Summary: Diagnosing PTSD in children is challenging when communication or emotional awareness is limited. New research from the University of South Florida uses privacy-preserving artificial intelligence to detect PTSD-related facial expression patterns during interviews, offering an objective tool to support clinicians.
The system avoids using raw video. Instead, it analyzes de-identified facial signals such as head pose, eye gaze, and mouth movement to identify expression patterns linked to trauma. Findings show clinician-led interviews reveal clearer signals than parent-child conversations.
Key Facts:
- Privacy-preserving AI: Analysis relies on de-identified facial movement data rather than identifiable video.
- Objective markers: Distinct facial expression patterns were associated with PTSD in children.
- Clinician sessions most revealing: Children tended to display stronger, more informative emotional expressions during sessions with clinicians compared with parents.
Source: University of South Florida
Diagnosing post-traumatic stress disorder (PTSD) in children is often difficult because many young patients have trouble describing emotions, lack the language to explain symptoms, or suppress distress altogether.
Researchers at the University of South Florida combined expertise in childhood trauma and machine learning to develop a system that helps clinicians detect PTSD symptoms from non-identifying facial movement data. The interdisciplinary project was led by Alison Salloum, professor in the USF School of Social Work, and Shaun Canavan, associate professor in the Bellini College of Artificial Intelligence, Cybersecurity and Computing.
Published in Pattern Recognition Letters, the study is the first to combine context-aware PTSD classification with strict privacy protections. Rather than processing identifying video frames, the team extracts and analyzes facial landmarks, head pose, eye gaze and action unit (AU) intensities—features that capture movement and expression without revealing identity.
Traditional PTSD diagnosis depends on clinician interviews and self-report measures. These methods can be subjective and limited by a child’s developmental stage, vocabulary, avoidance strategies or emotional suppression. The USF approach aims to complement clinical judgment with objective, low-cost measures that can track symptoms and recovery over time.
Salloum observed that many children display intense facial expressions during trauma-focused interviews even when they speak little. That observation led the team to ask whether machine learning could quantify those expressions in a way that reliably indicates PTSD risk.
Canavan adapted tools from his lab to create a privacy-first processing pipeline. “We do not use raw video,” he said. “We remove identity information and keep only signals about facial movement, then analyze those signals in context—specifically whether the child is speaking with a clinician or a guardian.”
For this early study, researchers compiled data from 18 recorded sessions in which children discussed emotional experiences. The dataset contained extensive frame-by-frame measurements—more than 100 minutes of video per child and roughly 185,000 frames in each video—allowing models to detect subtle muscle movements and temporal patterns associated with emotional expressiveness.
Results showed identifiable differences in facial movement patterns between children with and without PTSD. The analysis also demonstrated that clinician-child discussions produced clearer signals than conversations with guardians, aligning with prior evidence that children may be more emotionally forthcoming with therapists.
The authors emphasize that the AI is intended to supplement, not replace, clinicians. In practice, the system could provide real-time feedback during sessions, help monitor treatment progress, and reduce the need for repeated distressing interviews. Because the data are de-identified, the approach addresses privacy concerns that commonly arise when using video and machine learning in sensitive clinical settings.
The research team plans to expand this work to larger, more diverse groups to evaluate possible biases related to gender, culture and age. Young children, including preschoolers who rely on caregiver reports for diagnosis, are a priority for further study because nonverbal cues are often the primary signal of distress in that age group.
Although still preliminary, the study’s inclusion of children with complex clinical profiles—many with co-occurring conditions such as depression, anxiety or ADHD—mirrors clinical realities and supports the potential utility of the method in real-world settings.
“High-quality, ethically collected data are rare in this domain,” Canavan said. “Our privacy-first design and context-aware analysis provide promising, objective insights that could help clinicians identify and monitor PTSD in children.”
If validated through larger trials, this privacy-preserving, AI-driven approach could change how clinicians detect and follow PTSD in children—turning everyday video data into actionable, ethical support for mental health care.
About this AI and PTSD research news
Author: John Dudley
Source: University of South Florida
Contact: John Dudley – University of South Florida
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Multimodal, context-based dataset of children with Post Traumatic Stress Disorder” by Alison Salloum et al. Pattern Recognition Letters
Abstract
Multimodal, context-based dataset of children with Post Traumatic Stress Disorder
Clinical PTSD diagnosis often depends on subjective interpretation of symptoms within a specific context. AI-driven solutions for these sensitive areas must adopt similarly careful methodologies.
To support research and development, we present a de-identified, multimodal dataset of children clinically diagnosed with and without PTSD in multiple conversational contexts. The dataset is intended to facilitate future work that preserves participant privacy while enabling objective assessment.
Each participant completed several sessions with clinicians and/or guardians designed to elicit emotional responses. From session videos, we extract facial features that remove identity information, including facial landmarks, head pose, action units (AU), and eye gaze.
As a baseline, we evaluate PTSD classification using encoded AU intensity vectors, which capture a subject’s expressiveness. We train a transformer-based classifier that encodes low-dimensional AU vectors with a learnable Fourier representation and combine this encoding with a multilayer perceptron (MLP) mapping of AU intensities.
Our experiments demonstrate that this combined encoding outperforms individual components, and that incorporating conversational context (for example, clinician-child versus parent-child discussions) is essential for accurate classification of children’s videos.