A scientific team from the Universidad Carlos III de Madrid (UC3M) has refined the operation of Bindi, an emotion-detection system in wearables capable of identifying fear in real-time during abuse situations to trigger automatic aid protocols. Their latest breakthrough involves integrating deep learning algorithms—an advanced branch of Artificial Intelligence (AI)—enabling detection without the need for remote servers, thereby reducing energy consumption and the transmission of sensitive information.
The scientific study detailing these updates, published in the Journal of Biomedical and Health Informatics, was designed to operate on microcontrollers integrated into Body Area Networks (BANs). This strategy differs from other approaches that require sending information to cloud processors with high but costly computing capacity, the researchers explain.
“Our innovation stems from the fact that we do not input raw signals; instead, the system extracts 57 pre-selected features from physiological signals such as skin conductance, skin temperature, and blood pulse volume. It is precisely the combination of this feature-based representation with convolutional architecture that allows us to effectively capture the dynamics of physiological responses,” explains one of the study's authors, Laura Gutiérrez Martín, who defended her thesis on this topic at UC3M a few months ago, titled: Expert system for robust alarm detection during fear episodes in cyber-physical systems. “This allows us to drastically reduce computational requirements, making the model occupy less memory space than a photograph taken with a mobile phone,” she adds.
“What we are looking for is to detect fear before an assault occurs, in order to activate a support network that can intervene immediately,” explains another author of the study, Celia López Ongil, professor in the UC3M Department of Electronic Technology and Director of the Institute for Gender Studies (IEG).
The operation is simple and effective. When the system identifies a risk situation, it sends an automatic alert to a "guardian circle." If the individual does not confirm she is safe, the goal is for the system to contact the police directly. Furthermore, all recorded data is encrypted and stored on a secure server so it can be used as judicial evidence if necessary, the researcher explains.
The study achieved accuracy metrics of around 80%, representing a 26.4% improvement over previous versions. Beyond current results, the team continues to work on reducing consumption and enhancing the model, as the use of such devices can help individuals identify their emotional states and support them during subsequent psychological recovery.
“The advantage is that the system could also be extrapolated to other areas, such as the early detection of school bullying. However, technology alone will not solve gender-based violence or bullying. DeepBindi is a support tool that must be complemented with education and social measures”, concludes José Ángel Miranda Calero, also a member of the research team.
The DeepBindi project was developed within the framework of the multidisciplinary team UC3M4Safety and has received funding from the State Research Agency (AEI) and INCIBE. The team is now seeking to complete a large-scale pilot to validate the system in real environments and facilitate its implementation, while continuing research and technological development.
Bibliographic Reference: L. Gutiérrez-Martín, C. López-Ongil, J. A. Miranda-Calero (2026) "DeepBindi: An End-to-End Fear Detection System Optimized for Extreme-Edge Deployment," in IEEE Journal of Biomedical and Health Informatics, vol. 30, no. 1, pp. 688-699, Jan. 2026, doi: 10.1109/JBHI.2025.3587961. e-archivo UC3M: https://hdl.handle.net/10016/49576
Vídeo: https://youtu.be/fd5EkQIL2ko