of non-invasive multimodal sensors combined with deep learning algorithms for the early detection of physiological dysregulation states, including pain, stress, and hemodynamic deterioration, in patients over 13 years of age in clinical settings. A scoping review was conducted following the JBI and PRISMA-ScR guidelines. We included studies published between 2019 and 2025 in English or Spanish, identified through three databases and a secondary search. Twenty-seven studies were analyzed after duplicate removal and screening. Deep learning algorithms applied to electroencephalograms, electrocardiograms, photoplethysmography, and facial image signals showed high accuracy in predicting clinical events such as pain or hypotension. China and Australia had the highest number of included studies (n = 3), followed by South Korea, the United States, and Greece (n = 2 each). Retrospective and experimental designs predominated, with applications in intensive care units, operating rooms, and emergency rooms. These technologies represent an emerging strategy with high potential to improve early detection in clinical practice. However, further validation in real-world environments, optimization of implementation methods, and evaluation of their clinical impact are still needed.

Figures