Mental disorders are among the leading causes of disability worldwide, with severe consequences for individuals, their families, and society at large. Detecting mental disorders typically requires resource-intensive clinical interviews conducted by specialists. In addition, there is a global shortage of trained therapists. In the early stages of a mental disorder, when interventions are most effective, artificial intelligence could significantly improve diagnosis and treatment.
AI systems could support therapists by analyzing subtle signals in patients’ language, facial expressions, and choice of words. Training such systems, however, requires highly sensitive data from real therapy sessions. Speech and video data can reveal patient identities, and models trained on such data risk memorizing and unintentionally exposing personal information.
Researchers at the Ubiquitous Knowledge Processing (UKP) Lab at the Department of Computer Science at TU Darmstadt and at IIT Delhi have now published a Perspective article in Nature Computational Science that outlines a new path forward. They describe how AI systems for mental health can be designed in a way that preserves the confidentiality of patient information.
To achieve this, the authors propose a development pipeline for privacy-aware AI systems based on several approaches. These include the removal of personally identifiable information, anonymization of voice and facial data, the generation of synthetic data, and privacy-preserving training methods.
The first author of the study, Aishik Mandal, is part of the NLPsych group at the UKP Lab, a group of researchers working at the intersection of natural language processing (NLP) and mental health to develop data-driven solutions that support both those seeking and those providing help. Co-authors are Professor Tanmoy Chakraborty (IIT Delhi), who was a visiting researcher at the UKP Lab , and Professor Iryna Gurevych, head of the UKP Lab at TU Darmstadt.
More information:
Aishik Mandal et al, Towards privacy-aware mental health AI models, Nature Computational Science (2025). DOI: 10.1038/s43588-025-00875-w
The content is provided for information purposes only.